Knowledge management for a small consulting firm: a realistic method
In a small firm, knowledge capture only sticks if it happens during the engagement, not after it. Write down decisions and their rationale, reusable deliverables, pitfalls and contacts as you go, then close with a thirty-minute debrief. An AI assistant with access to that folder can draft these notes and the summary, and reread them when a similar engagement starts, as long as each client is kept separate.
In a firm of three to fifteen consultants, everyone agrees on “capturing what we learn from each engagement.” Then the next engagement starts, the deliverable goes to the client, and what the team learned stays in the heads of the people who were there. This guide explains why classic knowledge management struggles in small firms and offers a light method: record things during the engagement rather than after, close with a short debrief, and get help from an AI assistant without mixing clients.
Why knowledge capture fails in small firms
The wiki nobody feeds
The most common reflex is to open a wiki or a shared “knowledge base.” A few weeks later, the pages stop changing. This isn’t specific to small firms. Jonathan Grudin and Erika Shehan Poole, at Microsoft Research, studied thousands of enterprise wikis: they found many thriving wikis, but those remained a minority. Ofer Arazy and Arie Croitoru analyzed the activity of more than 33,000 wikis in a large organization over five years: most of them went inactive fairly quickly.
In a small firm, the cause is simple: feeding the wiki is a separate, non-billable task done after the fact. When the workload rises, it’s the first thing to go.
Knowledge stays in people’s heads
The most valuable part of an engagement isn’t the final deliverable, which the client owns anyway. It’s everything around it: why a scenario was ruled out, which person at the client actually gets things unstuck, which metric turned out to be misleading. Ed Yourdon put it this way in a Computerworld article: decision makers sometimes document what they decided, but rarely why. And the why is what helps on the next engagement.
Michael Nygard, who popularized “decision records” in software, describes the two risks when that context is missing: blindly accepting a decision that has become obsolete, or reversing it without understanding what justified it.
Engagements that end without a debrief
The debrief is often planned for “the end.” At the end, the consultant is already staffed elsewhere, the client has a new point of contact, and memories have faded. That’s why Yourdon recommends regular “minipostmortems” covering a few weeks of work, while the key people are still around and remember what they did and why.
Large firms have knowledge managers. A Consultor article also describes how a firm of about twenty consultants appointed people to spot, at the end of each project, what was worth keeping. A five-person firm doesn’t have that role: the method has to work without it.
A realistic method: record during the engagement
The principle: write little, often, while the information is fresh, in a fixed format. Four note types are enough.
| Note type | When to write it | What it contains |
|---|---|---|
| Decision | The day the decision is made | The decision, the context, the options ruled out and why |
| Template deliverable | When a deliverable is approved | The reusable template, without client data, and what worked well |
| Pitfall | As soon as you hit it | What happened, how to spot it earlier, how to avoid it |
| Contact | After a first useful exchange | Actual role, topics where the person is decisive, working preferences |
Decisions and their rationale
This is the most useful note and the one most often forgotten. Borrow Nygard’s format, cut down to a few lines:
- Decision: one sentence (“We limit the assessment to the three pilot sites”).
- Context: what was true at the time (budget, deadline, client constraint).
- Options ruled out: what was considered and why it wasn’t chosen.
- Consequences: what it means for what comes next.
Five lines written the same day beat a page rebuilt three months later. Nygard makes the point: long documents never get kept up to date; short, separate notes do.
Template deliverables
For each approved deliverable, ask one question: “Would I build this document again for another client?” If yes, keep a version stripped of client data (assessment framework, interview guide, steering committee agenda) with two lines on what worked well.
Pitfalls
A recorded pitfall saves a colleague a week. Stick to facts: “The client’s HR software exports truncate department names at 20 characters; check before matching against payroll.” No judgments about people.
Contacts
The address book already lives in the CRM. The contact note says something else: who really decides, who needs to be convinced before the committee meets, who prefers a short email to a meeting.
A short debrief at the end of the engagement
If notes were taken along the way, the debrief becomes a thirty-minute review. Keep the questions fixed:
- What would we do exactly the same way?
- What would we do differently, and when could we have seen it?
- Which deliverables are reusable, and in what form?
- Which decisions should be reread before an engagement of the same kind?
The result fits on one dated page, filed with the engagement. On a long engagement, also hold lighter checkpoints along the way, as Yourdon suggests.
What AI changes
The method above runs into one problem: even short, writing is still a chore. That’s where an AI assistant helps, on one condition: it needs access to the engagement context.
- Drafting notes: from meeting minutes or an exchange, the assistant proposes the decision note in the right format. The consultant fixes the rationale, which only they know, and approves it.
- Keeping the summary: at the end of the engagement, the assistant rereads the notes and drafts the debrief. The thirty-minute session is then spent correcting rather than writing.
- Rereading before a similar engagement: when an engagement of the same kind starts, the assistant rereads the decisions, pitfalls and template deliverables from past engagements and turns them into a list of points to watch.
The hard part is access to context. A standalone conversation with an assistant starts from scratch. Claude Projects, for example, let you upload documents and instructions that serve as context for every conversation in the project; they’re available on all accounts, and sharing them with other members requires a Team or Enterprise plan, according to Anthropic’s help center. Another route is MCP, an open standard for connecting AI applications to external data sources and tools: the assistant then reads and writes in a space maintained by the team.
Whatever the setup, keep the rule: the AI proposes, a human approves what goes into the firm’s memory.
Confidentiality between clients: one folder per engagement
Knowledge capture must never mean passing one client’s information to another. The Consultor article shows the issue weighs on large firms too, where consultants spend time anonymizing what they share. A few simple rules:
- One folder per engagement, with its own access. An invited client, or a subcontractor, sees only theirs.
- Two levels: the confidential engagement folder, and a firm-wide shared library that holds only anonymized material (template deliverables, generic pitfalls, methods).
- Moving something from one to the other is a deliberate act, done by a consultant during the debrief, not an automatic copy.
- The same separation for the AI assistant: while working on engagement A, it should see only folder A and the shared library. Check what the tool lets you restrict before you open your folders to it.
Example timeline (fictional case)
A six-consultant firm runs a ten-week procurement reorganization for a small manufacturing company.
- Week 1: the engagement folder is opened. After the kickoff meeting, the assistant drafts from the minutes two contact notes (the CFO makes the calls; the procurement manager knows the suppliers) and one decision note: scope limited to non-production purchasing, because the client is renegotiating its raw-materials contracts in parallel.
- Week 3: a consultant discovers that the supplier master list contains many duplicates. She dictates a three-line pitfall note.
- Week 5: twenty-minute checkpoint. The assistant presents the notes from the period; the team adds a forgotten decision (dropping a benchmark judged too time-consuming).
- Week 8: the client approves the buyer interview guide. A version with no names or figures goes into the shared library.
- Week 10: the assistant drafts the debrief from the notes. Thirty minutes of review by two people, one approved page.
- Six months later: a partner starts a procurement engagement at another small company. The assistant, which can only access the new folder and the shared library, rereads the interview guide, the duplicates pitfall and the template decisions, and proposes a list of questions for the kickoff meeting. Nothing from the first company shows up.
With Tavimi
Tavimi is the shared working memory for a team and its AI assistants, over MCP. For a consulting firm, it follows the method described here: one space per engagement, where you invite the relevant consultants by email, and the client if needed; each person sees only the spaces they’re invited to. Notes are organized into categories (decision, pitfall, reference by default) and each category enforces its required headings: a decision note without its rationale is rejected. Notes are versioned; nothing gets overwritten.
The assistant (Claude Code, Claude on claude.ai, Claude Desktop or any MCP client) reads and writes in the space, asks its questions on a task when it’s missing a decision instead of guessing, and consolidates notes into a summary document. When you connect an assistant, you check the spaces it can see, which extends the per-client separation. Tavimi doesn’t replace your practice management software: it tracks neither time nor billing. To see how it fits a consulting firm, visit the consulting firms page or read how it works.
FAQ
What people ask us
Does a five-consultant firm need a dedicated knowledge management tool?
Not necessarily. One folder per engagement, a few templates for short notes (decision, pitfall, contact) and a dated debrief are enough to start. The tool matters less than the habit of writing during the engagement and how easily you can find what was written.
How long does an end-of-engagement debrief take?
Thirty minutes is enough if notes were taken during the engagement: you reread what exists and answer a few fixed questions. If everything has to be rebuilt from memory, the session runs longer and the memories are less reliable.
Can you reuse one client's deliverable for another client?
You reuse the structure, the method and the questions, not the client's data or conclusions. The simplest way is to keep an anonymized version of the template deliverable, separate from the original engagement folder.
Can AI do the knowledge capture instead of the consultants?
It can draft the notes, keep the log and produce the summary if it has access to the engagement context. But what is worth keeping, and why a decision was made, is something only the consultants know: the AI proposes a text, a human approves it.
Sources
- Wikis at Work: Success Factors and Challenges for Sustainability of Enterprise Wikis (Grudin et Poole, Microsoft Research, 2010), accessed October 3, 2026
- The sustainability of corporate wikis: A time-series analysis of activity patterns (Arazy et Croitoru, ACM TMIS, 2010), accessed October 3, 2026
- Minipostmortems (Ed Yourdon, Computerworld), accessed October 3, 2026
- Documenting Architecture Decisions (Michael Nygard, Cognitect, 2011), accessed October 3, 2026
- Knowledge management dans le conseil : gain ou perte de temps ? (Consultor, 2015), accessed October 3, 2026
- What are projects? (aide Claude, Anthropic), accessed October 3, 2026
- What is the Model Context Protocol (MCP)? (modelcontextprotocol.io), accessed October 3, 2026