AI knowledge management for firms: can this answer leave the firm?
By Loïc Guyon (Tileo), founder of AI Jungle · Updated · 8 min read
On this page
AI knowledge management must keep old notes from becoming current client facts. Retrieval quality is only the first test. Before generated text enters an external message, a reviewer should be able to trace the answer to its source, confirm the client or mandate it belongs to, identify the person who owns the fact, and see when that fact was last confirmed. We call these checks the Four Proofs: Source, Scope, Owner, and Freshness. If all four support the exact statement, use it. If the evidence conflicts, verify it. If a proof is missing, keep the statement out and ask the question that would supply it.

What is AI knowledge management in a professional-services firm?
Knowledge management covers the work of identifying, categorizing, storing, archiving, and disseminating information, according to Upland's category overview. Stravito defines AI knowledge management as using AI to capture, organize, discover, and activate knowledge, with functions such as semantic search, summaries, and recommendations in its overview of AI knowledge management.
That definition gets a firm to retrieval. A partner still has another decision to make: may this sentence appear in a client update, candidate message, deal note, or recommendation?
For firms, an answer becomes usable only when the reviewer can return to the source, scope, owner, and freshness before external use. This matters in executive search first, and the same test fits recruiting, M&A advisory, and consulting work. Each field keeps the answer tied to the work that gave it meaning.
The trap. Connecting more documents can improve recall while also creating more chances to retrieve a note from the wrong mandate. Better retrieval does not settle the external-use decision.
What are the Four Proofs?
The Four Proofs turn a generated answer into an auditable object. The reviewer should see the evidence beside the answer, not reconstruct it after the message has been drafted.
- Source. Show the exact record that supports the answer and the relevant passage. A search result or generated summary alone is not the source.
- Scope. Name the client, mandate, project, or other boundary to which the source applies. A useful note in one search may be unsafe in another.
- Owner. Name the person accountable for the fact within that scope. The author of a note and the owner of the mandate may be different people.
- Freshness. Show the date on which the fact was confirmed, plus any stated review condition. A document date tells the reviewer when someone wrote the record. It does not, by itself, say that every statement remains current.
The four fields work together. A source without scope can support a sentence but attach it to the wrong mandate. Scope without an owner leaves no one accountable for resolving a conflict. An owner without freshness may confirm who knew the fact, but not whether it still applies.
How do you choose use, verify, or block?
Run the exact proposed sentence through the table below. Do not grade the document as a whole. The decision belongs to the outgoing claim, because one paragraph can contain both usable and blocked sentences.
| Decision | Evidence state | What the reviewer does | What may enter the external draft |
|---|---|---|---|
| Use | All Four Proofs support the exact claim, with no unresolved conflict. | Approve the exact wording. | Only the approved wording. |
| Verify | The proofs are visible, but two records disagree or a field is ambiguous. | Ask the owner a question tied to the conflict, then update the evidence. | Nothing from the disputed claim until verification resolves it. |
| Block | At least one proof is missing, or the answer cannot be bound to the proposed claim. | Keep the claim out and ask for the missing proof. | Nothing from that claim. |
The detail that protects the decision. Store the proposed question with the blocked claim. "Which mandate is this for?" is more useful than a generic warning because it names the proof needed to continue.
A reviewer can apply the method in one pass:
- Copy the exact sentence proposed for external use.
- Open the cited source and find the passage that supports it.
- Fill the scope, owner, and freshness fields without guessing.
- Check for another source that changes or contradicts the claim.
- Record use, verify, or block for that exact sentence.
Can the Four Proofs resolve two conflicting mandate notes?
The following exercise is entirely fictitious: the firm, people, roles, dates, numbers, documents, statements, and results do not describe a client result.
A 12-person executive search firm is working on a CFO mandate for a logistics company. A partner, Elena Park, reviews a client update prepared with an AI agent. Two meeting notes disagree about the candidate profile:
- An October 14, 2026 note says, "Cross-border logistics experience is required." It identifies the meeting, but it does not state which mandate the sentence covers.
- An October 28, 2026 note says, "Cross-border logistics experience is preferred, not required." It names the CFO logistics mandate as its scope and identifies the mandate owner, Marcus Reed.
The agent proposes this sentence: "The client requires cross-border logistics experience." Elena does not treat the newer note as an automatic winner. The older sentence may describe another mandate, or it may record a requirement that later changed. The Four Proofs expose the actual problem: the older record has a source, but its scope is missing.
Example decision. Elena marks the proposed sentence block. She asks the mandate owner: "For the CFO logistics mandate, is cross-border logistics experience required or preferred as of October 29, 2026?"
In this exercise, Marcus replies on October 29, 2026: "Preferred, not required, for the CFO logistics mandate." That confirmation supplies the scope, owner, and freshness for the answer. It also resolves the conflict without pretending the older note applied to this mandate.
Elena then reviews the revised wording. The approved final text in this exercise is exact:
For the CFO search, cross-border logistics experience is preferred, not required, as confirmed by the mandate owner on October 29, 2026.
The decision changes from block to use only after the new source supports all four fields. The agent did not settle the disagreement. It exposed the missing proof, proposed a precise question, and waited for a person to decide.
How do you audit a sample answer yourself?
Use the filled card as a model. The contents remain part of the exercise above. A complete card should let another reviewer repeat the decision without relying on the generated answer's confidence.
Copy this. Filled decision card for the CFO logistics exercise.
Answer / For the CFO search, cross-border logistics experience is preferred, not required, as confirmed by the mandate owner on October 29, 2026.
Source / October 29, 2026 confirmation from Marcus Reed, prompted by the conflict between the October 14 and October 28 meeting notes.
Scope / CFO search mandate for the logistics client.
Owner / Marcus Reed, mandate owner.
Freshness / Confirmed on October 29, 2026.
External-use decision / USE. Elena Park, partner, approved the exact answer above in this exercise.
Now copy the blank version and test a fictitious dossier. If a field cannot be completed from the dossier, do not fill it from memory. Record the missing proof and choose the corresponding decision.
Answer /
Source /
Scope /
Owner /
Freshness /
External-use decision / USE / VERIFY / BLOCK
Missing proof or conflict /
Question for the owner /
Exact wording approved /
Where does approval fit in a shared workspace?
Pitstop is the workspace where your team and your AI agents run the firm together. People and agents work on the same records across areas such as projects, clients and CRM, network, meetings and prep briefs, content, knowledge, finance, and validations. See the human-in-the-loop approval explanation for the related operating principle.
The operating principle is: "Nothing leaves without your yes, and you can prove it." A message stops for a person's approval bound to the exact content. If the text changes, that approval no longer applies. A hash-chained journal records the request and the decision. Agents may come from ChatGPT, Claude, or Codex through MCP. Pricing is per firm, never per seat.
The Four Proofs explain what a reviewer needs to judge; exact-content approval records which wording the person accepted.
FAQ
Is retrieval accuracy enough for AI knowledge management?
Retrieval accuracy does not make an answer safe for external use. Retrieval can find a relevant passage without proving that the passage belongs to the current mandate, has an accountable owner, or remains current. Apply the Four Proofs to the exact claim before external use.
What happens when one proof is missing?
Choose block for that claim. The agent may propose a question that would supply the missing field, but the disputed text stays out of the external draft until a person resolves it.
When should a reviewer choose verify instead of block?
Choose verify when the evidence is present but ambiguous or conflicting. Verification is the work of resolving that conflict. If a proof is absent, choose block.
Does a newer note always replace an older note?
No. A newer date does not prove that both notes share the same scope. First bind each note to its mandate, identify the owner, and ask whether the newer statement changes the older one.
Can this method work outside executive search?
Yes. The fields can be attached to an outgoing claim in recruiting, M&A advisory, or consulting. The scope label changes, but the review question stays the same: can the exact answer be traced and approved for this work?