AI Workflows  /  First Look

CoCounsel Legal AI: What Thomson Reuters Changed for Lawyers

Thomson Reuters just opened early access to a rebuilt CoCounsel Legal, the biggest overhaul of its flagship legal AI since it bought Casetext. It runs on Anthropic's Claude Agent SDK and works like a colleague that plans and executes instead of a menu of tasks. Here is what actually changed, what is real, and what is still a vendor promise.

By Anthony Guerriero  ·  Reviewed by The Leveraged Years Editorial Desk  ·  Published June 28, 2026  ·  Last updated June 28, 2026

On June 22, 2026, Thomson Reuters opened early access to the next generation of CoCounsel Legal, its flagship legal AI, rebuilt on Anthropic's Claude Agent SDK. The change is architectural: instead of discrete "skills" and structured workflows, a lawyer describes a matter in plain language and the system drafts a plan, pulls from the user's precedents plus Westlaw and Practical Law, drafts with traceable citations, and returns one iterable work product. Every US customer gets a toggle to the new experience now; general availability is planned for August 2026 in the US, and pricing was not disclosed. Primary source: LawNext (Bob Ambrogi) and Thomson Reuters.

What the new CoCounsel Legal actually is

The short version: Thomson Reuters rebuilt its flagship legal AI from the ground up, and the change is not a new feature. It is a change in what the product is.

The first generation of CoCounsel Legal, descended from the assistant Casetext launched in 2023 and folded into Thomson Reuters in its 2023 acquisition, was organized around discrete "skills" and structured workflows. You picked a skill, defined a workflow, or wrote a detailed prompt, and answers came back task by task. You did the orchestration. The tool did the steps you told it to.

The next generation, opened to early access on June 22, 2026, drops that model. Per Thomson Reuters, a lawyer now describes a matter in plain language and the system drafts a plan, works through the legal issues, pulls from the user's own precedents plus Westlaw and Practical Law, drafts with citations, and adapts as new facts surface. The output is a single, iterable work product rather than a string of piecemeal answers. Thomson Reuters' own framing is that the tool is meant to feel less like a menu of software functions and more like working with a colleague.

That is the real headline. CoCounsel went from a menu to an agent.

What changed, in one sentence: skills became an agent

Under the hood, the company says CoCounsel Legal was re-architected to incorporate Anthropic's Claude Agent SDK at the model layer, which Thomson Reuters characterizes as creating a legal agent with native access to Westlaw and Practical Law. CTO Joel Hron draws the line carefully: Anthropic's Claude supplies the reasoning at the foundation layer, but what the customer buys is the system built around it, the proprietary content, the domain expertise, and the governance.

That distinction matters for how you read this. The intelligence is Claude. The product is the wrapper: authoritative content grounded in from the first step, jurisdictional coverage, workflow integrations, and the editorial work of what the company says are thousands of attorney-editors. Hron's own line is that "models are table stakes" and "the real differentiator is the system around the model." Whether that holds is the open question every buyer should sit with.

The rollout: early access now, general availability in August

Be precise about status, because the marketing blurs it. As of the week of June 22, 2026, this is early access, not a full launch. Every CoCounsel Legal customer in the United States gets a toggle to switch into the new experience and back to the current version at will, and both versions stay available through the early-access period. Thomson Reuters says general availability is planned for August 2026 in the US, followed by Canada, the United Kingdom, and Australia.

Pricing was not specified. For a tool this central to a practice, that is not a small omission, and it is the first thing to pin down before you build a workflow around it.

The "fiduciary-grade" claim, and how to read it

The phrase Thomson Reuters leans on hardest is "fiduciary-grade AI," meaning AI built for work under regulatory oversight and duties of care, where accuracy, confidentiality, and accountability are the whole point. The company says the standard means four concrete things: outputs are grounded in authoritative content from the first step of reasoning rather than checked against sources afterward, every citation is intended to be traceable to a linked source, the reasoning process is inspectable, and confidential data is not used to train third-party foundation models.

For legal work, those are the right promises, and they map directly to the failure mode that has gotten lawyers sanctioned: confident text with fake or unverifiable citations. Grounding first and tracing every cite is the correct design.

Here is the discipline, though. These are design claims and vendor descriptions, not independent audit results. "Grounded from the first step" and "every citation is traceable" are how the system is built to behave, not a guarantee it never errs. The professional duty to verify does not transfer to the vendor because the marketing uses the word fiduciary. Read it as a serious and well-aimed standard that you still have to test on your own matters.

Workspaces, Brief Builder, and what is still rolling out

Thomson Reuters frames early access as a starting point, with three capabilities at different stages. Workspaces, described as a dedicated environment for each matter that carries documents, precedents, and prior positions across sessions and colleagues, is live today. Brief Builder, an agentic drafting tool for briefs and motions grounded in Westlaw and Practical Law with citation checking and issue spotting, is still rolling out. So is a firm and organizational intelligence layer, described as skill-authoring that lets a firm encode its own expertise so the system applies it consistently across matters and teams.

Translate that last one for your practice. The pitch is that your firm's way of handling a recurring matter becomes a reusable skill the agent applies the same way every time. That is genuinely useful if it works, and it is also exactly the kind of capability that is easy to demo and hard to operationalize. Put it on the list to test, not on the list to assume.

Where it fits next to general-purpose Claude

For a lawyer already using Claude directly, the obvious question is whether this replaces it. It does not, and the relationship is more interesting than replacement.

Both run on Anthropic's technology, but they are aimed at different jobs. General-purpose Claude is broad, fast, and yours to point at anything, from a first draft to a thinking partner, with no native tie to Westlaw or Practical Law. CoCounsel Legal narrows that same reasoning to legal work grounded in licensed authority with traceable citations, which is the part general tools cannot do because they do not have the content.

The connective tissue is worth knowing. Thomson Reuters frames CoCounsel as an emerging "agentic operating system" reachable through APIs, connectors, and the Model Context Protocol, and it says its MCP integration with Anthropic lets lawyers working inside Claude reach CoCounsel Legal without switching platforms. So the practical near-term picture for a firm is not one tool winning. It is general-purpose Claude for open-ended work and CoCounsel for the grounded, citable legal product, increasingly able to talk to each other through MCP.

The build-versus-buy question for a firm

Strip away the launch language and the decision in front of a firm is the old one in new clothes: buy the grounded, governed system, or build your own thinner workflow on a general model.

Buying CoCounsel Legal gets you native Westlaw and Practical Law, the editorial layer, and a vendor standing behind traceable citations, at a price not yet public. Building your own on general-purpose Claude gets you flexibility and lower license cost, but you supply the authoritative content, the verification step, and the governance yourself, which is real work and real risk in a domain where a bad cite gets sanctioned.

The honest read for most firms is that this is not ideological. It is a function of how much of your work is genuinely grounded in licensed legal authority versus open-ended reasoning, and how much verification capacity you already have. The firms that win with either path are the ones that decide deliberately, keep a human verification step no matter which they pick, and do not confuse a vendor's confidence with their own duty. That judgment, not the tool, is the durable skill.

Browse all AI tool workflows

Informational tool analysis for working professionals, not legal, medical, or financial advice. AI tools do not replace your professional judgment.