When Every Legal AI Has a Good Model, What Differentiates It?

Model Parity Won't Save a Legal AI Vendor — The Data Underneath It Will

On 21 September 2026, Artificial Lawyer published an op-ed by Deepak Kapoor, CEO of the Indian legal database Manupatra, arguing that as foundation models converge in capability, competitive advantage in legal AI is shifting away from the model itself and into what Kapoor calls "legal information architecture" — how legal material is collected, connected, kept current and made traceable back to its source. His core claim: a well-reasoned AI answer built on outdated or poorly structured material is still a wrong answer, so currency and provenance are themselves part of a system's intelligence, not just database housekeeping.

Does this argument transfer from legal research to litigation evidence?

Partly. Discovery has no equivalent of an overruled judgment, but it has an identical exposure: an AI answer is only as good as its ability to prove where it came from.

Kapoor's piece is written for research platforms tracking a moving target — statutes get amended, precedent gets distinguished or overruled, and a system has to know today's status, not last year's. A litigation collection is different: it's a fixed corpus, frozen at the point of gathering. There's no "currency" problem in the sense he means. But the underlying warning still lands for eDiscovery buyers, just translated — the risk isn't a stale statute, it's an AI-generated chronology or issue-coding pass that can't be traced back to the specific email, chat, or exhibit it was built from. Same failure mode, different mechanism.

What should a buyer ask their eDiscovery vendor instead of a legal research vendor?

Ask whether every AI-generated fact, date, or coded issue links to a specific document, page, and Bates number — not just a similar one.

Kapoor draws a distinction worth stealing wholesale: a system that finds "relevant authority" versus one that finds "semantically similar text." In discovery, that becomes the difference between a chronology entry that cites the actual exhibit a fact came from, and one that's a plausible-sounding synthesis nobody can pin to a document if opposing counsel challenges it at deposition or in a Daubert-style fight over the AI's methodology.

Should you be sceptical that this argument comes from a legal database vendor?

Yes — Manupatra sells exactly the layer its CEO says now matters most, which doesn't make the argument wrong, but it is self-serving.

No benchmark, error rate, or comparative figure is offered anywhere in the piece — it's a positioning essay, not a study. That's fine as far as it goes, but it means the reader should treat "structured, current, provenance-backed data" as a checklist to interrogate vendors with, not as a claim already proven by anyone, including Relevant.

Frequently asked questions

Does this change how firms should evaluate eDiscovery AI in 2026?

It reinforces a question worth asking regardless of vendor: can the system show its source for every AI-generated answer, or only its confidence?

Is a bigger context window or a "better" model still worth paying for?

Kapoor's argument, and the eDiscovery parallel, is that model choice matters less once several vendors can access comparably capable models — the data layer becomes the differentiator.

Does this relate to other coverage on this site?

Yes — it follows the site's recurring theme, also raised in "AI Gains Slim for Most Staff, Is Legal Different?", that legal AI's value depends on verifiable grounding, not raw model performance.

Source: Artificial Lawyer, "When Every Legal AI Has a Good Model, What Differentiates It?" by Deepak Kapoor, 21 September 2026.