She found the precedent on a Tuesday morning. It was buried in a matter folder from three years back, exactly where instinct said it would be, and it answered the question precisely. The advice went out that afternoon. Two weeks later, a colleague flagged that the statutory position the memo relied on had been amended in 2024. The advice was wrong. The precedent was real; the law it assumed was not.
That scenario sits at the centre of a debate Thomson Reuters reopened on 17 June 2026 when it announced general availability of its CoCounsel-DeepJudge integration, which lets lawyers query their firm's own documents, precedents, and past decisions from within the CoCounsel interface. The proposition behind the release was that firms gain an edge by applying their own precedent, perspective, and past decisions to new matters, and that raw model capability has stopped being the differentiator.
One of the largest incumbents in legal technology has said, publicly, that raw model capability has stopped being the point.
We think that is the right diagnosis. The argument Thomson Reuters is now making is one we have been making for a while: the value in legal AI is grounding. Frontier models generate fluent text. They summarise, restructure, and produce plausible-looking output at speed. What they cannot do on their own is tell you whether any of it is true, traceable, or grounded in the materials that govern your matter. So the firm that spent years telling the market that CoCounsel ran on the best available models is now telling that same market that models alone were never enough. That is the right conclusion.
The question worth pressing is whether the treatment follows.
Retrieval and verification are different things. Grounding an output in a document does not, by itself, confirm that the output accurately represents what the document says. The underlying model is still generating. Hallucinations in retrieval-augmented systems are well-documented: the system retrieves a real document and then mischaracterises what it contains. The citation exists; the summary is wrong.
There is a second, less-discussed risk. A firm's precedent bank is not a validated corpus. It contains advice prepared under old law, settlements that never went to judgment, internal memos written before a statute was amended. The senior associate's Tuesday precedent was real. The statutory position it assumed was not. Grounding confidently in institutional knowledge that is itself outdated does not reduce error; it gives outdated positions a veneer of authority. Retrieval over an unaudited corpus is grounding of a kind, but a practitioner who cannot see the audit trail cannot know which kind they have.
The Federal Court's GPN-AI, in force since April 2026, makes these distinctions enforceable rather than theoretical. Personal verification of cited authorities is required before filing, and the standard is specific: the authority must exist and support the stated proposition. Retrieval from a document corpus satisfies the first condition. Whether the AI's characterisation of that document accurately reflects its holding is a separate question, one the practitioner signing the document must be able to answer.
That is the standard against which any grounding claim should be tested. Can you see, at the output level, which passage in which source generated which claim? Can you trace the reasoning back to its origin? Can you produce that trail to a supervising partner, a client, or a court?
This is where the category reads differently from the way Thomson Reuters is framing it. The value is grounding. But grounding is only as reliable as the citation trail behind it, and buying a retrieval integration is not the same as buying verifiability all the way down. Habeas's Search Engine scans over 300,000 Australian cases and pieces of legislation in seconds, from a closed dataset of legitimate Australian legal sources, with results that are verifiable and traceable rather than generated from a corpus that may or may not reflect current law. The Document Stores feature applies that same citation discipline to a practitioner's own uploaded materials: matter documents, briefs, internal advice, all queryable, all with a traceable path from output to source. Foundational research that used to take a full morning can be completed in minutes, and the citation trail a partner follows at the end is the same trail the system used to construct the answer.
Thomson Reuters entering this space with the message that grounding beats horsepower is, for those of us building grounded research tools, market validation. The most capitalised incumbent in legal technology has concluded that cited, traceable, retrieval-grounded output is where the value sits. We agree. The question each practitioner now needs to answer is whether the grounding they are buying is verifiable at the level the GPN-AI requires.
The senior associate's Tuesday memo was grounded in something. The something was wrong. The integration that answers her problem is one where the output shows her, before the advice goes out, exactly where the law currently stands.
See it for yourself at habeas.ai.
If you want to try for yourself or get in contact, book a demo with us here. We also offer the capacity for self-serve individuals to sign up, and subscribe or register a free trial at app.habeas.ai.
The legal research in this article was conducted and every citation verified using Habeas, the Australian legal AI research platform.
Hero image: Tara Winstead on Pexels
