The firm knowledge base: stop answering the same question every week
A second-year associate asks the group chat how the firm handles a particular filing. Two people answer within the hour and they do not agree. A partner settles it from memory. The real answer is in a memo written three years ago by someone who has since left, saved in a folder named after a matter number nobody remembers.
The same question will be asked again in about six weeks. The firm knows the answer. It cannot get to it.
What a firm knowledge base is
A knowledge base is the firm’s own written material, indexed so that a question in plain language comes back with the answer and the document it came from.
The material is what already exists: procedures, checklists, precedent documents, prior memos, filing guides, the notes a partner wrote for a training session, the email explaining why the firm stopped doing something a particular way. Most of it is written already. It is spread across a document management system, a shared drive, three inboxes and one person’s desktop.
Four properties do the work.
It answers only from what the firm put in it. Ask about a filing deadline and the answer comes from the firm’s procedure for that filing, not from what a general model has read on the subject. Where the firm’s practice differs from the textbook, and it usually does, the firm’s practice is what comes back.
Every answer carries its source. The answer is a starting point. The document is the authority. Anyone who needs to rely on it opens the source and reads the paragraph themselves. Nobody is asked to trust a paraphrase.
It says when it does not know. A question the material does not cover returns nothing rather than a plausible guess. That behaviour is the whole system, and it is the first thing worth testing: ask it something the firm has never written down and watch what happens. If it invents an answer, it is not usable in a professional practice.
Access follows the folder. A document nobody outside the tax team can open does not become an answer for the rest of the firm. Permissions come across with the material rather than being rebuilt by hand.
Nothing leaves the firm’s environment, and client files are never used to train anything.
Does a firm your size need it
The test is repetition. How often does someone answer a question they have answered before, and how long does it take the asker to find it alone?
A firm where everyone has been there eight years does not need this. The knowledge is distributed and the answers are quick. The picture changes with hiring. Every new person spends their first months asking questions that have documented answers, and every question costs two people’s attention rather than one.
The second signal is a single point of failure. If one person is the firm’s institutional memory, the firm has a risk it has never priced. That person goes on holiday and work slows. That person leaves and twenty years of practice goes with them.
The third is silent divergence. Two associates producing the same filing differently, because each learned it from a different colleague, is the same problem as the missing memo. The knowledge exists. The routing is broken.
The pattern holds up at volume. In one archive project, 103,382 files on a firm’s drive were indexed and cross-linked to their case files, and 4,517 appearing parties and 2,182 act types were extracted into structured records. That was a back catalogue nobody could search, turned into one that answers questions.
There is one case where the answer is honestly no. If the firm has never written its procedures down, there is nothing to index. That is not a failed project, it is a different first phase: writing down what the firm does. Indexing an empty shelf produces a system that confidently knows nothing.
What it costs
The knowledge base is built as a phase in a plan, at a fixed price agreed before work starts. The number depends on how much material there is, what state it is in, and how many places it lives in. A firm with a document management system and a written procedures manual is a smaller build than one with fifteen years of shared drive.
That is why this work starts with the audit. The audit maps what the firm has, prices the build as a fixed phase, and credits its own fee against that phase if the firm goes ahead. The plan belongs to the firm either way.
Running costs are model usage, which is low for question answering, and storage, which is negligible.
The cost that gets underestimated is curation. Not everything on the drive should be in the knowledge base. Superseded procedures, drafts that were never adopted and the version of a checklist from before the rules changed are all worse than nothing, because the system will answer from them in good faith. Someone at the firm has to decide what is current. That work is unglamorous and it determines whether the answers are right.
What the first week looks like
The first week starts with questions, not documents.
We collect the questions the firm actually asks. The group chat history is the best source. So is the person who gets interrupted most, who can usually list the top twenty from memory. Those questions become the acceptance test, written down before anything is built.
Then the material is traced backwards from the questions. Which document answers this one? Is it current? Where does it live? This is where the gaps surface, and they are useful findings on their own. A question that no document answers is either undocumented practice or a procedure that needs writing.
A narrow slice gets indexed first, usually one practice area or one process end to end. Narrow and correct beats broad and unreliable, and it gives the firm something to judge in days rather than months.
Then the staff who asked the questions ask them again, of the system. Every answer is checked against the source. A thin or wrong answer is almost always a gap in the material rather than a fault in the search, and the fix is to correct the document, which improves the firm whether or not the system exists.
Go-live waits for the people who answer these questions today to agree the system answers them properly. That rule holds in every build. Systems that staff route around are worse than no system.
The full scope is on the service page: /services/drafting-knowledge. To ask a fictional firm’s procedures manual a question and see the sources it answers from, the demo is here: /demos/knowledge.