When a legal AI vendor builds its own model instead of renting one from a frontier lab, the questions you put at renewal change. Thomson Reuters announced on 24 August 2026 that it has trained a model of its own, named Thomson, on the content behind Westlaw, Practical Law, Checkpoint and Reuters. It reaches customers first inside CoCounsel Legal, powering the tabular analysis feature that pulls a structured answer out of a stack of documents.
The company puts the investment at about 40 million dollars in talent and compute, built on an open-source foundation and then specialised. Its chief technology officer, Joel Hron, describes the method as starting from a strong foundation and specialising it deeply for the work that matters. Less than a tenth of the company's content has gone into training so far, which tells you plainly that the product you buy this autumn will not be the product you renew next year.
Owning the model shortens the chain
Most legal AI products until now have been a layer of software sitting over somebody else's model. That mattered for your due diligence, because the promises about accuracy, security and data handling ran through your supplier to a third party you had no contract with and no leverage over. When the supplier owns the model, that chain shortens to one company, one agreement and one set of undertakings you can hold to account.
The trade is that outside evidence becomes harder to find. The performance claims here are the company's own early evaluations, which report the model doing better than general models on tasks that draw on Thomson Reuters content and less well on open web tasks. Full technical benchmarking has yet to be published. That is a fair thing for a vendor to say about its own work and a weak thing for a buyer to rely on. The same caution belongs on every demonstration, as the buying discipline set out here works through.
What to ask before you sign
Start with which model answers which task inside the product, and what happens when that changes. A firm running a specialist tool alongside a general assistant already has to keep track of who does what, a question the multi-model piece takes further. Ask for written notice when the model behind a feature is swapped or retrained, because your supervision arrangements were built around the behaviour of the old one.
Then take up the data question. Mr Hron says customer data plays no part in training and that internal experts recreate the relevant tasks instead. Treat that as the opening position for the contract rather than the answer, and ask for the same words in the agreement, alongside the sub-processor list and the place where the model runs. A vendor training on its own archive rather than on the open web helps you here, because the provenance of the training material is something it can describe.
Finally, ask what evidence exists for accuracy on the work you do. Tabular analysis across a bundle of leases is a different task from finding the authority that decides a point, and a model that scores well on one tells you little about the other. Ask for the evaluation that covers your workload, and where none exists, run your own on matters you have already closed and know the answers to.
Your duty stays where it was
None of this shifts the obligation. Whether the model was trained in California or from a legal publisher's own archive, the solicitor who signs the advice carries responsibility for it, which is what the SRA warning notice of 17 August restates. A model trained on Westlaw content will read as more authoritative to a fee earner in a hurry, and the verification habit has to hold against that impression rather than relax because the badge on the tool looks familiar.
Before your next renewal, write down three lines. Name the model behind each feature you rely on. Name what the contract says about your client data and who else touches it. Name the test you will run yourself, on closed matters, before the tool touches live work. A supplier that owns its model can answer all three, and one that will not answer them has told you something worth knowing.
The announcement sits in the Thomson Reuters press release, which is open to read and asks nothing of you.
If a renewal quote is on your desk and you want those three lines answered before you sign it, that is work we do: send us the proposal.