Corrections. Spot a factual error, misquote, mischaracterization, broken link, or anything else that needs correcting? Please email allsourcelegalai@substack.com. If a material issue is confirmed, the publication will aim to append a dated correction or clarification promptly, ordinarily within 72 hours where practicable.

The first two sub-pieces of Part 2 covered five roles inside private practice and the corporation. This one covers the three roles where the architecture meets its hardest constraints: the litigation support specialist who runs the data infrastructure the connectors reach into, the government attorney who works inside procurement and security rules the private sector does not face, and the public-interest or legal aid lawyer who serves the clients least able to absorb an error.

A quick refresher on the vocabulary that travels across every role section:

• The stack consists of three layers, as defined in Part 1 and used throughout this series.

• Work surface means Microsoft 365 with Claude inside Word, Excel, and PowerPoint, with Outlook in public beta as of the May 7 announcement.

• Workflow layer means Anthropic’s twelve practice-area legal plugins and twenty-plus connectors.

• Integration layer means CoCounsel Legal, grounded in Westlaw, or LexisNexis’s Lexis+ with Protégé, which integrates the Claude legal plugin suite inside the Lexis environment. Integration-layer access means whether the role has either one in the working environment.

• Adjacent systems sit at the edges: document management, eDiscovery, and contract management platforms from outside the May 7 to May 13 announcements, increasingly wired into the stack through standardized connections.

• eDiscovery, or electronic discovery, is the identification, collection, processing, review, and production of electronically stored information in litigation.

• Grounded retrieval means the technical mechanism by which a connected AI system searches a known database before generating output.

• MCP, the Model Context Protocol, is a standardized, open-protocol connection that allows AI systems to access external systems without custom integrations.

Product-status descriptions in this piece reflect the cited public record as of the publication date; vendor deployments, beta status, and integration scope can change quickly.

Three lessons from the prior articles travel through every role section. The architecture is still stabilizing in public. Announcement language often exceeds the scope of deployment. Operational adoption is outpacing verification discipline, and unchecked AI output has already reached court filings, which is reason enough to learn the checking habits each role section below describes.

In these three roles, the lessons take a different shape than they did at the firm. The first five roles asked what the architecture does to the work. These three also ask how the work’s constraints affect the architecture. The litigation support specialist works where data governance is actually practiced. The government attorney works where the tools often cannot follow. The public-interest lawyer works where the budget for any of it may not exist.

The first five roles asked what the architecture does to the work. These three ask what the work’s constraints do to the architecture.

The role profiles that follow are analytical composites built from the architecture documented in Part 1 and the deployment patterns described in the public record. They are illustrative, not survey findings, and not claims about every organization or rollout.

Role 6: The litigation support specialist

On a Monday, the litigation support specialist at a midsize firm has a production going out the door, a load file that will not validate, two attorneys asking why their search results differ, and a new matter arriving as four hundred gigabytes of mixed email and chat data. The specialist’s day does not start in Word. It starts in the systems that the rest of the firm’s AI stack is trying to reach.

This is the role that has been doing structured, tool-mediated legal work the longest. eDiscovery platforms, document management systems, processing engines, and review databases were the specialist’s daily environment for decades before the May announcements. For every other role in this series, the architecture arrived as something new on top of the work. For the litigation support specialist, the architecture arrived as new connections into systems the specialist already runs.

The role

The litigation support specialist, sometimes titled eDiscovery specialist, litigation support manager, or practice support analyst, runs the data infrastructure of a case. Collections, processing, early case assessment, search, review database administration, productions, load files, exhibit management, and trial support. The specialist is the person attorneys call when the data does not behave, and the person who knows what is actually in the database rather than what everyone assumes it is.

The specialist is also, in many firms, the closest thing the building has to a working expert in how software-mediated legal work actually fails: bad extractions, dropped attachments, deduplication errors, search syntax that silently returns the wrong universe. Quality control on tool output is not a new discipline at this desk. It is the job description.

Work surface

The specialist works in the same Microsoft 365 environment as the firm, and Claude in Excel may be the most immediately useful piece of it: tracking collections, building production logs, reconciling exhibit lists, and analyzing review metrics are spreadsheet work, and the specialist lives in spreadsheets. Claude in Outlook helps with the constant coordination among case teams, vendors, and opposing litigation support contacts.

But the work surface is the smaller part of this role’s story. The specialist’s center of gravity is the adjacent systems, and that is where the May architecture changes the most.

Workflow layer and the adjacent systems

For the other roles, the workflow layer means the practice-area plugins. For the specialist, the more consequential development is that the connectors reached the specialist’s actual daily environment first. The launch’s eDiscovery and litigation connectors include Relativity, Everlaw, and Consilio, and the Relativity connector was built for exactly this desk. Relativity’s own description of the integration is administrative orchestration: a system administrator can ask Claude in natural language to stand up a matter, provision a workspace, modify templates, configure permissions, and validate access, with Claude calling the tools. Every action is executed under the authenticated user’s identity and is fully audited in RelativityOne, with the user who triggered it attributed.

Read that description against the rest of this series, and the pattern is unmistakable. At the attorney roles, the architecture touched drafting and research. In this role, the first thing it touches is the administrative work that makes everything else possible, and it arrived wrapped in the vocabulary this desk already enforces: identity, permissions, audit, attribution, defensibility. Relativity built the connector around the governance realities the specialist lives in, which is both a convenience and a signal of who the integration expects to run it.

A second integration followed nine days later, on the other side of the pipeline. On May 21, Relativity announced an integration with the Claude Compliance API, bringing Claude Enterprise activity into RelativityOne as collectible data, alongside the platform’s existing native collection from ChatGPT Enterprise and Gemini Enterprise. That sentence deserves a slow read at this desk. The same AI layer the firm is deploying as a productivity tool is also becoming a data source that the specialist may have to collect from. Employee AI conversations are enterprise data, enterprise data becomes legal data, and the role that turns one into the other, defensibly, is this one. In this series, the specialist is the clearest example of a role sitting on both sides of that line.

Employee AI conversations are enterprise data, enterprise data becomes legal data, and the role that turns one into the other, defensibly, is this one.

The document management side matters less at this desk than the announcements might suggest. iManage’s MCP server, announced May 14, connects AI systems to governed firm documents, but the document management system is primarily the lawyers’ world; the specialist’s involvement with it is typically light. The connections that change the specialist’s day are the ones into the review platform, and those are the ones that shipped first.

When an AI layer can reach into a review platform, the questions that follow are the specialist’s questions: which workspaces are exposed to the connection, whose permissions govern what the AI can retrieve, how access is logged, what happens to confidentiality designations and protective-order restrictions when retrieval runs through a new pathway, and whether the connection’s behavior is defensible if a production built on it is ever challenged. The Relativity connector’s identity-bound, fully audited design answers some of those questions by construction. The specialist is the person who verifies that the construction holds in practice.

The litigation plugin’s named agents also touch the specialist’s lane. A Chronology Builder and a Privilege Log Reviewer, both present in the repository, sit on work the specialist often shares with associates and paralegals. The specialist who knows what these agents do and what data they need to do it is positioned to be the person who configures, governs, and quality-controls them, rather than the person surprised by them.

Integration layer

The specialist typically does not run CoCounsel Legal or Lexis+ with Protégé for legal research; that is attorney work. The specialist’s integration story is different: the specialist is often the person other people ask to evaluate, connect, secure, and troubleshoot whatever the firm plugs in. When the firm considers a new connector, the practical questions about authentication, data flow, export behavior, and audit trail tend to land at this desk, whether or not the title says so.

That makes the specialist’s verification role distinctive. The attorney checks whether a citation is real. The specialist checks whether the pipeline that produced it is sound: whether the connector reached the right repository, whether the search hit the full universe, and whether the output format preserved what the input contained. Both checks matter. Only one of them appears in most discussions of legal AI verification, and it is not the specialist’s.

What this role should know

The specialist’s leverage in this architecture is governance fluency. Permissions, logging, matter walls, data handling, and defensible processes are the areas that the new connection pathways emphasize, and this role already understands them better than anyone else in the building. The specialist who can explain to a partner what an MCP connection actually exposes, in plain language, becomes more central to the firm’s AI decisions, not less.

The skills that transfer are quality-control skills. The discipline the specialist applies to a production is the same discipline AI-assisted workflows need: validate, sample, reconcile, and document. The specialist has run that discipline on tool output for years. The architecture raises the volume of tool output without changing what sound checking looks like.

The risk to watch is a silent scope change. Part 1 documented a launch-window commit that removed a vendor’s references from the public configuration on launch day itself. Connector configurations, default settings, and exposed capabilities can shift without the people relying on them noticing. The specialist’s habit of testing what a system actually does, rather than trusting its documentation, is the right one for this environment and worth teaching to everyone else.

The next twelve months: what this role can expect

• More connector evaluations landing on this desk, as firms wire AI layers into review and document platforms and someone has to test what each connection actually does.

• More governance questions arriving from attorneys and firm leadership, and more weight given to the people who can answer them in plain language.

• Growing demand for translators who can sit between the attorneys and the architecture and explain each side to the other.

• A rising profile for the role itself: the desks that administer the systems the AI reaches into are not on the margin of this transition. They are at the junction of it.

At a glance: what the architecture touches at this role

• Claude in Excel touches the core tracking work: collections logs, production logs, exhibit reconciliation, and review metrics. Claude in Outlook touches vendor and case-team coordination.

• The connectors reached this desk first: Relativity, Everlaw, and Consilio are on the launch list, and the Relativity connector is an administrative orchestration built for this role, identity-bound and fully audited.

• The May 21 Compliance API integration puts the specialist on both sides of the line: Claude is a layer above RelativityOne and a collectible data source inside it, with Claude Enterprise activity now natively collectible alongside ChatGPT Enterprise and Gemini Enterprise.

• Governance territory is the touched surface: workspace exposure, permissions, access logging, confidentiality designations, and whether a connection’s behavior is defensible if challenged.

• The specialist’s verification role is the pipeline, not the citation: whether the connector reached the right workspace, hit the full universe, and preserved the data.

Role 7: The government attorney

On a Tuesday, the government attorney, a prosecutor in a county office, or counsel at a state agency has a motion due, a records request to respond to, and a caseload that would be three lawyers’ work at a firm. Down the hall, the IT department’s answer to most new software is the same answer it gave last year: not until it clears security review.

The government attorney performs the same professional duties as every other lawyer in this series, but within a set of constraints that none of the private-practice roles face. Procurement rules govern what can be bought. Security and data-handling rules govern what can be connected. Records laws govern what must be retained and disclosed. The May architecture was built to connect to everything. Government practice is built to control what connects.

The May architecture was built to connect to everything. Government practice is built to control what connects.

The role

Government attorneys span a wide range: prosecutors, public defenders employed by the government, agency counsel, city and county attorneys, and attorneys general's staff. What they share is practice inside public institutions, where the client is the government or the public, the caseloads are often heavy, the support staff is thin, and the technology environment is decided by procurement cycles and security offices rather than by a practice group’s budget.

The daily docket looks different across that range, and most of it happens away from a keyboard. A prosecutor’s day runs through charging decisions, plea negotiations, court appearances, grand jury work, and meetings with investigators, victims, witnesses, and defense counsel. Agency counsel draft regulations, interpret statutes, set up enforcement procedures, and argue for the government. Attorneys general and their staff counsel state agencies, issue opinions, and defend the state’s laws and lawsuits. Public defenders carry the defense side of the same courtrooms, with caseloads and resources that have more in common with the legal aid desk in the next section than with any firm role. The writing that AI tools touch is the part of the day that happens between courtrooms, interviews, and hearings: the motions, briefs, opinions, and advice memos. That is exactly why the time pressure on it is so heavy.

The duties are the same duties this series has described at every role: competence, candor, confidentiality, and the obligation to check what goes out the door under the attorney’s name. What differs is the environment in which those duties operate.

Work surface

Whether the government attorney has Claude in the Microsoft 365 environment depends entirely on what the institution has authorized. Government deployment of commercial AI tools runs through procurement, security assessment, and data-handling review, processes that exist for good reasons and that move at institutional speed. Some offices will have authorized tools available. Others will have none, and a written policy saying so.

That makes the government attorney’s first question different from every other role’s. The associate asks how to use the tools well. The government attorney asks, first, what is actually authorized here, and the answer is in the office’s policy, not in a vendor’s announcement. Using an unauthorized consumer AI tool on government matter data is not a productivity shortcut; it is a data-handling problem, a records problem, and potentially a confidentiality problem, all at once.

Workflow layer

Where an institution has authorized a deployment, the workflow layer’s logic applies the same way it does at a firm: plugins encode workflows, connectors define reach, and the configuration deserves the same scrutiny. The public nature of the claude-for-legal repository is useful here in a way that is easy to miss: a government office can inspect exactly what a plugin does, what connectors it assumes, and what data pathways it opens, before any procurement conversation begins. For institutions that must justify technology decisions on the record, a publicly inspectable configuration is a genuine aid.

The constraint cuts the other way too. The architecture’s strength is connection, and connection is precisely what government data-handling rules scrutinize. An office evaluating the stack is not evaluating a chatbot; it is evaluating every pathway the connectors open, and every place where matter data could travel. That evaluation is slower than a firm’s, and it should be.

Integration layer

Integration-layer access in government offices varies as widely as everything else. Some offices carry research subscriptions comparable to those of a midsize firm. Others rely on lower-cost research tools and public sources. The repository’s default research connectors, as configured in the public claude-for-legal templates, include CourtListener, the free public docket and opinion database, which matters more at this role than at any firm role: where budget constrains the integration layer, the free public layer is what grounding looks like.

The verification lesson is unchanged by the budget. Grounded retrieval against any source, commercial or public, can reduce the risk of fabricated or unsupported citations when the retrieval source, configuration, and output are checked, but it does not remove the need to read the authority. A government attorney’s filing carries the same basic signature-and-verification logic, whether under Rule 11, a state analog, a criminal rule, an agency rule, or a court’s inherent authority, and the checking habits described across this series apply with full force at a desk with a triple caseload and no associate to share them.

What this role should know

The government attorney’s institutional constraints are not only burdens; some are protections. Procurement review, security assessment, and authorized tool policies are the institutional version of the governance that in-house counsel, in the previous sub-piece, has to build from scratch. The government attorney who understands what the office’s policy actually permits and why is working within the constraint rather than around it.

Records and disclosure obligations deserve specific attention. Government practice runs on retention schedules, public-records laws, and, for prosecutors, disclosure duties to the defense. Any use of AI in government legal work should be evaluated under the office’s governing records, retention, disclosure, procurement, security, and ethics rules: what a tool retains, what becomes a record, and what may be discoverable. These are questions to resolve with the office’s records and ethics guidance before adoption, not after.

The caseload pressure that makes AI assistance attractive at this role is the same pressure that makes shortcuts dangerous. The honest framing is the one this series has used throughout: the tools compress drafting time, and they do not compress the checking. An office adopting authorized tools gains the most by pairing them, from the start, with the verification habits the heavy caseload tempts everyone to skip.

The next twelve months: what this role can expect

• More government-specific deployment conversations, as vendors pursue public-sector authorization and offices weigh what to permit.

• More office policies moving from silence to explicit rules, and more value in knowing what the office’s policy actually says before the question comes up mid-matter.

• Widening variation between well-resourced offices and thinly resourced ones, in tools and in training.

• A steady advantage for the government attorneys who learn the architecture’s logic now, from the public record, so the judgment is in place when the institution’s authorization arrives.

At a glance: what the architecture touches at this role

• Whether Claude touches this desk at all depends on institutional authorization: procurement, security review, and data-handling rules decide it, not the vendor announcement.

• Where authorized, the work surface and workflow layer touch drafting and caseload work the same way they do at a firm.

• The public claude-for-legal repository touches the procurement process itself: the office can inspect plugin configuration and connector pathways before any purchase.

• CourtListener as a default free research connector touches the grounding options where budget constrains commercial research subscriptions.

• Records, retention, and disclosure obligations are touched by any AI use: what a tool retains, what becomes a record, and what may be discoverable.

Role 8: The public-interest or legal aid lawyer

On a Wednesday, the legal aid lawyer has nine client appointments, an eviction hearing, an intake backlog the office triages weekly, and a budget in which a research subscription competes with a staff position. The clients are people for whom one legal error can mean a lost home, a lost benefit, or a lost case that cannot be appealed for lack of resources.

This is the role where the architecture’s promise and its risk are both at their sharpest. The promise: tools that compress drafting and research time are worth the most where the gap between legal need and legal capacity is widest, and nowhere is that gap wider than in legal aid. The risk: the clients at this desk are the least able to absorb the cost of an unchecked error, and the office is the least staffed to catch one.

The role

Public-interest and legal aid lawyers practice in housing, benefits, family, immigration, consumer, and civil rights, usually at nonprofit organizations funded by grants and public money. Caseloads are high, support staff thin, and the no-second-reader problem from the solo lawyer’s desk in the first sub-piece is the norm here, too, with less budget behind it.

The day is wider than the caseload. Intake and client interviews sit at the front of it: receiving low-income clients, hearing them out, and either representing them or advising them on their rights. Litigation runs alongside, sometimes aimed beyond the individual case to enforce rights or build precedent. Then comes the rest of the role that rarely makes a job posting headline: community education and know-your-rights outreach, policy and legislative advocacy, supervising junior attorneys, paralegals, interns, and volunteers, and documenting every case and every hour in the case management system, because the grants that fund the office require it. The drafting and documentation load runs heavy here, as it does across the profession, with client communication close behind.

The duties do not scale down with the budget. Competence, candor, confidentiality, and verification obligations apply in full at a legal aid desk, and the professional stakes of the work are amplified by the client’s stakes in it.

Work surface

Where the organization runs Microsoft 365, the work surface arrives the way it does everywhere else, and the productivity case is real: client letters, declarations, intake summaries, community education materials, and grant reporting are exactly the document-heavy, time-starved work the work surface compresses. The economics matter here more than anywhere in this series, and the launch record speaks to them directly: Anthropic’s legal launch included discounted access for legal aid organizations and public defenders through its Claude for Nonprofits program.

The confidentiality discipline from the in-house section of the previous sub-piece applies with particular force at this desk. Legal aid client data is often among the most sensitive personal information in any legal practice: immigration status, family violence, health, and benefits records. The duty to understand what any tool retains and where data travels comes before the productivity gain, and an organization-level policy beats individual improvisation.

Workflow layer

The public repository includes material aimed at the thin-budget end of practice. The connector defaults lean on free public sources, with CourtListener among the research connectors, and the repository’s legal clinic plugin, built for law school clinics, lists agents such as a Clinic Intake agent, a Case Memo Scaffold, and a Client Letter Drafter that map closely onto legal aid workflows, even though the plugin’s named audience is clinical education. The fit is not exact, and an organization adapting clinic-oriented workflows to legal aid practice should treat them as starting points for configuration, not as finished tools to trust.

The open question at this role is who does the configuring. Firms have litigation support specialists and IT departments. Legal aid organizations often have one overextended operations person. The architecture’s open structure means the workflows are inspectable and adaptable in principle; the constraint is the staff time to do it. The gap between what the public configuration offers and what a legal aid office can deploy is now easier to see and describe than it used to be, which puts it squarely in front of the funders, bar foundations, and access-to-justice technology groups that decide where support goes.

Integration layer

Integration-layer access is the architecture’s most expensive tier, and at many legal aid organizations, it will be the missing one. Commercial grounded-research subscriptions carry costs that many legal aid budgets cannot absorb. Where they are absent, the free public layer carries the load: public dockets, public opinions, and the office’s own brief banks and form libraries.

The verification picture follows the budget. Grounded retrieval, against commercial databases or public sources, can reduce the risk of fabricated or unsupported citations when the retrieval source, configuration, and output are checked; it does not remove the need to read the authority. The legal aid lawyer drafting with work-surface tools and researching against public sources is working at the architecture’s thinnest tier, where the checking habits matter most precisely because the institutional backstops are fewest.

What this role should know

The highest-value uses at this desk are those that compress time without touching legal authority: intake summaries; client communication in plain language; declaration drafting from interview notes; translation support with human review when legal rights, declarations, court filings, or client instructions are involved; and grant reporting. These capture the productivity gain at the lowest verification cost, because the lawyer is checking facts they already know, rather than an authority system retrieved.

Research assistance deserves the utmost care at this very desk. The combination of high caseloads, no second reader, and clients who cannot absorb an error is the combination in which an unchecked citation does the most damage. The habit this series has described at every role, read the authority before relying on it, is not a counsel of perfection here. It is the floor.

Read the authority before relying on it. At the legal aid desk, that habit is not a counsel of perfection. It is the floor.

The organization-level move is a short, written AI policy, even for a small office, that outlines which tools are approved, what client data may be entered into them, and what checks are required before anything reaches a court or a client. The in-house counsel section of the previous sub-piece describes the logic; a legal aid office needs the same thing on a smaller scale, and writing it down once beats deciding it case by case under a deadline.

The next twelve months: what this role can expect

• More attention from the access-to-justice community to what the open architecture makes possible at low cost, and more pilots aimed at thin-budget practice.

• More pressure on funders and bar foundations to support the configuration and training work the architecture assumes someone can do.

• Growing divergence between organizations that adopt with a policy and habits in place and organizations where adoption happens one improvising lawyer at a time.

• The same stakes that define the desk: the clients here have the most to gain from tools that extend thin capacity, and the most riding on the discipline with which those tools are used.

At a glance: what the architecture touches at this role

• The work surface touches the document-heavy, time-starved work at the architecture’s least expensive tier: client letters, declarations, intake summaries, grant reporting.

• Any tool use touches highly sensitive client data: immigration status, family violence, health, and benefits records. The confidentiality discipline comes ahead of the productivity gain.

• The legal-clinic plugin’s agents, including Clinic Intake, Case Memo Scaffold, and Client Letter Drafter, map onto legal aid workflows as adaptable starting points rather than finished tools.

• CourtListener and public sources are the integration layer here: grounding at the free tier, where the checking habits matter most because the institutional backstops are fewest.

• The highest-value touches avoid legal authority entirely: plain-language client communication, declarations from interview notes, translation support with human review, and reporting.

Looking ahead

That completes the eight roles. Across three sub-pieces, the same architecture has landed eight different ways: on the solo lawyer’s economics, the paralegal’s changing task list, the associate’s drafting day, the partner’s signature, in-house counsel’s double burden, the specialist’s governance territory, the government attorney’s constraints, and the legal aid lawyer’s stakes. The architecture is the same. The working reality never is.

Part 3 reorganizes the same facts one more way: by practice area. Litigation, corporate, employment, privacy, regulatory, intellectual property, and adjacent fields, each read against what the May architecture actually shipped.

Sources

• Anthropic, “Collaborate with Claude across Excel, PowerPoint, Word and Outlook,” May 7, 2026

• Anthropic, “Claude for the legal industry,” May 12, 2026

• Anthropic claude-for-legal GitHub repository

• Reuters, “Anthropic expands Claude’s AI tools for law firms, lawyers,” May 12, 2026

• Thomson Reuters, “Thomson Reuters and Anthropic Expand Partnership to Connect Claude with CoCounsel Legal,” May 12, 2026

• LexisNexis, “LexisNexis Expands Lexis+ with Protégé by Integrating Anthropic’s Claude Legal Plugin Suite,” May 13, 2026

• iManage, “iManage MCP Server is now Available to Connect Governed Knowledge to the Broader AI Ecosystem,” May 14, 2026

• Relativity, “Making Administration Conversational with RelativityOne and Claude,” May 2026

• Relativity, “Relativity Adds Collection of Claude Enterprise Data with Claude Compliance API Integration,” May 21, 2026

• LawSites, “Anthropic Goes All-In on Legal, Releasing More Than 20 Connectors and 12 Practice-Area Plugins for Claude,” May 12, 2026

• CourtListener, Free Law Project

• American Bar Association, Model Rule 1.1, Competence

• American Bar Association, Model Rule 1.6, Confidentiality of Information

• American Bar Association Standing Committee on Ethics and Professional Responsibility, Formal Opinion 512, “Generative Artificial Intelligence Tools,” July 29, 2024

• Damien Charlotin, “AI Hallucination Cases Database“

• Federal Rule of Civil Procedure 11

A Note on Methodology: This piece is based on the sources cited below, including company announcements, public repository records, press releases, and legal authorities. High-risk factual assertions, dates, and direct quotations were checked against the cited materials. The role profiles are analytical composites derived from the cited architecture and public reporting; they are illustrative, not statistical claims about all deployments. Where the architecture has not yet been deployed at a given role, that is identified directly. Where public announcement language differs from documented implementation scope, this article separates the broader framing from the narrower operational claim. This article discusses the ABA Model Rules as a general reference framework. Professional-responsibility duties vary by jurisdiction and may be affected by state rules, local court rules, agency rules, ethics opinions, protective orders, client instructions, procurement requirements, records-retention rules, and organizational policies.

AI tools used in preparation. This article was drafted and revised with assistance from Claude and ChatGPT. Factual claims were checked against primary sources using Perplexity and ChatGPT. Copyediting was performed with Gemini. Editorial judgments, framing, and analytical conclusions are the author’s.

This article is for informational and educational purposes only. Nothing herein constitutes legal advice, and reading this article does not create an attorney-client relationship. Readers should consult qualified legal counsel for advice specific to their situation.

Copyright © 2026 Robert Meyers-Lussier. All rights reserved.

Part 2 of a three-part series. Part 1 walked the components. Part 2 walks the roles. Part 3 will walk the practice areas. Part 2 publishes in three sub-pieces. This is the second. The first covered the solo or small-firm lawyer and the paralegal. The third covers litigation support specialists, government attorneys, and public-interest or legal aid lawyers.

Corrections. Spot a factual error, misquote, mischaracterization, broken link, or anything else that needs correcting? Please email allsourcelegalai@substack.com. If a material issue is confirmed, the publication will aim to append a dated correction or clarification promptly, ordinarily within 72 hours where practicable.

The first sub-piece of Part 2 covered the two roles in which the productivity tier lands first and integration-layer access is uneven, emerging, or absent: the solo or small-firm lawyer and the paralegal.

This sub-piece covers three roles where the picture inverts. Integration-layer access is often standard. The work surface is taken for granted. And the supervisory chain spans more than one person, which means the exposure no longer sits on a single desk.

A quick refresher on the vocabulary that travels across every role section. The stack is three layers. Work surface refers to Microsoft 365 with Claude in Word, Excel, PowerPoint, and Outlook. Workflow layer means Anthropic’s twelve practice-area legal plugins and twenty-plus connectors. Integration layer means CoCounsel Legal, grounded in Westlaw or Lexis+, with Protégé connected to Lexis. Adjacent systems sit at the edges: document management, eDiscovery, and contract management platforms that come from outside the May 7 to May 13 announcements but are increasingly wired into the stack through standardized connections. Integration-layer access means whether the role has CoCounsel Legal or Lexis+ with Protégé in the working environment. Grounded retrieval means the technical mechanism by which a connected AI system searches a known database before generating output. MCP, the Model Context Protocol, is a standardized, open protocol that allows AI systems to connect to external systems without custom integrations. Product-status descriptions in this piece reflect the cited public record as of the publication date; vendor deployments, beta status, and integration scope can change quickly.

Three lessons from the prior articles travel through every role section. The architecture is still stabilizing in public. Announcement language often exceeds the scope of deployment. Operational adoption is outpacing the verification discipline. The reported phantom-quotations filing in a federal layoffs case, drafted partly through Claude Console, is the working example of that third lesson reaching a court filing.

At the three roles in this sub-piece, those lessons stop being individual concerns and become institutional ones. The associate produces work that another person signs. The partner signs work other people produced. The in-house counsel sets the rules governing how the entire workforce uses the stack. The verification chain now has links.

The architecture is the same. The exposure moves.

The role profiles that follow are analytical composites built from the architecture documented in Part 1 and the deployment patterns described in the public record. They are illustrative, not survey findings, and not claims about every firm, department, or rollout.

Role 3: The litigation associate at a midsize-to-large firm

On a Wednesday at the firm, the second-year associate has a research memo due to the partner by the end of the day and a section of a summary-judgment brief due tomorrow. At a midsize-to-large firm, all three layers of the stack are often available to the associate, though what is actually provisioned varies by practice group, license tier, and the firm’s rollout progress. In a fully provisioned setup, the work surface is open in Word and Outlook. The workflow layer is deployed: the firm has rolled out the litigation plugin, which runs beneath the drafting environment. The integration layer is connected: CoCounsel Legal with Westlaw Deep Research, or Lexis+ with Protégé on the Lexis side, is one pane away, with the specific capability depending on which product the firm has configured. The adjacent systems are wired in: iManage holds every prior memo, brief, and template the firm has produced on this kind of matter.

This is the role the architecture was built for. In firms that have deployed these tools, the associate has the full stack and the institutional support to run it. The associate also has the least control over how the output is used, because the associate is often not the one who signs the filing.

The role

The litigation associate researches, drafts, and revises. Memos, brief sections, deposition outlines, discovery responses, chronologies, and fact summaries. The associate is the person who reads the cases, builds the argument, and produces the first complete version of the document that the partner will review, revise, and sign.

The associate is measured on output and billable hours, which puts the associate in the role under the most direct pressure to adopt tools that produce output faster. A partner who tells associates to be careful with AI is also a partner who expects the memo by the end of the day. The associate lives inside that tension every working week.

The associate is often not the signing attorney on the filing. The partner usually is. But the associate produces the draft that the partner relies on, which means the quality of the associate’s verification determines what the partner is signing, whether the partner knows it or not.

Work surface

Same Microsoft 365 environment as the rest of the firm, fully provisioned rather than the low-admin setup at the solo desk. Claude in Word for drafting memos and brief sections. Claude in Excel for damages models, deposition designations, and document review tracking. Claude in Outlook for the constant internal coordination an associate runs with partners, paralegals, co-counsel, and the client’s in-house team.

What this looks like on a working day. The partner asks for a memo on whether a particular affirmative defense survives a recent appellate decision. The associate opens Word with Claude, frames the question, and gets a first structural pass in minutes rather than the half day a cold start used to take. The associate then does the actual legal work: reading the appellate decision in full, checking whether Claude’s framing of the holding matches the court’s language, tracing every citation Claude proposed back to the reporter.

The speed gain is real and front-loaded. The first draft arrives faster. The verification work that follows is the same work it always has been, and in this role, it is the work that separates a usable memo from a liability.

Workflow layer

This is the role where the workflow layer is most likely to be fully deployed and actively used. The firm has configured the litigation plugin, managed the identity and tenant administration the plugin assumes, and set the policy oversight the solo desk could not. The associate runs inside that configured environment.

The associate can also read the claude-for-legal repository directly, the same public configuration the solo lawyer can read, and the same configuration that changed inside the launch window when the May 12 commit titled “Remove Lexis references at partner request” modified the visible stack. For the associate, the lesson is operational. The plugin the firm deployed encodes a particular sequence of steps and a particular set of connector assumptions. When the firm updates the plugin, the sequence can change. The associate who understands what the plugin is actually doing underneath the drafting environment is better positioned than the associate who treats it as a black box.

The workflow layer now also includes the named agents Anthropic has published in the repository: job-style tools such as a Brief Section Drafter, a Chronology Builder, a Deposition Prep agent, and a Privilege Log Reviewer, each runnable with a single command. These map directly onto associate tasks. The associate who knows which agents exist for litigation work has a more complete picture of what the firm’s deployment can actually do.

Integration layer

This is where the associate’s experience diverges most sharply from the solo lawyer’s. At a midsize-to-large firm, integration-layer access is often standard. The associate runs CoCounsel Legal with Westlaw Deep Research or Lexis+ with Protégé as part of the normal research workflow, with the firm covering the subscription costs that the solo desk could not absorb.

Grounded retrieval changes the failure mode but does not remove it. When the associate runs a research question through CoCounsel Legal, the system searches Westlaw before generating the draft, which lowers the chance that a fabricated citation reaches the draft. It does not lower it to zero. The associate still has to read the cases the system returns, confirm that they say what the system says they say, and confirm that they are still good law. The grounded tier is faster and more reliable than ungrounded drafting. It is not a substitute for reading the cases.

The grounded tier is faster and more reliable than ungrounded drafting. It is not a substitute for reading the cases.

The operational-scope lesson from Part 1 applies directly to this role. The CoCounsel Legal capability exposed through Claude is Westlaw Deep Research, specifically, a research workflow that runs a search against Westlaw and returns a report, exposed through a limited set of tool calls, not the full CoCounsel Legal platform. The associate who needs drafting support, file summarization, or other CoCounsel Legal capabilities is directed back to the native CoCounsel environment. The associate should know which capabilities are actually available in which environment before building a workflow based on an assumption about what the integration does.

What this role should know

The associate’s verification work is the foundation on which the partner’s signature rests. Every citation in an associate-produced draft has to be read against the actual reporter before the draft goes to the partner. Every quotation has to be checked against the source. Every proposition of law has to be traced to a case that actually states it. The partner will review the draft, but the partner is reviewing it on the assumption that the associate already did that work. When that assumption is wrong, the chain breaks at a point the partner cannot see.

AI hallucination in court filings is a documented and growing concern, tracked in more than 1,000 cases across public databases such as Damien Charlotin’s AI Hallucination Cases Database. The database covers more than lawyer sanctions; it includes pro se litigants and a range of outcomes. Many of the lawyer-involved cases follow a consistent pattern. The work was produced with AI assistance, the citations were not independently verified, and the filing reached the court with fabricated or misattributed authority. The associate is frequently the person who produced the underlying draft. Making AI use visible to the supervising partner is one of the most protective practices available in this role, because it tells the partner’s verification step where to focus.

The associate has professional exposure that the supervisory structure does not provide. The duties of competence, candor, and verification under the rules of professional conduct attach to the associate as a lawyer, not only to the partner who signs. A partner’s instruction to produce work fast does not eliminate the associate’s own obligation to check it. The associate who files AI-assisted work without checking the authority, or who passes unchecked work up the chain, knowing the partner may not catch it, is acting against duties the associate holds independently.

The next twelve months

More firms are moving from pilot deployments to full litigation-plugin rollouts. More associate workflows are built on the named agents in the repository. More firm-level policy on how associates flag AI use when work moves up to the signing attorney. More pressure on the billable-hour model as the work surface compresses the time a first draft takes, and more attention to what that compression does to associate training and development. The verification work does not compress. It is the part of the job the architecture does not do.

Role 4: The litigation partner at a midsize-to-large firm

On a Thursday, the partner has four matters moving at once and a brief to sign by close of business. The partner did not draft the brief. An associate did so, with AI assistance, within the firm’s deployed stack. A paralegal assembled the exhibits and flagged which sections came from Claude and which citations came from CoCounsel Legal. The brief is on the partner’s screen now, and the partner’s signature is what turns it from a draft into a filing.

The partner sees the least of the stack day-to-day and carries the most exposure to it. The partner may not open Claude in Word at all. The partner reviews the output that other people produced inside the architecture, and the partner signs.

The role

The litigation partner supervises, signs, and carries the relationship with the client and the court. The partner sets strategy, reviews the work associates and paralegals produce, makes the judgment calls, and puts a signature on the documents that go to the court. The partner is often the lead signing attorney. The partner is the person the court can hold responsible when a filing contains fabricated authority.

The partner’s exposure is structural, and it runs in two directions. Under the federal rules, the signature on a filing certifies that the partner has conducted a reasonable inquiry and that the claims and contentions are warranted, not that the partner personally guarantees every assertion in the abstract. Under the rules of professional conduct, the partner carries supervisory responsibility for the lawyers and nonlawyers the partner directs. The partner is signing for the associate’s research, the paralegal’s assembly, and the partner’s own judgment, all at once.

Work surface

The partner’s use of the work surface is often lighter and more selective than the associate’s. Claude in Outlook to triage a heavy inbox and draft responses to opposing counsel and clients. Claude in Word to revise rather than to draft from scratch, tightening an associate’s section or reworking an argument. Claude in PowerPoint for client presentations and pitch materials. The partner is more likely to use the work surface as a revision and communication tool than as a primary drafting environment.

What this looks like on a working day. The associate’s brief lands in the partner’s inbox at two in the afternoon. The partner reads it. The argument is sound. The structure is right. The citations look correct. Here is the decision the architecture has made sharper: the partner can sign on the strength of the read, or the partner can verify the citations independently before signing. The work surface did not create that decision, but the speed of AI-assisted drafting has increased the volume of work flowing up to the partner’s signature, which makes the decision recur more often and under more time pressure.

Workflow layer

The partner is often less likely than the associate to interact directly with the workflow layer and is among the most affected by its configuration. The partner generally does not run the litigation plugin. The associate does. But the partner is the person who should understand, at a policy level, what the firm’s deployed plugins do, what the named agents are configured to produce, and where the verification gaps are. A partner who signs work produced through a workflow the partner does not understand is signing on trust alone.

The architecture-stabilizing lesson presents itself to the partner as a governance question. The May 12 Lexis removal commit showed that the stack's visible configuration can change within a launch window. For the partner, the implication is that the firm’s AI deployment is not a fixed thing the partner can understand once and then ignore. It is a moving system, and the partner’s supervisory obligation attaches to whatever it is currently doing.

Integration layer

The partner relies on integration-layer access without necessarily operating it. When the associate runs CoCounsel Legal with Westlaw Deep Research and weaves the grounded citations into the brief, the partner inherits the benefit and the residual risk. Grounded retrieval reduces the rate at which fabricated citations reach the draft. It does not eliminate the partner’s obligation to verify before signing. A grounded citation is more likely to be real than an ungrounded one, but “more likely to be real” is not the standard the partner certifies to when the partner signs.

The signing decision is where every layer of the stack converges. The work surface produced the draft faster. The workflow layer structured it. The integration layer grounded the research. The adjacent systems supplied the firm precedent. All of that lands on the partner’s signature, and none of it discharges the partner’s certification obligation.

Every layer of the stack lands on the partner’s signature, and none of it discharges the partner’s certification obligation.

What this role should know

The partner’s signature is the certification the entire chain rests on, and AI assistance does not displace the lawyer’s duty of reasonable inquiry and supervision. A lawyer who signs a filing certifies that a reasonable inquiry was made. AI-assisted drafting does not change what a reasonable inquiry requires; if anything, the documented pattern of AI hallucination in filings raises the baseline of what reasonable inquiry now includes. The partner who signs an associate’s AI-assisted brief without confirming the citations were verified is certifying to something the partner has not confirmed.

The partner carries supervisory exposure on top of the signing exposure. Under the rules of professional conduct, the partner is responsible for ensuring that the lawyers and nonlawyers the partner supervises conform to the rules. That responsibility now extends to how associates and paralegals use AI. A partner who has not set clear expectations for how AI-assisted work is verified and flagged inside the team is carrying a supervisory gap that the architecture makes more consequential, because the architecture puts AI inside more of the work.

AI hallucinations create a real and growing sanction and malpractice risk that the signing attorney bears most directly, documented across a rising count of cases in public trackers. The lesson for the partner is not the specifics of any one case. It is that the courts have consistently placed responsibility on the signing attorney regardless of who did the drafting and regardless of whether AI was involved. The partner who treats verification as the associate’s job alone has misread where the exposure sits. The signature is the partner’s. So is the certification.

The most protective practice at this role is requiring AI use to be made visible up the chain, so the partner’s verification step is informed rather than blind. If the partner does not know which parts of a draft came from AI output, the partner cannot focus where the checking is needed. The firms reducing this risk are the ones building the flagging practice into the workflow, not the ones relying on individual diligence under deadline pressure.

The next twelve months

More firm-level AI governance is moving from informal guidance to written policy. More partner attention to what the firm’s deployed workflow actually does, driven by supervisory exposure rather than curiosity. More malpractice and professional liability attention to AI-assisted filings. More divergence between firms that build verification into the workflow and firms that leave it to individual diligence. The decision to sign is not getting easier. The volume of work reaching it is going up.

Role 5: In-house counsel

On a Friday, the in-house counsel at a regional employer has two jobs running in parallel. One is the legal work: reviewing a vendor contract, advising the business on a regulatory question, managing outside counsel on a piece of litigation. The other is the policy work: the company’s employees are already using AI tools, the company is being asked to approve new AI features in the products it buys, and the in-house counsel is the person expected to set the rules. Both jobs run through the same stack, and the in-house counsel sits on both sides of it.

This role is different from the three litigation roles because the in-house counsel is not only a user of the architecture. The in-house counsel is also the person who decides how an entire workforce is allowed to use it.

The role, as user

As a practitioner, the in-house counsel does much of what the other roles do, often with a smaller team and a broader subject-matter range. Contract review, regulatory advice, employment questions, commercial disputes, board and governance support, and management of the outside counsel who handle what the in-house team does not. The in-house counsel drafts, reviews, and advises, frequently as a department of one or a few, which means the in-house counsel often has the solo lawyer’s no-second-reader problem inside a corporate structure.

The role, as policy-setter

The second job is what makes this role distinct. The in-house counsel is often the person the business turns to when it asks whether employees can use a given AI tool, whether a contract’s AI clause is acceptable, whether a vendor’s new AI feature creates a data or privilege risk, and what the company’s policy on AI use should be. Some companies route those questions through privacy, security, IT, procurement, or compliance instead, but in many companies, they land with in-house counsel. The in-house counsel is not setting policy for a legal team of a few people. The in-house counsel is often shaping policy for an entire workforce, most of which is not legally trained, and much of which is already using AI tools, whether or not a policy exists.

That dual position is the defining feature of the role. The in-house counsel feels the burden of checking as a user and the governance burden as a policy-setter at the same time.

The in-house counsel feels the checking burden as a user and owns the governance burden as a policy-setter, at the same time.

Work surface

In-house counsel often works within the company’s standard productivity environment, which, for many companies, is Microsoft 365 with Claude. Claude in Word for contract markups and advice memos. Claude in Excel for tracking matters, budgets, and outside counsel spend. Claude in Outlook for the heavy internal communication and an in-house role that runs across business units. Claude in PowerPoint for board and leadership presentations.

The work-surface question for this role is not only how the in-house counsel uses it. It is how everyone else in the company uses it. While Claude is available across Microsoft 365 in the company’s tenant, in-house counsel’s own use is just one of hundreds or thousands. The marketing team is drafting in Word with Claude. The finance team is modeling in Excel with Claude. The sales team is in Outlook with Claude. The in-house counsel is the person who has to think about what all of that means for privilege, confidentiality, data handling, and record retention across the business.

Workflow layer

The in-house counsel encounters the workflow layer from two directions. As a user, the in-house counsel may run practice-area plugins relevant to corporate and commercial work, and the named agents in the repository map onto in-house tasks: a Vendor Agreement Reviewer, a DPA Reviewer for data processing agreements, a DSAR Responder for data subject access requests, and an AI Use Case Triager. As a policy-setter, the in-house counsel is the person who has to decide which of these workflows the company sanctions, which it restricts, and what verification the company requires before relying on AI-assisted work product.

The architecture-stabilizing lesson is sharper for the in-house counsel than for any other role, because the in-house counsel is setting policy against a moving target. A policy written around what the stack did in May may not describe what it does in August. The in-house counsel who writes an AI policy as a fixed document, rather than as a reviewed and updated one, is governing a system that does not hold still.

Integration layer

Whether in-house counsel has integration-layer access depends on the company's size and litigation posture. A large in-house department with active litigation may carry CoCounsel Legal or Lexis+ with Protégé. A smaller department may rely on the work surface and outside counsel for grounded research. Either way, the obligation to read the authority is the same as every other role’s. Grounded retrieval reduces the fabrication rate; it does not remove the need to read the authority.

The integration-layer question unique to this role is vendor management. The in-house counsel is frequently the person reviewing the contracts and data-processing terms for the AI tools the company buys, including the legal-research tools the legal team itself uses. The in-house counsel is reading the CoCounsel Legal or Lexis+ with Protégé terms not only as a user deciding whether to trust the output, but also as counsel deciding whether the data-handling, confidentiality, and retention terms are acceptable for the company’s matters.

What this role should know

The in-house counsel carries the verification burden twice. Once, as a practitioner, where every AI-assisted contract markup and advice memo needs the same checking as any legal work needs, often without a second reader inside the department. And once as a policy-setter, where the in-house counsel is responsible for whether the broader workforce’s AI use creates risk the company has not addressed through vendor terms, internal policy, and employee training.

The privilege and confidentiality questions are the ones most specific to this role. When employees across the business use AI tools, in-house counsel must consider whether confidential or privileged information is being entered into systems whose retention, training, and access terms the company has not vetted. AI use can create privilege and confidentiality risk when sensitive information enters tools that the legal department does not control and has not reviewed, against the backdrop of a lawyer’s duty to protect client confidences. These are governance questions the architecture has made urgent because it has put AI inside the everyday productivity tools the whole company already uses.

The policy-setting role is where the in-house counsel can do the most good and carry the most exposure. A clear, current, and enforced AI policy that sets expectations for verification, confidentiality, and acceptable use is one of the most protective measures in-house counsel can build. A policy that is written once and never updated, or written for the legal team but not for the workforce, leaves the gap open. The in-house counsel is the person positioned to close it, and the person accountable if it stays open.

AI hallucination in legal filings is a risk that in-house counsel manages from the client side, through the outside counsel relationship. The in-house counsel who sets expectations with outside counsel about AI use and how its output is checked is managing a risk that, when it goes wrong, lands on the company’s matter and the company’s outcome. The growing record of AI hallucination cases is a reason for the in-house counsel to make verification an explicit term of the outside counsel relationship, not an assumed one.

The next twelve months

More in-house departments are writing and revising formal AI policies. More attention to privilege and confidentiality as AI use in the workforce scales within the Microsoft 365 environment. More vendor-management scrutiny of the AI terms in the contracts companies sign. More in-house counsel are setting explicit AI-verification expectations with outside counsel. More pressure on small in-house departments that carry the user burden and the policy burden at the same time, with limited resources. The architecture put AI in front of the whole workforce. The in-house counsel is the person who has to govern that and use it at once.

Looking ahead

The third and final sub-piece of Part 2 covers the three roles where the architecture meets the parts of the system with the fewest resources and the highest stakes: the litigation support specialist and eDiscovery professional who runs the data infrastructure of a case, the government attorney who works inside procurement and security constraints the private sector does not face, and the public-interest or legal aid lawyer who serves the clients least able to absorb an error. The architecture is the same. The constraints are not.

Sources

• Anthropic, “Collaborate with Claude across Excel, PowerPoint, Word and Outlook,” May 7, 2026

• Anthropic, “Claude for the legal industry,” May 12, 2026

• Anthropic claude-for-legal GitHub repository, including commit 210fd7a, “Remove Lexis references at partner request”, May 12, 2026

• Reuters, “Anthropic expands Claude’s AI tools for law firms, lawyers,” May 12, 2026

• Thomson Reuters, CoCounsel Legal MCP Connector Guide

• LexisNexis, “LexisNexis Expands Lexis+ with Protégé by Integrating Anthropic’s Claude Legal Plugin Suite,” May 13, 2026

• iManage, “iManage MCP Server is now Available to Connect Governed Knowledge to the Broader AI Ecosystem,” May 14, 2026

• Bloomberg Law, “Lawyers Apologize for Fake AI Quotes in Trump Mass Layoffs Case,” May 2026

• Damien Charlotin, “AI Hallucination Cases Database“

• American Bar Association, Model Rule 1.1, Competence

• American Bar Association, Model Rule 1.6, Confidentiality of Information

• American Bar Association, Model Rule 5.1, Responsibilities of Partners, Managers, and Supervisory Lawyers

• American Bar Association, Model Rule 5.3, Responsibilities Regarding Nonlawyer Assistance

• Federal Rule of Civil Procedure 11

A Note on Methodology: This piece is based on the sources cited below, including company announcements, public repository records, press releases, court-related reporting, and legal authorities. High-risk factual assertions, dates, and direct quotations were checked against the cited materials. The role profiles are analytical composites derived from the cited architecture and public reporting; they are illustrative, not statistical claims about all deployments. Where the architecture has not yet been deployed at a given role, that is identified directly. Where public announcement language differs from documented implementation scope, this article separates the broader framing from the narrower operational claim.

AI tools used in preparation. This article was drafted and revised with assistance from Claude and ChatGPT. Factual claims were checked against primary sources using Perplexity, ChatGPT, and DeepSeek. Copyediting was performed with Gemini. Editorial judgments, framing, and analytical conclusions are the author’s.

This article is for informational and educational purposes only. Nothing herein constitutes legal advice, and reading this article does not create an attorney-client relationship. Readers should consult qualified legal counsel for advice specific to their situation.

Copyright © 2026 Robert Meyers-Lussier. All rights reserved.

Part 2 of a three-part series. Part 1 walked the components. Part 2 walks the roles. Part 3 will walk the practice areas. Part 2 publishes in three sub-pieces. This is the first. The next sub-piece covers litigation associates, litigation partners, and in-house counsel. The third sub-piece covers litigation support specialists, government attorneys, and public-interest or legal aid lawyers.

Corrections. Spot a factual error, misquote, mischaracterization, broken link, or anything else that needs correcting? Please email allsourcelegalai@substack.com. Confirmed corrections will be posted within 72 hours, with a note indicating what changed and when.

Part 1 of this series walked through the architecture. Microsoft 365 with Claude across Word, Excel, PowerPoint, and Outlook as the work surface. Twelve Anthropic legal plugins and twenty-plus connectors as the workflow layer. CoCounsel Legal is grounded in Westlaw and Lexis+ with Protégé connected to Lexis as the integration layer. Three layers of the stack. The May 14 iManage MCP server and surrounding announcements at the adjacent-system edges, outside the stack, but increasingly wired into it.

The workflow layer that arrived in full on May 12 was an expansion of the original Claude for Legal Cowork plugin Anthropic released in February 2026. The architecture has been building since the start of the year.

Part 2 walks the people. The architecture is the same across the legal workplace. The lived experience of it is not.

A solo lawyer taking small-business clients in Sacramento works inside one slice of the stack. A litigation paralegal at a midsize-to-large firm in Chicago works inside another. A litigation partner signing filings sees a different cut again. An in-house counsel managing employee AI policy at a regional employer sees another.

This article begins the role walk. It covers the two roles in which the productivity tier lands first and in which integration-layer access remains uneven, emerging, or operationally constrained. The remaining six roles follow across the next two sub-pieces.

A few terms travel across every role section. The stack itself is three layers. Work surface refers to Microsoft 365 with Claude in Word, Excel, PowerPoint, and Outlook. Workflow layer means Anthropic’s twelve practice-area legal plugins and twenty-plus connectors. Integration layer means CoCounsel Legal, grounded in Westlaw or Lexis+, with Protégé connected to Lexis.

Adjacent systems sit at the edges of the stack rather than inside it: document management, eDiscovery, and contract management platforms that come from outside the May 7 to May 13 announcements but are increasingly integrating with the stack through standardized connections. Integration-layer access means whether the role has CoCounsel Legal or Lexis+ with Protégé in the working environment. Grounded retrieval means the technical mechanism by which a connected AI system searches a known database before generating output. MCP, the Model Context Protocol, is a standardized, open-protocol connection that allows AI systems to access external systems without custom integrations.

Three lessons from the prior articles travel through every role section that follows.

The architecture is still stabilizing in public. The May 12 commit titled "Remove Lexis references at partner request" showed that the visible stack can change within a launch window.

Announcement language often exceeds the deployed scope. Thomson Reuters CTO Joel Hron’s Reuters clarification showed that the Claude integration does not replace CoCounsel Legal.

Operational adoption is moving faster than the verification discipline. The May 6 motion in American Federation of Government Employees v. Trump, drafted in part by Claude Console and containing phantom quotations, showed what happens when that gap reaches a court filing.

Each role experiences those three lessons differently.

Role 1: The solo or small-firm lawyer

On a Tuesday night in late May, the solo lawyer opens Word at the kitchen table. The brief is due Wednesday at four. Claude is in the side panel. The entire stack on this desk is a single layer: the work surface.

The workflow layer, Anthropic’s twelve practice-area legal plugins and twenty-plus connectors, is technically reachable but practically unconfigured at this scale. The integration layer, CoCounsel Legal, grounded in Westlaw and Lexis+, with Protégé connected to Lexis, has historically been priced above the firm’s research budget at this scale, though Thomson Reuters’ March 2026 partnership with Smokeball is the first announced move to bring integration-layer capabilities into the small-firm price band. The adjacent systems at the edges are consumer file storage, OneDrive, Dropbox, or a small-firm practice management tool such as Clio or Smokeball, and an eDiscovery vendor the lawyer calls only when a matter requires one.

The stack at this desk is still not the stack a large firm sees. But it is no longer accurate to say the upper layers are simply absent.

The stack at this desk is still not the stack a large firm sees. But it is no longer accurate to say the upper layers are simply absent.

The role

The solo lawyer handles every part of a matter. Intake. Conflict checks. Drafting. Citation. Filing. Calendar. Billing. No associate to delegate the brief to. No paralegal to pull the cases. No partner downstream to catch the citation that does not exist. The work is done at one desk, often after hours, and the same person who drafts the motion is the one whose signature appears on it. Small-firm lawyers operate on the same structural model with one or two colleagues, sometimes a part-time paralegal, often the same software stack.

Work surface

Microsoft 365 with Claude is built into Word, Excel, PowerPoint, and Outlook. Many solo and small-firm lawyers operate on Microsoft 365 Personal, Family, Business Basic, or another low-admin productivity setup. Claude for Small Business, announced on May 13, is not a legal product, but it is positioned for the kind of low-admin small-business environment many solo practices inhabit. Claude in Outlook entered public beta in the same window, which puts assistance directly inside the inbox, where most client communication already happens.

What this looks like on a working night. The brief in Word is a response to a motion to compel. The solo drafts the procedural background paragraph, then asks Claude to tighten it. Claude returns a tighter version. The solo edits, accepts. The next paragraph is the standard of review. The solo asks Claude to draft it. Claude generates three sentences and proposes a case citation for the controlling standard.

The solo stops. The citation looks plausible. The case name is plausible. The reporter citation is plausible. The solo opens a separate tab, goes to Google Scholar, and looks for the case. If the case exists and says what Claude said it says, the citation goes in the brief. If it does not, the citation is fabricated, and the brief almost went to the docket with a hallucination in it. This is the rhythm of drafting at this seat: continuous AI assistance, continuous verification, one verifier, one signer.

This is the rhythm of drafting at this seat: continuous AI assistance, continuous verification, one verifier, one signer.

Excel with Claude turns a damages framework into a working model with formula assistance. PowerPoint generates mediation visuals and client presentation drafts. Outlook beta lets the solo drop a long client email at nine at night into Claude for triage and a draft response. The unit cost of the work-surface tier is closer to consumer software than to enterprise legal tooling, which is what makes it reachable for this seat at all.

Workflow layer

Anthropic’s twelve practice-area legal plugins are generally not deployed for solo practices. Plugin governance assumes information technology administration: configured tenants, managed identity, and policy oversight. A one-attorney or two-attorney practice typically does not have that administration in place.

One thing the solo lawyer can do that most legal AI products have never permitted is look at the workflow layer directly. Anthropic published the legal plugin and connector configurations in a public repository called claude-for-legal. A solo lawyer who is technically curious can read the configuration logic for each plugin, identify which connectors each plugin assumes, and track how configurations change over time. Whether the solo actually uses the plugin layer or not, the visibility itself is a change from the closed legal AI products that came before.

The Lexis removal commit on May 12 modified 97 files in that repository in roughly a single afternoon. The lesson for the solo lawyer is not the specific commit; it is that the visible stack can change inside a launch window, and any configuration the solo invests time in today may shift under her tomorrow. A configuration that called Lexis on May 11 was configured to call something else by May 13. A solo considering whether to spend a weekend setting up connectors should know that the connectors themselves are a moving target.

Integration layer

Historically out of reach at this scale. CoCounsel Legal and Lexis+ with Protégé have been priced for firms with research budgets. Many solo lawyers carry a basic Westlaw or Lexis subscription, or no commercial research subscription at all. Free or low-cost legal research, Google Scholar, court records, and consumer-tier Claude or ChatGPT often fill the gap.

The picture is moving. Thomson Reuters and Smokeball announced a partnership on March 25, 2026, to integrate Smokeball’s practice management platform with Thomson Reuters CoCounsel Legal. The announcement language is comprehensive. Thomson Reuters wrote that the partnership is “giving legal professionals in small law firms something that simply hasn’t existed in the market until now.” Both companies described a combined experience that would unite Smokeball’s practice management with CoCounsel Legal’s authoritative legal research, document analysis, and drafting.

The initially deployed scope is narrower than the announcement. The first capability is a real-time data connector that syncs legal documents from Smokeball into CoCounsel Legal in bulk rather than one file at a time. Further phases are described in a roadmap toward matter data, Westlaw Advantage, and Practical Law inside a single interface, U.S.-only at initial scope, but those phases are not yet shipped.

The investigation pattern from Part 1, applied to Joel Hron’s clarification on CoCounsel Legal in Claude, applies here. The announcement is real; the operational scope is narrower; the working solo should verify what is actually deployed before paying for the integration uplift.

The procurement decision rests on understanding the differences among practice-management integration, document-sync integration, legal research integration, document-analysis capability, drafting support, and full content-and-workflow access. These are not the same thing. A solo evaluating the Smokeball partnership today is evaluating a document connector with a roadmap. A solo evaluating it twelve months from now may be evaluating something materially more comprehensive.

The directional shift matters at this seat regardless of where the deployment lands. For the first time, the integration layer is being deliberately routed toward the small-firm market through the practice-management software those lawyers already use.

The strategic move is not accidental. The solo and small-firm segment is the permanent unfinished business of the legal AI market. The vast majority of U.S. lawyers practice at firms with fewer than 10 attorneys, and enterprise-tier legal information products have historically been priced above what the segment can absorb.

The May 2026 architecture changes the calculus from three directions at once. Anthropic’s Claude for Small Business reaches the segment at consumer-software price points. Microsoft 365 with Claude lands the work surface inside the productivity tools, the segment already uses. Thomson Reuters’ Smokeball partnership routes the integration layer through the practice-management software that the segment already pays for. Three routes to the same seat, run by three vendors with different competitive bets.

A fourth front is already open inside the same chokepoint: Clio has embedded its own AI layer (Clio Duo) into Clio Manage and added Vincent, vLex’s AI research tool, into Clio Work. In April 2026, Clio went further and made Clio Work available as a standalone AI workspace for solo, small, and midsize firms, with no Clio Manage subscription required. Clio is not just defending its practice-management base; it is turning legal research and analysis into a separate entry point for the same small-firm market the others are pursuing.

The practice-management platform becomes the strategic chokepoint because it is the operating-system layer of the small firm’s working day. For the solo lawyer, the question is which route delivers actual value first, not which announcement language is loudest.

The practice-management platform becomes the strategic chokepoint, because it is the operating-system layer of the small firm’s working day.

For the smaller number of solos who do subscribe to CoCounsel Legal directly today and are considering the Claude integration announced May 12, the operational scope is narrower than the announcement language suggested. Joel Hron’s clarification via Reuters made clear that the integration does not replace CoCounsel Legal and does not provide standalone access to the full content-and-workflow system. The first capability exposed through the Claude connector is Westlaw Deep Research, surfaced through four tool calls: start research, check status, get report, and follow up. For drafting, file summarization, exhaustive search, citator work, claims explorer, and litigation analytics, the solo continues to use CoCounsel Legal directly or Westlaw natively.

The procurement decision rests on what is exposed, not on what the announcement described. A solo who reads “CoCounsel Legal now in Claude” and pays the integration uplift expecting the full system has paid for a specific Westlaw research workflow with four tool calls, not the full platform.

Adjacent systems

Document management at this scale is OneDrive, Google Drive, Dropbox, or a small-firm practice management system such as Clio or Smokeball. iManage and NetDocuments are not in this picture; those are firm-scale platforms used at midsize-to-large firms and inside larger corporate legal departments. eDiscovery, the discovery process applied to electronically stored information, is typically outsourced to a vendor on a per-case basis.

The broader pattern worth watching is the spread of MCP-style integrations across professional software environments. iManage’s May 14 server is one example. The May 12 NetDocuments MCP server is another. Smokeball’s Thomson Reuters partnership puts a small-firm practice management tool on the same trajectory; the small-firm practice management vendors the solo lawyer already uses, Clio chief among them, may add equivalent integrations over the next twelve months, which would change what Claude can do inside the systems the solo already pays for.

What this role might do differently now

The brief above the solo’s hands on a Tuesday night is faster to produce than it was a year ago. The standard-of-review section, the procedural background, the introduction, and the discussion of the standard exhibit. All are reachable inside Word with Claude's assistance and a measure of the lawyer’s time the solo did not previously have. The damages model in Excel takes only twenty minutes, where it used to take ninety. The client presentation in PowerPoint comes together in an evening rather than a weekend. Triage of the Outlook inbox is more manageable when Claude can draft routine responses for the lawyer to edit and send.

The discipline that does not move is verification. Every citation Claude proposes has to be checked against the actual reporter before the brief is filed. Every factual claim has to be traced. The faster the drafting goes, the more important the discipline becomes. The solo who treats Claude’s output as ready-to-file is exposing the practice to sanctions; the solo who treats Claude’s output as a first draft that needs the same verification any first draft needs is operating safely.

Configure connectors and plugins with caution, on the understanding that what is configured today may shift, and that announcement language about what an integration does often exceeds what is actually deployed at this scale.

What this role should know

Verification is the central functional concern at this seat. Without a live legal-research connector or another verified retrieval source beneath it, Claude within Word should be treated as drafting based on model output rather than grounded in legal authority. Every cited case in an AI-assisted draft must be read by the actual reporter before the brief is filed. The lawyer is the only verifier in the chain. The lawyer is also the signer, which means that Federal Rule of Civil Procedure 11 applies directly here.

The verification reality at this seat is structurally harder than at firms with more readers. American Federation of Government Employees v. Trump, the May 6 motion to quash filed in the Northern District of California, was drafted in part by Claude Console and contained phantom quotations. That motion went through a firm with associates, paralegals, and a partner review before it reached the docket. The phantom quotations got through anyway. At a one-attorney practice, there is no other reader. Every error that reaches the docket reaches it without a second pair of eyes.

Integration-layer access does not eliminate the need for verification. Sullivan & Cromwell, one of the most prestigious firms in the country, acknowledged to a federal bankruptcy judge in April that a recent filing contained AI hallucinations flagged by opposing counsel. The solo lawyer is not behind for not having CoCounsel Legal or Lexis+ with Protégé. The integration layer is faster and more reliable than ungrounded drafting, but it does not remove the obligation to read every cited case in full before filing.

The financial reality at this seat is concrete. One of the largest AI hallucination sanctions in U.S. legal history landed on a small-practice context. In Couvrette v. Wisnovsky, U.S. Magistrate Judge Mark D. Clarke of the District of Oregon issued two orders, on December 12, 2025 and March 23, 2026, sanctioning San Diego attorney Stephen Brigandi a total of approximately $96,000, with combined penalties to him and co-counsel Tim Murphy totaling $110,204.38, after three summary-judgment briefs contained 15 hallucinated case citations and 8 fabricated quotations falsely attributed to legitimate authorities. The court struck the briefs without leave to refile and dismissed the plaintiffs' claims with prejudice.

The court found that the lead plaintiff, a self-represented serial litigator, likely generated the AI-tainted briefs herself and that the lawyers signed them without verification. The verification chain broke at the signature line, regardless of who typed the prompt. The solo lawyer signing AI-assisted work without independent verification is signing onto that risk profile.

The next twelve months

Claude for Small Business and adjacent productivity offerings will continue widening access to the work-surface tier at lower price points. The integration layer is, for the first time, beginning to reach toward this seat through the Thomson Reuters and Smokeball partnership; the operational scope and pricing of that reach are the open questions worth watching over the next two quarters.

The verification burden on this role will not decrease as the work surface becomes easier to use and as integration-layer access becomes more accessible. State mandatory technology continuing legal education requirements are spreading, and several jurisdictions have already adopted them. The architecture is landing inside Word at the solo seat. The work of verifying what comes out of it remains the work it always was.

Role 2: The paralegal or legal assistant

On a Tuesday afternoon at a midsize-to-large firm, the paralegal is among the most active operational users of the building's architecture. The work surface is open: Word with Claude on the left of the screen, Outlook with Claude in the corner, Excel with Claude running a damages model in another tab. The workflow layer is live: the Anthropic plugin the firm deployed for this practice area runs beneath the drafting environment, encoding the firm’s standard sequence for this kind of matter. The integration layer is connected: a CoCounsel Legal Westlaw Deep Research report is returning grounded citations that the paralegal weaves into the draft, or Lexis+ with Protégé is doing the equivalent work on the Lexis side. The adjacent-system edges are wired in: the iManage MCP server is searching firm precedent in the background without requiring the paralegal to leave the drafting environment.

All three layers of the stack operate simultaneously in ways most attorneys do not see, because attorneys do not run the workflows. Attorneys review the output. The paralegal seat is where the May 7 to May 13 architecture often operates at volume.

The role

The paralegal at a midsize-to-large firm wears more hats than almost any other seat in the building. Citation pulls and document assembly. Client intake and screening calls. Court calendaring, filing logistics, service of process, and procedural interactions with the clerk’s office. Exhibit preparation, trial binders, and the physical and digital infrastructure of a matter as it moves toward a hearing. Vendor coordination across court reporters, copy services, expert witnesses, and process servers. Discovery coordination, including collecting documents from clients, organizing productions, and managing the matter’s eDiscovery workspace. Subpoena preparation and tracking. Billing entries, sometimes for attorneys who hate doing their own. Junior paralegal supervision and training. And, increasingly, first-draft research and first-draft drafting that the attorney reviews and signs.

On many matters, the paralegal is the most consistent person across the life of the case. The attorney rotates among matters. The paralegal lives inside one matter for months and knows it cold.

The paralegal does not sign filings; the supervising attorney does. The paralegal’s working day, however, is often where the firm’s actual AI workflow lives in practice.

Work surface

Same Microsoft 365 environment as the supervising attorney. Claude in Word for drafts, exhibits, and memos. Claude in Excel for damages models, witness lists, document logs, and matter budgets. Claude in Outlook for calendar coordination, drafts of client communication that the supervising attorney reviews before sending, and routing the dozens of small administrative emails a matter produces in a week. Paralegals are often the earliest daily users in the firm because the workflow runs through their seat first; they are also often the first to discover what works and what does not.

What this looks like on a working day. The partner forwards a case assignment at nine. By nine-thirty, the paralegal has opened iManage and run a search for related prior matters; the search runs through the firm’s MCP server, returns a list of relevant precedents, and Claude, inside iManage, summarizes the most relevant memos without the paralegal leaving the environment. By ten, the paralegal is in Word with Claude drafting the procedural background section, pulling citations from CoCounsel Legal’s Westlaw Deep Research report running in the connector pane. By eleven, the draft is in shape to circulate. By eleven-thirty, the paralegal has flagged in the document which sections were drafted with Claude's assistance, which citations came from CoCounsel, and which paragraphs are original work; the attorney’s verification step then knows where to focus.

The work surface has changed the rhythm of the morning. The verification chain still depends on the paralegal making AI use legible up the chain.

Workflow layer

Where firms have deployed Anthropic’s twelve practice-area plugins, paralegals are often the daily users, especially on litigation and contracts work. The plugin layer is where firm-specific workflows live, and paralegals tend to be the people who actually run them.

The workflow layer is publicly inspectable for the first time. Anthropic’s Claude-for-legal repository exposes the configuration logic for each plugin and each connector. A paralegal at a firm where information technology or knowledge management is deploying a plugin may be able to look at the plugin’s configuration directly, where firm access permits, see which connectors the plugin assumes, see the default behavior, and track configuration changes through the repository’s commit history. Most legal AI products have not exposed their configuration logic this way. The visibility is part of the architecture now.

The Lexis removal commit on May 12 is one demonstration that the configurations change in real time, sometimes in ways the working user needs to track. A plugin configured to call Lexis on May 11 was configured to call something else on May 13. A paralegal running a workflow today should understand that the workflow’s connector behavior may have shifted since the documentation was last read.

Integration layer

Where the firm has CoCounsel Legal or Lexis+ with Protégé, the paralegal typically has the same access as the attorneys. Drafting with grounded retrieval, in which the AI system searches Westlaw or Lexis before generating, is meaningfully different from drafting from general training data. Paralegals supporting brief preparation are often the first people in the firm to run that workflow at volume.

The two integration-layer surfaces are not equally documented. Thomson Reuters published a public CoCounsel Legal MCP Connector Guide that describes exactly which tool calls are exposed through the Claude integration for Westlaw Deep Research: start research, check status, get report, follow up. For drafting, file summarization, exhaustive search, claims explorer, and litigation analytics, the guide directs the paralegal back to CoCounsel directly or to Westlaw natively.

LexisNexis announced its Claude integration into Lexis+ with Protégé one day after Thomson Reuters did, but the publicly visible operational scope of the LexisNexis integration is narrower. A paralegal running workflows through both integrations operates with less ability to anticipate the behavior’s side. The CoCounsel side has a documented surface to plan against. On the Lexis side, until more documentation appears, there are fewer documented capabilities to anticipate. The working assumption should be that the Lexis side will surprise more often than the Thomson Reuters side.

Adjacent systems

Every firm-scale seat lives inside the document management system, often called the DMS. Associates, partners, litigation support specialists, and the paralegal all touch the DMS daily. The paralegal touches it most often, at the highest volume, and across the widest range of tasks: pulling prior memos, locating motion templates, retrieving deposition outlines, surfacing related matters, organizing exhibits, and assembling productions.

The practical shift in May 2026 is that the document management system itself is becoming an AI-reachable working context rather than a passive repository. iManage announced its MCP server on May 14. NetDocuments published its own MCP collaboration with Anthropic on May 20. The shift lands at the paralegal seat first because the paralegal is the heaviest daily user.

Before these MCP-style DMS integrations, the paralegal searched iManage or NetDocuments manually for the morning’s research needs, opened Claude separately in another window, copied or retyped relevant content into Claude, asked Claude to summarize or draft, then took the result back into Word or the matter folder. Three or four context switches per task. The MCP server collapses those steps. Claude can now search firm work product, summarize prior memos, and surface related matters within the DMS environment via the same protocol it uses to access external systems.

The rhythm of the morning changes. The number of tabs open at once drops. The cognitive overhead of moving content between systems drops.

The two announcements together mean that the firm-side knowledge layer at most large and midsize firms in the United States is now reachable from Claude through the same protocol. The architectural bets differ. iManage is positioning its MCP server as a vendor-neutral substrate that any AI tool can connect through, including Harvey, Legora, ChatGPT, Claude, Microsoft Copilot, and proprietary firm tools. NetDocuments is exposing its content graph through MCP alongside its existing first-party AI surface, in a configuration that competes more directly with the third-party AI tools at the edges.

The paralegal at a firm running iManage and the paralegal at a firm running NetDocuments are both now working within MCP-enabled DMS environments, but what each surface actually exposes through the protocol and how each governs ethical walls and permissions within the AI-accessible layer may differ in ways that affect the working day. The paralegal evaluating which DMS workflows to deploy with Claude should test rather than assume.

The eDiscovery side is moving in parallel. Relativity, Logikull, DISCO, and proprietary review platforms are adding Claude integrations and newer generative AI features alongside their established workflows. The paralegal supporting first-pass review work is often the first person in the firm to encounter the changes and evaluate whether the new features should be used on a given matter. Contract management platforms such as Ironclad and ContractPodAi are adding AI features along similar lines, though those are more often used inside corporate legal departments than at outside firms.

What this role might do differently now

Pull and assemble citations through CoCounsel Legal or Lexis+ with Protégé directly inside Word, without leaving the drafting environment. Use the iManage MCP server to surface relevant firm precedent when starting a new matter or brief, without context switching out of iManage. Use Claude in Excel for matter tracking, document logs, witness lists, and budget management with formula assistance. Use Claude in Outlook for calendar coordination and drafting client communications for the supervising attorney to review and send. Run practice-area plugins where the firm has deployed them.

Flag AI use explicitly when handing work up the chain. The simplest practice is to annotate the draft directly: which paragraphs were drafted with Claude's assistance, which citations came from CoCounsel or Lexis+ with Protégé, and which sentences are first drafts versus reviewed text. The attorney’s verification step then provides a focus, rather than treating the entire document as equally suspect or equally trusted.

At the ceiling of what is possible at this seat: build custom workflows on the Anthropic API. Mark Pike, Anthropic’s Associate General Counsel, described one such case at the May 12 launch interview with Artificial Lawyer. A paralegal on a four-person pro bono team facing an AmLaw 200 firm in an elder abuse trial built a tool on the Anthropic API that sat at counsel’s table during trial, pulling in cross-examination lines as the trial unfolded and helping counsel react in real time. The team walked out with a large jury verdict.

Not every paralegal will build API-based real-time trial tools. The working principle holds at every scale: where firm policy permits, the paralegal is often the person who actually deploys AI workflows, while the supervising attorney remains the verifier and signer.

The paralegal is often the person who actually deploys AI workflows, while the supervising attorney remains the verifier and signer.

What this role should know

Rule 5.3 was written for an era of human nonlawyer assistance. AI-assisted work falls inside the same supervisory obligation, but the verification surface is now larger and faster-moving than the rule contemplates. The paralegal is not the signer on a court filing; the supervising attorney is, and the supervising attorney’s responsibility for verification does not transfer downward.

The paralegal’s structural leverage is making AI use visible up the chain. Flagging which content came from Claude, which citations came from CoCounsel Legal or Lexis+ with Protégé, and which sentences are first drafts versus reviewed text creates a record and gives the attorney’s verification step a place to focus. Flagging AI use is becoming a protective practice consistent with Rule 5.3 supervision obligations.

The harder situation is the one the rule does not address. The paralegal flagged everything correctly, and the attorney signed without verifying the citations. The brief was filed with the docket with unverified content. The paralegal did everything right, and the chain still broke. The paralegal cannot force the attorney to verify. The available moves are firm-level: raise the pattern directly with the supervising attorney, escalate to a managing attorney if the pattern continues, and document the flagging practice in the matter file. At firms with general counsel or risk management functions, those are also paths. The article cannot tell any paralegal which escalation move is right at their firm. It can acknowledge that doing everything right in the paralegal seat is sometimes not enough on its own, and that firm-level governance is where the residual risk lies.

The AFGE v. Trump filing is the working example. That motion was drafted in part by Claude Console at a firm with paralegals, associates, and partner review. The phantom quotations got through anyway.

The lesson for the paralegal seat is that the verification chain only works if AI use is visible to the verifier. If the attorney does not know which parts of the draft came from AI output, the attorney cannot direct the verification effort accurately. The paralegal who makes AI use legible up the chain protects both seats.

Integration-layer access does not eliminate hallucinations. Grounded retrieval through CoCounsel Legal or Lexis+ with Protégé reduces the likelihood that a fabricated citation reaches the draft, but it does not remove the verification step. Sullivan & Cromwell’s bankruptcy admission shows that even firms with full integration-layer access produce filings that opposing counsel can flag for AI hallucinations. The grounded tier is better than the ungrounded tier. Neither tier ends the verification work.

Announcement language exceeds the deployed scope. The paralegal asked to run a workflow through the CoCounsel Legal integration in Claude is running Westlaw Deep Research, specifically, exposed through four tool calls, not the full CoCounsel system.

The May 12 Lexis removal commit is a separate reminder that what is wired into a plugin today can be wired out tomorrow. The architecture is still stabilizing. The paralegal working inside it should plan workflows around what is currently deployed, not around what the announcements described.

The next twelve months

More paralegal-originated AI work is flowing up to attorneys. More iManage, NetDocuments, and similar adjacent-system integrations are putting more AI inside the platforms paralegals already use daily. More vendor documentation of the operational scope as the integration layer matures. More firm-level policy on how to flag AI use up the chain.

Looking ahead

The next sub-piece covers the three roles where integration-layer access is often standard, the work-surface tier is taken for granted, and the supervisory chain runs through multiple seats: the litigation associate at a midsize-to-large firm, the litigation partner who signs the filing, and the in-house counsel who advises a corporation on AI policy for the entire workforce. The architecture is the same. The exposure shifts.

Sources

• Anthropic, “Collaborate with Claude across Excel, PowerPoint, Word and Outlook,” May 7, 2026

• Anthropic, “Claude for the legal industry,” May 12, 2026

• Anthropic, “Introducing Claude for Small Business,” May 13, 2026

• Anthropic claude-for-legal GitHub repository, including commit 210fd7a, “Remove Lexis references at partner request”, May 12, 2026

• Reuters, “Anthropic expands Claude’s AI tools for law firms, lawyers,” May 12, 2026

• Thomson Reuters, CoCounsel Legal MCP Connector Guide

• Above the Law, “Finally, Something For Small Law Firms,” covering the Thomson Reuters and Smokeball partnership, May 2026

• LawSites/LawNext, “Exclusive: Smokeball and Thomson Reuters Partner to Integrate CoCounsel Legal AI with Practice Management Platform,” March 25, 2026

• Thomson Reuters Institute, “A Partnership to Build the Premier Legal Technology Ecosystem,” March 25, 2026

• Smokeball, “Smokeball and Thomson Reuters Partner to Create the Premier Legal Technology Ecosystem,” March 25, 2026

• LexisNexis, “LexisNexis Expands Lexis+ with Protégé by Integrating Anthropic’s Claude Legal Plugin Suite,” May 13, 2026

• iManage, “iManage MCP Server is now Available to Connect Governed Knowledge to the Broader AI Ecosystem,” May 14, 2026

• NetDocuments, “NetDocuments Collaborates with Anthropic on Legal Industry MCP,” May 20, 2026, by Manish Rai, VP Product Marketing

• ABA Journal, “Federal judge hands down $110K penalty against 2 lawyers for AI errors in court documents,” April 17, 2026, covering the Brigandi sanctions order

• Bloomberg Law, “Lawyers Apologize for Fake AI Quotes in Trump Mass Layoffs Case,” May 2026

• Fortune, “Even as hallucinations show up in legal filings, Big Law goes all in on AI with new Anthropic release,” May 12, 2026

• Artificial Lawyer, “Claude For Legal Launches, May Reshape the Legal Tech World,” May 12, 2026, including Mark Pike interview

• LawNext, “Clio Work, Clio’s AI Workspace, Is Now Available To Solo and Smaller Law Firms As A Standalone Product,” April 21, 2026

• American Bar Association, Model Rule 5.3, Responsibilities Regarding Nonlawyer Assistance

• Federal Rule of Civil Procedure 11

A Note on Methodology: Every dated event, named participant, and quoted statement in this piece traces to a primary source, a public repository record, a company announcement, a press release, a court filing, or verified secondary reporting that quotes or identifies a primary actor. The role profiles walked through above are constructed from the architecture documented in Part 1 of this series and from the working-environment patterns observed at each seat. Where the architecture has not yet been deployed at a given seat, that is identified directly. Where public announcement language differs from documented implementation scope, this article separates the broader framing from the narrower operational claim.

AI tools used in preparation. This article was drafted and revised with assistance from Claude and ChatGPT. Factual claims were checked against primary sources using Perplexity, ChatGPT, and DeepSeek. Copy edit was performed with Gemini. Editorial judgments, framing, and analytical conclusions are the author’s.

This article is for informational and educational purposes only. Nothing herein constitutes legal advice, and reading this article does not create an attorney-client relationship. Readers should consult qualified legal counsel for advice specific to their situation.

Copyright © 2026 Robert Meyers-Lussier. All rights reserved.

Corrections. Spot a factual error, misquote, mischaracterization, broken link, or anything else that needs correcting? Please email allsourcelegalai@substack.com. Confirmed corrections will be posted within 72 hours, with a note indicating what changed and when.

All-Source Verification Tracker

Documented AI hallucination cases in court filings, worldwide: 1,459 United States cases: 1,008 Named legal AI products appearing in documented cases: ChatGPT, Claude, Gemini, Westlaw AI, Lexis+ AI, MX2.law (most database entries record the tool as “Implied”)

Source: Damien Charlotin’s AI Hallucination Cases Database, last updated May 22, 2026.

Between May 7 and May 13, a cluster of legal AI announcements landed in rapid succession. Anthropic expanded Claude into Microsoft 365, launched legal plugins and connectors, announced a major CoCounsel Legal integration with Thomson Reuters, and then appeared inside Lexis+ with Protégé through LexisNexis.

Most coverage treated these as separate stories. Read separately, they sounded like overlapping product launches competing for attention inside an already crowded legal AI market; taken together, they reveal something larger.

The announcements are best understood not as isolated launches but as layers of a single emerging legal AI stack. Microsoft 365 became the work surface. Claude legal plugins became the workflow layer. Connectors became the pathways into outside systems. Thomson Reuters and LexisNexis became vendor-controlled research environments attached to that stack.

Once those layers are separated, the week becomes much easier to understand. Part of the confusion surrounding the May announcements is that companies were often describing different layers of the stack at the same time. One sentence might describe a legal plugin. The next might describe a research environment behind the plugin. Another might describe a capability available only inside a separate vendor-controlled system. The result was a wave of headlines that sounded larger, smaller, or more complete than the underlying architecture actually was.

This article explains what was actually built. Not what vendors promised, not what some readers expected, not what the market speculated might happen later. This article is about what was actually launched.

The Microsoft 365 Layer

The Work Surface

The May legal AI launch did not begin in legal software. It began in productivity software.

On May 7, Anthropic announced Claude for Microsoft 365. Claude for Excel, PowerPoint, and Word became generally available on paid plans. Claude for Outlook entered public beta. The announcement also described cross-application context and deployment through Microsoft AppSource and the Microsoft admin center.

The significance is not that Claude can appear inside Office applications. The significance is where Claude was now positioned. Lawyers draft in Word. Legal operations teams work in spreadsheets. Partners review decks in PowerPoint. In-house counsel manage communications in Outlook. The same AI layer later attached to legal systems was already being positioned inside the everyday environment where professional work begins.

For years, the legal AI conversation revolved around standalone tools. Legal professionals imagined opening a separate AI product to perform a separate AI task. The underlying assumption was that AI would sit beside existing workflows rather than inside them.

The Microsoft 365 announcement pointed toward a different model. Instead of positioning AI as a separate destination, Anthropic was beginning to position Claude as part of the work surface itself. Software environments shape behavior. A standalone tool competes for attention. A work surface becomes part of routine professional activity.

A standalone tool competes for attention. A work surface becomes part of routine professional activity.

By the time the legal launch arrived on May 12, Claude was no longer being framed primarily as a chatbot sitting outside legal work. It was being framed as part of the environment where drafting, reviewing, analysis, communication, and coordination already happen.

The Legal Plugins and Connectors

The Workflow Layer

If Microsoft 365 became the work surface, the legal plugins and connectors became the workflow layer sitting on top of it. This is where Anthropic’s legal push became concrete.

Anthropic’s May 12 legal-industry announcement described more than 20 new MCP connectors and 12 legal plugins tailored to practice areas and roles. Reuters reported that the release added access within Claude to legal platforms including Thomson Reuters, LexisNexis, Harvey, Box, Everlaw, and DocuSign.

A legal plugin is a preconfigured legal workflow for Claude. It bundles assumptions about a practice area, the systems that matter in it, the sources likely to be useful, and the outputs the work typically produces. A litigation, employment, privacy, or regulatory plugin does not simply rename a general-purpose AI model for a legal audience. It organizes a working environment around a specific kind of legal task.

The connectors are what make those environments operational. If plugins define how Claude behaves, connectors define what Claude can reach. They are the access pathways between Claude and outside systems: collaboration tools, document repositories, public legal resources, and vendor-controlled legal information platforms. The connectors reveal the real architecture of the stack more clearly than the marketing language does. They show which systems Claude is being wired into, and which systems remain outside the default configuration.

This is the layer where the architecture becomes easier to see.

Claude is not becoming a lawyer. Claude is being wired into legal information systems. The plugins are not legal expertise in abstract form. They are structured configurations connecting an AI reasoning system to workflows, repositories, communication environments, and legal information sources.

Claude is not becoming a lawyer. Claude is being wired into legal information systems.

The open-source structure of the repository made that shift more important. Much of the legal AI industry has historically operated behind closed systems. Vendors announced capabilities while the underlying workflow logic remained opaque. With Anthropic’s public claude-for-legal repository, portions of the legal workflow layer became publicly inspectable infrastructure. Firms, developers, and observers can inspect configuration logic, trace architectural changes, and see how some legal workflows are being represented as code.

That visibility became especially important because the launch-week repository did not remain static.

Watching the Stack Change in Real Time

One of the most revealing moments of the launch week arrived through a GitHub commit. On May 12, a commit titled “Remove Lexis references at partner request” modified 97 files and removed Lexis+ Protégé references from default plugin configurations across the legal plugins. Repository documentation changed alongside the configuration files. The current raw connector documentation no longer lists Lexis among shipped default connectors.

The timing mattered. A firm examining the repository earlier in the day would have encountered one version of the legal stack. A firm examining it later would have encountered another. The launch was not a static product release. It was an actively shifting architecture.

This should not be framed as scandal. The commit showed how quickly the visible configuration of the legal AI stack could change during the launch itself. It also showed why announcement language, repository state, configuration defaults, and operational deployment should not be treated as identical.

Even after the removal commit landed, the broader repository structure illustrated how quickly configuration state and surrounding documentation can diverge in fast-moving deployments. That kind of drift is not unusual in rapidly evolving software environments, but it matters in legal AI because legal professionals, firms, and vendors may rely on very specific claims about what a system can access and what it is actually configured to do.

The Lexis removal is more than a repository footnote. It is a live example of the larger point. The legal AI stack visible during launch week was already evolving while the public was still trying to understand what had launched.

Thomson Reuters and CoCounsel Legal

The Vendor Bridge

If the plugins and connectors formed the workflow layer, the Thomson Reuters integration demonstrated how that layer could connect to a major legal research environment. On May 12, Thomson Reuters announced that the next generation of CoCounsel Legal had been rebuilt on Anthropic’s Claude Agent SDK. The same day, Reuters reported Anthropic’s direct integration involving CoCounsel Legal.

Many readers interpreted the announcement as implying something close to full CoCounsel functionality operating directly through Claude. The public record now supports a narrower day-one reading. Reuters reported that Thomson Reuters CTO Joel Hron clarified that the Claude integration did not replace CoCounsel Legal or provide standalone access to the full underlying content-and-workflow system. Thomson Reuters’ public CoCounsel Legal MCP connector guide later made the operational scope even clearer.

The first publicly documented capability exposed through the Claude connector was Westlaw-backed deep legal research. The connector guide described four exposed operations for that workflow: starting deep legal research, checking its status, retrieving a report, and running a follow-up against an existing conversation. The same guide said users should continue using CoCounsel directly for other CoCounsel Legal capabilities, including drafting documents and summarizing files.

The difference was not contradiction. It was scope. The announcement described the environment behind the integration; the connector guide described what Claude could actually do inside it.

That gap between environment and exposed functionality became one of the defining themes of the launch week. The CoCounsel integration was real. It was also more bounded than many readers initially understood.

The important shift was not that Claude suddenly absorbed all of CoCounsel. It was that a major legal research environment visibly attached itself to the emerging AI stack. The bridge mattered even where the bridge was narrower than the headlines implied.

Lexis+ With Protégé

The Second Vendor Environment

The LexisNexis announcement arrived one day later. On May 13, LexisNexis announced that it had integrated Anthropic’s Claude legal plugin suite into Lexis+ with Protégé. LexisNexis described the integration as grounded in its legal content repository and operating inside Lexis+ with Protégé’s private, secure technology environment.

The structure looked familiar: a major legal information environment, a conversational interface layer, Anthropic integration, and a repositioning around Claude.

But the Lexis announcement differed from the Thomson Reuters announcement in one important respect: much less was publicly visible about the operational scope. Compared to Thomson Reuters’ CoCounsel connector guide, reviewed public Lexis materials did not expose an equivalent tool-by-tool MCP connector surface.

That uncertainty itself became part of the story. Different vendors were converging toward the same architectural direction at different speeds and with different levels of transparency. The Thomson Reuters announcement became clearer because public documentation narrowed the exposed connector surface. The LexisNexis announcement remained more compressed. The technical scope was harder to evaluate from public materials alone.

The Lexis story also carried additional significance because Lexis references had existed inside the Anthropic repository before their rapid removal on May 12. The following day, LexisNexis announced its own integration through its own environment. The timing makes them difficult to read in isolation, though the more important point is structural.

By May 13, the legal information market no longer looked like a set of isolated AI experiments. It looked like multiple major legal information vendors attempting to establish their position inside the same emerging architecture. For decades, legal information systems largely operated as separate destinations. Westlaw was Westlaw. Lexis was Lexis. Document systems, communication systems, and productivity software lived in separate environments.

The May announcements suggested the beginning of a different model. Not one unified legal AI platform, but a connected stack.

One Week Later

The Pattern Continued

A week after the original announcements, the broader pattern had not weakened. It had intensified.

New hallucination-related legal incidents continued surfacing almost immediately. Reuters and Bloomberg Law reported that Jason Greaves of Binnall Law Group apologized to Judge Susan Illston in the Northern District of California after AI-generated phantom quotations appeared in a May 6 motion to quash a subpoena filed in AFGE v. Trump, the litigation involving the Trump administration’s mass layoffs of federal employees. Public court filings and reporting identified Anthropic’s Claude Console as part of the workflow used to generate the draft filing.

The incident connects directly to the architecture described throughout this article. As the legal AI stack became more operationally integrated into drafting and research environments, the verification problem became more operationally urgent. The issue is no longer hypothetical future misuse. Legal professionals are already working inside these systems while governance, review discipline, and institutional safeguards remain uneven.

The incident does not show that Anthropic’s May 12 legal launch caused the filing problem. The filing predated the launch by six days. The chronology supports a narrower and more important point: operational adoption is moving faster than verification discipline in real legal workflows.

Anthropic continued moving deeper into infrastructure rather than retreating back toward standalone chatbot positioning. On May 13, Anthropic announced Claude for Small Business, extending Claude into business platforms through ready-to-run workflows. On May 18, Anthropic announced the acquisition of Stainless, a company focused on SDK and MCP server tooling. Anthropic framed the move around the idea that agents are only as capable as the systems they can reach.

Also on May 14, iManage announced its MCP server, a standardized open-protocol connection that lets AI systems, including Harvey, Legora, ChatGPT, Claude, and Microsoft Copilot, reach governed iManage content without custom integrations or bulk data exports. iManage is one of the most widely deployed document management systems in law firms. The MCP server release means that the firm-side knowledge layer (where matter files, client documents, work product, and historical drafts actually live) is now directly addressable through the same MCP protocol that Claude uses to reach external systems. That extension matters because most firm AI work eventually has to touch firm-controlled documents, not just public sources. The iManage move closes a layer of the stack that Anthropic's own connector layer does not cover by default.

The story is no longer simply about a chatbot competing with other chatbots. It is increasingly about who controls the workflow layer, the integration layer, and the surrounding operational environment.

The same pattern became visible outside the legal sector. Anthropic announced a global alliance with KPMG integrating Claude into KPMG’s Digital Gateway platform and making Claude available across KPMG’s workforce. Anthropic announced an expanded PwC partnership involving Claude Code and Claude for client work. EPAM announced a multi-year Anthropic partnership focused on enterprise AI delivery.

The architecture emerging in legal practice may actually be part of a broader professional-services transition. Law firms, accounting firms, advisory practices, and enterprise operations are converging around the same underlying AI infrastructure questions: workflow integration, system access, verification, governance, and operational control.

The week after the launch strengthened rather than weakened the article’s central argument. The stack did not freeze after May 13. It continued consolidating.

For a firm trying to understand these announcements in real time, the practical problem was not abstract. A partner, knowledge-management lead, or litigation-support team could read the launch coverage and think a full research platform had moved into Claude, then discover that the actual exposed capability was narrower, vendor-specific, or still changing in configuration. That is the real operational shift: legal AI is no longer just something lawyers test. It is becoming something firms must interpret, govern, and verify as part of the working environment.

What All of This Adds Up To

The Stack Becomes Visible

Before May 2026, legal AI was discussed as a collection of isolated tools. One company built a chatbot. Another built a drafting assistant. Another built a legal research feature. Another announced document summarization. The industry conversation often treated these developments as parallel experiments.

The May announcements changed that. They made the stack visible. What appeared that week was not a collection of AI products. It was a work surface, a workflow layer, an access layer, and vendor-controlled research environments beginning to attach themselves to the same underlying architecture.

The market reaction reflected a broader concern that AI integration pressure might move upward into traditional software and legal-information environments. Earlier Anthropic legal-plugin announcements had already triggered Reuters-reported selloffs across segments of the software and services sector. That market history is useful context, not proof that investors correctly priced the May legal-AI announcements in real time.

The May announcements also exposed uneven transparency across the emerging stack. Some operational boundaries became clearer only after Reuters clarifications and connector documentation. Some repository structures changed in real time. Some vendor capabilities were described broadly while remaining narrower operationally. None of that necessarily means the announcements were misleading. It means the architecture is still stabilizing.

And the stabilization process is now happening in public.

From Tool to Environment

The legal AI conversation has often revolved around tools: which chatbot is better, which drafting system is faster, which research product is more accurate, which vendor will dominate. Those questions still matter. The May announcements pointed toward something larger.

Legal professionals may eventually stop experiencing AI primarily as a separate tool. They may instead experience it as part of the environment where professional work already happens: drafting, research, communication, analysis, and workflow coordination.

What launched between May 7 and May 13 was not one product. It was the first clear view of a legal AI environment taking shape across productivity software, workflow systems, research platforms, and enterprise infrastructure. The stack had been forming for months. That week, it became visible.

The stack had been forming for months. That week, it became visible.

Sources

• Anthropic, “Collaborate with Claude across Excel, PowerPoint, Word and Outlook,” May 7, 2026

• Anthropic, “Claude for the legal industry,” May 12, 2026

• Reuters, “Anthropic expands Claude’s AI tools for law firms, lawyers,” May 12, 2026

• Thomson Reuters, “Thomson Reuters and Anthropic Expand Partnership to Connect Claude with CoCounsel Legal,” May 12, 2026

• LexisNexis, “LexisNexis Expands Lexis+ with Protégé by Integrating Anthropic’s Claude Legal Plugin Suite,” May 13, 2026

• Anthropic claude-for-legal GitHub repository, including commit 210fd7a, “Remove Lexis references at partner request,” May 12, 2026

• Thomson Reuters, CoCounsel Legal MCP Connector Guide

• Reuters, “Lawyer apologizes for ‘phantom’ AI quotes in Trump layoffs case,” May 18, 2026

• Bloomberg Law, “Lawyers Apologize for Fake AI Quotes in Trump Mass Layoffs Case,” May 2026

• CourtListener docket, American Federation of Government Employees, AFL-CIO v. Trump, No. 3:25-cv-03698 (N.D. Cal.)

• Anthropic, “Claude for Small Business,” May 13, 2026

• Anthropic, “Anthropic acquires Stainless,” May 18, 2026

• Anthropic and KPMG global alliance announcement, May 2026

• Anthropic and PwC expanded partnership announcement, May 2026

• iManage, “iManage MCP Server is now Available to Connect Governed Knowledge to the Broader AI Ecosystem,” May 14, 2026

• EPAM and Anthropic strategic partnership announcement, May 2026

• Reuters market coverage on Anthropic legal plug-ins and software-sector selloffs, February 2026

All-Source Verification Tracker

Documented AI hallucination cases in court filings, worldwide: 1,459 United States cases: 1,008 Named legal AI products appearing in documented cases: ChatGPT, Claude, Gemini, Westlaw AI, Lexis+ AI, MX2.law (most database entries record the tool as “Implied”)

Source: Damien Charlotin’s AI Hallucination Cases Database, last updated May 22, 2026.

Every increment of the counter is documented evidence that current adoption training is not closing the gap. Each new case is another working lawyer who used an AI product in a court filing, failed to catch a fabricated citation or a misrepresented authority, and faced sanctions, disqualification, or referral as a result. Adoption training teaches lawyers how to prompt and how to integrate. Verification training teaches lawyers how to test AI output for the failure patterns the documented record now describes in detail. The stack walked through in this article puts AI deeper into legal work. The tracker shows what is happening at the practice end of that stack today.

A Note on Methodology: Every dated event, named participant, and quoted statement in this piece traces to a primary source, a public repository record, a company announcement, a press release, a court filing, or verified secondary reporting that quotes or identifies a primary actor. Where public announcement language differs from documented implementation scope, this article separates the broader framing from the narrower operational claim.

AI tools used in preparation. This article was drafted and revised with assistance from Claude and ChatGPT. Factual claims were checked against primary sources using Perplexity, ChatGPT, and DeepSeek. Copy edit was performed with Gemini. Editorial judgments, framing, and analytical conclusions are the author’s.

This article is for informational and educational purposes only. Nothing herein constitutes legal advice, and reading this article does not create an attorney-client relationship. Readers should consult qualified legal counsel for advice specific to their situation.

Copyright © 2026 Robert Meyers-Lussier. All rights reserved.

If you work in legal practice and have not yet changed how you think about AI, the market may already have changed around you. What happened between May 12 and May 13 was not just another legal tech launch. It was a public demonstration that one AI provider can now sit inside Microsoft 365, plug into legal workflows, and appear at the center of products tied to both Thomson Reuters and LexisNexis.

That does not mean every lawyer, paralegal, legal operations professional, or court-facing worker needs to panic. It does mean the old assumption, that legal AI would arrive slowly enough for the profession to watch first and adapt later, looks much weaker than it did a week ago.

The change many expected to arrive over several years appears to have arrived, at least structurally, in a matter of days.

For the largest law firms, legal AI has already been a working concern for years. Many firms in the Am Law 100 and 200 have been evaluating, piloting, restricting, and deploying AI tools in one form or another since 2023. But for much of the rest of legal practice, last week may be the first unmistakable signal that legal AI is no longer a distant development or a specialized BigLaw experiment.

This article addresses one question: What happened on May 12 and May 13, and what happened in the four months that led to those announcements? Two follow-up questions will be taken up in subsequent pieces: what this touches on in the reader’s own work and what the reader now needs to learn.

The intended reader is not only the BigLaw partner who has been tracking legal AI announcements since 2023. It is also the litigation partner in Atlanta, the corporate associate in Cleveland, the solo practitioner in Sacramento, the in-house counsel in Pittsburgh, the paralegal, and every other person working in legal practice who has been doing the work the way they have always done it.

What Happened

On May 12, Anthropic announced a dedicated legal push for Claude, including 12 practice-area plugins and more than 20 new connectors linking Claude to legal software systems. This did not happen in isolation. It built on Anthropic’s May 7 expansion of Claude for Microsoft 365, which made Claude generally available across Word, Excel, and PowerPoint, with Outlook in public beta.

That matters because the story is not just “AI for lawyers.” The more important development is that Claude is now positioned to move across the software environments where legal work already lives: office productivity tools, document systems, discovery platforms, contract workflows, and research products. In other words, the new contest is not only about who has the best chatbot. It is about who becomes the connective layer across legal work.

Within hours of Anthropic’s legal announcement, Thomson Reuters said the next generation of CoCounsel Legal had been rebuilt on Anthropic’s Claude Agent SDK, introducing an integration that connects Claude directly to CoCounsel Legal. The following day, LexisNexis announced that it had integrated Anthropic’s Claude legal plugin suite into Lexis+ with Protégé.

The announcement language was already being recalibrated within days. Thomson Reuters CTO Joel Hron clarified through Reuters that the Claude integration does not replace the CoCounsel Legal platform and does not provide standalone access to the underlying content-and-workflow system. The first capability available inside Claude is Westlaw Deep Research, exposed through a limited set of tool calls, not the full CoCounsel system. The gap between announcement language and actual deliverables is the kind of distinction this publication will track piece by piece.

In less than 24 hours, the two dominant legal research and information providers publicly linked major legal AI offerings to the same underlying AI company. That is the central development. For decades, the legal information market has been organized around separate platforms: Westlaw, Lexis, firm systems, discovery environments, contract repositories, and specialized tools. The May 12 and May 13 announcements suggest the beginning of a different structure, one in which a single AI layer can reach across many of those systems at once through connectors, plugins, and embedded workflows.

That is not merely a product story. It is an infrastructure story.

The Four Months Before

The May 12 launch looks less surprising when placed in sequence. In late January and early February, Anthropic’s legal-adjacent plugin moves were already provoking reactions well beyond the legal press. On February 3, shares of Thomson Reuters, RELX, and Wolters Kluwer fell sharply after investors interpreted Anthropic’s legal plugin push as a real competitive threat to incumbent information businesses. Based on our reconstruction from public pricing data and share counts, the three companies lost an estimated $20 billion or more in combined market capitalization in the trading sessions that followed.

By late April, Anthropic had gone a step further, announcing a long-term co-development partnership with Freshfields to build AI legal workflows and deploy Claude across the firm globally. That was a sign that Anthropic was not just offering a general model to legal users. It was moving toward firm-specific workflow design within one of the world’s largest law firms.

Then, on May 7, Anthropic expanded Claude’s Microsoft 365 footprint. A few days later, it introduced the legal-specific layer. Then, Thomson Reuters publicly deepened its relationship with Anthropic. Then LexisNexis did the same.

The public adoption picture at the time of the May 12 launch extended beyond Freshfields. Fortune‘s launch coverage reported that Quinn Emanuel Urquhart & Sullivan, Holland & Knight, and Crosby Legal were also using Claude on live matters, citing confirmations from the firms alongside Anthropic. A plurality of major firms, not a single flagship adopter, was publicly committed.

The launch also arrived against an active backdrop. In April, Fortune reported that Sullivan & Cromwell, one of the most prestigious law firms in the United States, acknowledged to a federal bankruptcy judge that a recent filing contained AI hallucinations flagged by opposing counsel. Fortune‘s launch coverage placed that incident directly against the picture of BigLaw firms publicly committing to Claude. Anthropic’s stated answer to the hallucination problem, per Fortune, is grounding: the new connector architecture restricts Claude to drawing from live external sources rather than generating from memory. Whether grounding actually changes the hallucination picture is among the questions this publication will track piece by piece.

Seen one by one, these developments could be mistaken for ordinary product updates. Seen together, they look more like a compressed market transition. The period for slow observation may not be over everywhere, but it is harder now to argue that the profession is still in an early, purely experimental stage.

What Comes Next

The ground under legal practice has shifted. The building still stands. The foundation underneath it is no longer what it was a week ago.

What this means for any individual legal professional still depends on role, practice area, client expectations, technology environment, and risk tolerance. A legal aid attorney, a litigation support manager, an in-house contract team, and a BigLaw M&A partner will not experience these shifts in the same way.

The next pieces in this series will organize the same set of facts differently. One will walk through the named components of the launch and explain what each connector or plugin appears to do. Another will organize the changes by legal role. A third will sort them by practice area, including litigation, corporate, employment, privacy, regulatory, intellectual property, tax, and adjacent fields implicated by the launch materials.

What these articles will not do is decide what the change means for any particular practice, matter, client, or professional judgment. They will document what is now in place. The reader does the rest.

Sources

• Anthropic, “Claude for the legal industry,” May 12, 2026

• Thomson Reuters press release, CoCounsel Legal and Anthropic integration, May 12, 2026

• LexisNexis press release, Lexis+ with Protégé and Claude legal plugin suite, May 13, 2026

• Anthropic, “Collaborate with Claude across Excel, PowerPoint, Word and Outlook,” May 7, 2026

• Reuters, “Anthropic expands Claude’s AI tools for law firms, lawyers,” May 12, 2026

• Reuters, “AI concerns pummel European software stocks,” February 3, 2026

• Freshfields, “Freshfields and Anthropic team up to co-build AI legal workflows,” April 23, 2026

• Fortune, “Even as hallucinations show up in legal filings, Big Law goes all in on AI with new Anthropic release,” May 12, 2026

A Note on Methodology: Every dated event, named participant, and quoted statement in this piece traces either to a primary source (company announcement, press release, court opinion, public filing) or to verified secondary reporting that quotes a primary actor. Where a number rests only on trade press characterization without a primary disclosure behind it, this publication will either trace the number to its primary source or leave it out. The three-company market capitalization decline cited above was reconstructed by this publication from public pricing data and share counts, not taken from trade press summary.

AI tools used in preparation. This article was drafted with Claude (Anthropic); fact-checked and link-verified with Perplexity, ChatGPT, and DeepSeek; and copy edited with Gemini. All factual claims trace to the cited primary sources or to verified secondary reporting. Editorial judgments, framing, and analytical conclusions are the author’s.

This publication is for informational and educational purposes only. Nothing herein constitutes legal advice, and reading this publication does not create an attorney-client relationship. Readers should consult qualified legal counsel for advice specific to their situation.

Copyright © 2026 Bob Meyers. All rights reserved.