CoCounsel

Every lawyer who has tried a general-purpose chatbot on a legal question has had the same experience. The answer is fluent, well structured, correctly formatted — and somewhere in it is a case that does not exist.

That is not a bug to be patched. A model generating text will generate plausible citations the same way it generates plausible sentences, because from the inside those are the same operation.

CoCounsel's response is to change what the system reasons from. Instead of the open web, it works from Westlaw, Practical Law and your own documents, and it returns citations you can click.

CoCounsel: Legal AI Grounded In Westlaw

• Thomson Reuters' own product, reasoning from Westlaw content, Practical Law guidance and your organisation's documents.
• Linked citations you can open and verify, which changes the failure mode from invented authority to retrievable authority.
• Tabular Analysis reviews up to 10,000 documents against 100 questions in one table.
• Westlaw Brief Builder takes a litigation brief from issue framing to argument structure, with the attorney in control.
• Percentages published with an asterisk — 76%, 64% and 33% figures that describe survey agreement, not measured outcomes.

What Does Grounding In Westlaw Change?

Thomson Reuters describes CoCounsel Legal as the only AI grounded in Westlaw content, Practical Law guidance and your firm's own documents, handling research, analysis and drafting from a single query.

The important word is grounded, and it means something specific: the system retrieves from a defined corpus and reasons over what it retrieved, rather than generating from everything it has ever absorbed.

That changes the character of the mistakes. An ungrounded model can produce a citation to nothing; a grounded one can produce a citation to something real but wrong for your matter. The second is a legal judgement error, which lawyers are trained to catch. The first is a fabrication, which nothing in professional training prepares you for.

The corpus itself is the other half. Westlaw and Practical Law are licensed, edited and maintained, with editorial treatment layered on top of primary law — which is a meaningfully different input from whatever version of a case happens to be indexed publicly.

None of this removes the obligation to verify. It changes what verification consists of, from checking whether the authority exists to checking whether it says what the tool claims — which is the work a lawyer should have been doing anyway.

An ungrounded model can cite nothing. A grounded one can cite something real but wrong. The second is a judgement error lawyers are trained to catch; the first is not.

Why Citations You Can Click Matter

The site states that drafts come with linked citations for verification, and that is a design decision worth understanding rather than skimming past.

Verification friction determines whether verification happens. If checking a citation means leaving the tool, opening a database, running a search and reading a case, most people will check the first two and assume the rest.

A linked citation collapses that to one click, which is the difference between a policy that says everything must be verified and a practice where it actually is.

It also makes the tool auditable after the fact. When someone asks in six months how a conclusion was reached, the chain from assertion to authority is still there rather than reconstructed from memory.

This is the same principle behind grounded drafting in other regulated fields — patent work in DeepIP and contract review in LegalOn both live or die on whether the output points back at something checkable.

Reviewing 10,000 Documents Against 100 Questions

Tabular Analysis is described as reviewing up to 10,000 documents against 100 questions, with results delivered in one dynamic table you can filter.

The shape of that feature tells you what it is for. This is diligence, discovery and portfolio review — the work where the difficulty is not any individual document but the volume and the consistency of the questions asked of each one.

Consistency is the real gain. A team of associates reviewing ten thousand documents will interpret question seven slightly differently on day four than on day one, and nobody will notice until it matters.

The table format matters too. Answers arranged in a grid are comparable, sortable and reviewable in aggregate, which lets you spot the outliers — the three contracts that answered differently — rather than reading everything again.

What it does not do is remove the need to read the ones that matter. The right use is triage: let the table tell you which two hundred documents deserve human attention, then give them proper attention rather than skim ten thousand.

• Verify what a citation says, not merely that it exists — that is the new work.
• Use bulk review as triage, not as a substitute for reading what matters.
• Read the asterisk on any published percentage before quoting it internally.
• Check whether your own documents are used to train anything, in writing.
• Pilot on a closed matter where you already know the right answer.

Reading The Percentages On The Page

The product page leads with three figures: 76% agree it improved research and drafting quality, 64% agree it helped reduce risk, and a 33% average reduction in time spent on document review, research and drafting. All three carry an asterisk.

Two of those are agreement statistics, which measure what users reported believing rather than what was measured. That is not worthless — practitioner perception is a real signal about whether a tool gets adopted — but it is a different thing from an outcome study.

The third, the 33% time reduction, is the one that would matter most for a business case, and it is also the one most sensitive to how it was collected. Self-reported time savings are notoriously generous, and averages hide the distribution.

None of this makes the figures dishonest. They are published with an asterisk, which is the correct way to present survey data, and the honest reading is that they indicate direction rather than magnitude.

The practical response is to generate your own number. Run a defined task both ways on a closed matter, measure the actual difference, and use that for your business case rather than a vendor's average.

Where CoCounsel Sits Among Legal Tools

Legal AI has fragmented into categories that look similar and solve different problems, and the distinctions matter when you are choosing what to buy.

Research and drafting assistants — CoCounsel's category — start from a legal question and produce reasoned, cited work product. The differentiator is the corpus behind them.

Contract review sits elsewhere. Assessing whether a clause is acceptable against your playbook is a comparison task rather than a research task, which is why dedicated reviewers such as Ivo and Luminance exist as separate products rather than features.

Contract storage is a third thing again, and the most commonly conflated. Being able to find and track your executed agreements — what platforms like ContractSafe do — is an operations problem, not a legal reasoning one.

The fourth category is intelligence about risk that has not yet become a matter, which is a genuinely different proposition from anything above. Confusing these four is the most common way firms buy the wrong thing and conclude that legal AI does not work.

Introducing CoCounsel To A Practice

Adoption failures in legal AI are rarely technical. They happen because the tool was introduced without a rule about what it may be used for and who checks the output.

This sequence establishes both before anything touches live client work.

1. Pilot on a closed matter where you already know the correct answer.
2. Compare the tool's output against what the file actually shows, citation by citation.
3. Write down which task types are permitted and which are not, before wider rollout.
4. Establish who verifies output and record that the verification happened.
5. Get the data handling terms in writing, particularly around your own uploaded documents.
6. Measure your own time saving on a defined task rather than adopting a published average.