News in Practice

AI and legal research: searching case law from inside the matter, not beside it

The same need, two different actions. In one, the question leaves the matter and the answer comes back diminished. In the other, it never leaves.

For lawyers and in-house counsel who research case law regularly. What the round trip actually costs, and what remains the province of a specialised database.

A piece of legal research always arises inside a matter: about a document, an argument from the other side, a clause. It is nonetheless carried out outside it, in an environment that knows neither the matter nor the reason for the question. That displacement looks harmless because it has become natural. It produces three effects most practitioners have stopped noticing.

Regime 1: research conducted beside the matter

You leave the matter. The question, which bore on a specific fact, has to be reformulated in general terms: the context does not travel, so it must be replaced by keywords.

First effect: loss of context. The question put to the database is not the question you were asking. It is an impoverished version, stripped of what made it precise, and the answer therefore addresses a question adjacent to yours.

Second effect: reconstruction on return. You get three decisions. You now have to establish what each changes for your matter, which means mentally rebuilding the link between what has just been found and what you knew before leaving. That work is the real work, and it is done without any support.

Third effect: no trace. What you searched, what you set aside and why, survives nowhere. Six months later someone else, or you, will ask the same question and retrace the same route, unaware it has already been done.

These three effects compound, and the third is the costliest at the level of the organisation, because it is the only one whose effect grows over time: every unpreserved search is a search that will be redone.

They also fall unevenly by seniority. An experienced practitioner compensates better for the loss of context, because they know what needs restating. A junior absorbs it in full, without knowing that they are absorbing it.

Regime 2: research conducted from inside the matter

The question keeps its context, because it has not left the place where that context sits. It bears on the clause in article 7 of the contract in front of you, not on a general notion.

The most visible difference is not speed, it is the precision of what comes back. A search that knows which matter it comes from can set aside what does not apply, which no keyword reformulation permits.

That precision has a direct consequence for volume: contextualised research retains few decisions and retains them better, whereas keyword research produces many, most of which will have to be discarded on reading.

The most durable difference lies elsewhere: the result attaches to the matter. The decision retained, those set aside, the reason for setting them aside become elements of the matter rather than the contents of a closed browser tab. That is what makes it possible, six months later, to consult instead of redoing.

What research beside the matter costs is not the time spent searching. It is the context lost on the way out and the link rebuilt on the way back.

The question that comes back

It is worth looking at what happens the second time, because that is where the gap between the two regimes widens.

In the first regime, the second search on the same subject is identical to the first. Nothing indicates that it has been done, nor what was retained, nor what was set aside and why. On an eighteen-month matter, the same question may be handled three or four times, by different people, with slightly divergent results.

That divergence is more costly than the repetition. Two notes from the same firm retaining two different decisions on the same question do not necessarily contradict each other on the merits, and they produce the same effect as a contradiction when they end up side by side in front of a client or an opponent.

In the second regime, the second search starts from the first. It may confirm it, refine it, or find that a later decision has altered it. It is the same work, carried out on a known base.

What verification requires

Research integrated into the matter dispenses with no checking, and that must be said plainly because it is the point at which too much is usually promised.

A reference must remain verifiable: court, division, date, number, link to the official decision. An approximate reference is not a reference, and a system that does not produce verifiable ones shifts the burden onto whoever must check them.

That requirement is not peculiar to recent systems: it applies identically to research carried out by hand, where a mis-transcribed reference produces exactly the same effect.

A retained decision must have been checked against what followed it: a later ruling, a clarification by a higher court, a contrary decision. A result presenting a judgment without having looked for what has affected it since is more dangerous than useful.

A third requirement bears on what was not found. A system that finds nothing must say so as such, and not leave the impression that nothing exists. Absence of results is ambiguous information: it may mean that no decision exists, that the query was wrong, or that the decision is not in the corpus searched. Those three situations call for different responses.

These requirements are verifiable by the practitioner themselves, and they constitute the only serious criterion for assessing any research system.

They have ceased to be theoretical. Between December 2025 and February 2026, several French courts, both civil and administrative, identified in filings references to case law that did not exist. One expressly invited drafters to satisfy themselves that their citations were not hallucinations; another reminded counsel that it is for them to verify the decisions cited before coming to court.

What is striking in those cases is not that references were invented, which was foreseeable, but that they passed review. They were plausible: right court, correct format, a holding consistent with the argument. Nothing signalled to the reviewer that they needed opening.

That is why the verifiability of a reference is not a matter of convenience. A reference that cannot be opened in one action will not be opened, and an unopened reference is an unverified one.

What remains the province of a specialised database

Research conducted from inside a matter is targeted by construction: it starts from a precise question and retains a limited number of genuinely relevant decisions. That is what makes it useful, and it is also its limit.

That limit should be stated by the provider rather than discovered by the user, and a system claiming otherwise would deserve suspicion.

It replaces neither exhaustive research across a complete corpus, nor a full editorial citator, nor systematic coverage of every court, nor an alerting service, nor the assurance that no relevant decision has been missed. Those needs are real, and they are met by tools built for them.

It should be added that those needs are not evenly distributed within an organisation. A litigation department and a transactional department do not have the same relationship to exhaustiveness, and the trade-off is better made team by team than globally.

The question to ask is therefore not what can be removed, but what proportion of one's research falls into each category. In most organisations, targeted research on a specific question dominates by a wide margin, and exhaustiveness is required on a small number of matters. That ratio, particular to each organisation, is what should drive the decision.