The legal AI conversation has been dominated by models. Which one reasons better? Which has the larger context window? Which performs best on the benchmarks? Which is fastest? These are important questions.
But as legal AI moves from experimentation into everyday professional work, a different question may matter more: what is the AI actually working with?
Reasoning is only as good as the material it reasons over. A model can be excellent at working across documents, and still have limited use if the legal material underneath it is incomplete, outdated, poorly structured or hard to verify.
Legal Data Is Not Just a Collection of Documents
At first glance, the raw material of legal research looks straightforward: judgments, legislation, regulations and other legal materials. But a system lawyers can actually rely on is more than a repository. Law is connected.
- A judgment cites earlier judgments — and is later followed, distinguished or overruled.
- A statutory provision is amended; an Act is repealed.
- A notification quietly changes how a provision operates.
- Different authorities acquire different significance depending on the court, the jurisdiction, the issue and the point in time.
For decades, legal publishers and research platforms have done the largely invisible work of organising these relationships. That work becomes more important in the age of AI, not less. An AI system does not merely need access to legal text — it benefits from information that has been collected, classified, connected, updated and enriched in ways that reflect how lawyers actually research the law.
Generative AI does not make legal databases less relevant. It may make the quality of the database underneath the AI considerably more important.
The Model Is Only One Part of the Stack
The foundation models available to legal technology companies are improving extraordinarily quickly. Companies can increasingly choose between several of them, and change them as capabilities evolve.
That shifts where durable differentiation is created. If many platforms can reach increasingly capable models, the harder questions become:
- Access — what information can the model reach, and how current is it?
- Structure — has that information been organised the way law is organised?
- Authority — can the system identify relevant authority, not just semantically similar text?
- Verification — can the lawyer return to the original source and determine whether the authority remains good law?
These are not merely model questions. They are legal information architecture questions.
Currency Is Part of Intelligence
Law changes. A beautifully reasoned AI response built on an outdated statutory provision is still the wrong answer for the lawyer advising a client today. A judgment may have been overruled or distinguished. A provision may have been amended or repealed. A new decision may materially change the legal position.
In legal AI, currency is not simply a database maintenance issue. It is part of the intelligence of the system.
A more reliable division of labour
That leaves the lawyer free to inspect the authority behind the answer, which is where the professional judgment actually belongs.
Provenance May Matter As Much As the Answer
Generative AI has made producing plausible answers remarkably easy. Legal practice sets a higher threshold: the lawyer does not only need an answer, but a reason to rely on it.
Where did the proposition come from? Which judgment supports it? What does the judgment actually say? Does the cited paragraph support the proposition? Has the authority subsequently been treated differently?
This is why provenance and verification need to be core parts of legal AI architecture, rather than features added after generation.
The workflow that holds up
AI can compress the first four stages considerably. The final stage remains fundamentally professional.
The Other Half of the Data Equation Belongs to the Lawyer
There is another form of data that matters enormously: the lawyer's own documents. Legal work usually involves bringing together two information environments — the external legal universe of cases, legislation and regulation, and the matter-specific universe of contracts, pleadings, correspondence, evidence, opinions, orders and client information.
Much of the potential of legal AI lies in connecting the two.
Take something as ordinary as this question
The contract provides the factual context and the language to be analysed. The legal information system provides the applicable legislation and authorities. AI can help connect them: identify the issues, locate the relevant law, compare authorities against the facts, organise the analysis and prepare a first draft.
The usefulness comes not simply from generating text, but from being able to work across matter context and legal context at the same time. That is a far more interesting role for AI than being a better chatbot.
The value is not in the text it produces. It is in the two contexts it holds together.
From Legal AI Tools to a Legal Work Layer
This is also why the next phase of legal AI will increasingly be about workflow. Lawyers rarely perform isolated tasks. A research question leads to cases. Cases need to be read. Their relevance needs to be assessed against the facts. The analysis may become an opinion, which may then lead to drafting, negotiation, litigation strategy or client advice.
Agentic AI is interesting because it allows technology to work across a sequence of tasks while carrying information from one stage into the next. But agentic capability makes the quality of the underlying information more important, not less. An agent that can execute ten steps autonomously can also propagate an error across ten steps.
What This Means for Established Legal Information Businesses
Manupatra has been thinking about this question while building ManuWorks.ai. The starting point is unusual: the AI layer has been added to a legal information system built and editorially enriched over 26 years. ManuWorks therefore sits above an existing body of legal content, search infrastructure and relationships between legal materials, while also allowing lawyers to work with their own documents.
But the broader lesson is not about any particular product. It is about how legal AI may evolve. In the first phase of generative AI, much of the attention was on what the model could produce. The next phase may be about what the model can reliably work with, what it can do across that information, and how easily a professional can verify the result.
“Data Is the Moat” — Restated
Owning a large quantity of legal text is not the moat. The moat is legal information that is authoritative, current, structured, connected and capable of being interrogated by machines, while preserving a clear path back to the source for the human professional.
The real value emerges when that legal knowledge can be combined with the lawyer's own matter-specific information, and an AI layer capable of working across both.
That may ultimately be where legal AI becomes genuinely useful — not when AI knows everything, or attempts to replace the lawyer's judgment, but when it can do more of the groundwork required to put the right information, with the right context and the right sources, in front of the lawyer at the right point in the work.
The model matters.
What sits underneath it may matter more.