AI knowledge base for consultancies: what goes first
5 minutes read
Most firms start a knowledge base by loading everything they have, and then get poor answers to every question. A small library of the right material, added in the right order, is useful from the first week.
An AI knowledge base for consultancies should start with the firm's methods and standards, because they are reused on almost every project, change slowly, and belong to the firm outright. Proposals and project write-ups come second. Raw project folders, email, and chat come last, and many firms never need to add them at all.
The common mistake is to begin with volume. Loading every file the firm has feels thorough, and it produces a library where a superseded draft counts for as much as the approved version. An AI tool reading that library will quote whichever one it finds.
Why starting small works better
A small library is better because you can trust every entry in it. Twenty documents that someone has read and approved will improve a draft. Twenty thousand that nobody has looked at since they were saved will mostly add noise, and occasionally add something wrong.
A small start also makes the first result easy to judge. When a draft cites one of twenty known documents, the consultant reading it knows the source and can tell at once whether it was used well.
What to add, in order
Add material in the order below. Each step is useful by itself, so there is no need to finish the list before people start working.
| Order | What to add | Why it comes here |
|---|---|---|
| 1 | Methods and frameworks the firm uses | Reused on most projects and owned by the firm |
| 2 | Quality standards and house style | Shapes every draft, and is short enough to approve in an hour |
| 3 | Recent proposals that won | Shows how the firm describes its own work |
| 4 | Project write-ups with client detail removed | Keeps the lessons and leaves out the client material |
| 5 | Research notes and reading lists | Useful, but dates quickly and needs an owner to keep current |
Stop after step two if that is all the time available. Methods and standards alone change the structure and tone of what comes back, which is the part reviewers spend most time correcting.
What to leave out
Leave out anything nobody would want quoted. That covers drafts that were replaced, internal debate about a client, personal notes, and old versions of documents that have a current one.
Client material is a separate case. It should not be in the firm-wide library at all. It belongs with the client's own project, readable only by the people who can already open it, and used only on that client's work. The connected knowledge guide explains the split between what the firm owns and what a client owns.
Give every entry an owner
Each entry needs one person who approves it and whom people ask when it gets out of date. Without an owner, a library fills with documents that were right when they were added and that nobody has checked since.
The owner's job is light. They confirm the entry is the current version, that it contains nothing that should not be reused, and that it still reflects how the firm works. Anything not yet approved can still be in the library, as long as it is marked as a draft so readers and tools treat it accordingly.
An owner also settles arguments. When two versions of a method exist, the owner decides which one the firm uses. The library then has one version, and the disagreement does not come back in every draft.
How to tell whether the library is useful
The test of a knowledge base is whether drafts cite it and whether the citations are right. After the first few weeks, look at a handful of drafts and check three things: that firm entries appear in the sources, that they were the right entries for the task, and that reviewers are correcting structure and tone less than before.
If entries are never cited, either the library has the wrong material or its entries describe that material badly. If the wrong entries are cited, two of them overlap and one should go. Both are cheaper to fix at twenty entries than at two thousand.
How this looks in Hebno
Hebno keeps the firm's methods, standards, and notes in one library. An owner approves each entry, and unapproved entries are marked as drafts. People can ask the library a question in plain language, a research run pulls in the passage it needs, and every answer cites its entry. Client files are connected separately, under the permissions they already have. The knowledge feature page shows the library.
For how the library fits with the firm's other tools, see connecting your firm's tools and data to AI.
