Legal knowledge management: turning research into a compounding asset

In the traditional model, research is an event. A question arises on a matter, someone researches it, an answer is produced, it is used, and it is filed in the matter folder. The matter closes. Eleven months later the same question arises on a different matter and is researched again from zero, frequently by a different person who does not know the first answer exists.

The cost of this is enormous and almost entirely invisible, because nobody is counting. Ask a practice how many times the same question was independently researched in the last year and the honest answer is that they have no idea. Firms that start measuring are consistently surprised by what they find.

In the stock model, every answer is an asset with a life beyond its originating matter. It is filed by question rather than by matter, tagged, dated, and updated when the underlying law moves. The question that arises eleven months later starts from the existing answer and asks only what has changed. The practice gets faster every year it operates, and that improvement lives in the institution rather than in individuals who eventually leave.

A stock is only worth building if it is consulted, and consultation has to be a rule rather than a good intention. The rule: before researching any question from scratch, check the stock. Every time. Including when you are confident there is nothing there, because that confidence is exactly what keeps the same question being answered three times.

The corollary is equally important: when the stock is checked and comes up empty, the answer produced goes into the stock. Without that, the stock only ever contains what someone remembered to contribute, which is a small and shrinking subset of what the practice actually knows.

This is the same structural pattern as the evidence-anchoring rule - a small, absolute, cheap-to-follow rule that converts an individual habit into an institutional property. Practices consistently find these rules easier to sustain than large process programmes, precisely because they are small enough to actually follow on a busy day.

There is an important asymmetry in what language models do well. They are extraordinary at retrieval and synthesis over a well-structured body of material, and they are unreliable at invention over nothing at all. Almost every legal AI failure story is an instance of the second. Almost every quiet success is an instance of the first.

A curated research stock is precisely the well-structured body that makes the reliable mode possible. An agent asked 'what is our position on this' against a tagged, dated, anchored stock retrieves the prior answer, surfaces the authorities, and flags what has changed since - a task it performs well and verifiably. The same agent asked the same question against nothing generates a plausible answer with fabricated citations.

This is why knowledge management, long treated as a low-status administrative function, has become one of the highest-leverage investments a practice can make. The stock is not merely a filing improvement. It is the substrate that determines whether AI in your practice operates in its reliable mode or its dangerous one.

A stock that might be out of date is a stock nobody trusts, and an untrusted stock is not consulted, at which point it decays to zero value regardless of how much is in it. Currency is therefore not housekeeping - it is the thing that determines whether the asset exists at all.

One quarterly pass is generally sufficient. Sort entries by age and by how often they have been retrieved, and review the top of that list. For each: is the law still as stated, and is the answer still the answer? Update, or mark superseded with a pointer to the current position. An entry marked 'superseded, see X' is more valuable than a silently stale one, because it tells the reader something true.