Superintelligent Agents for Executive Search Firms: Candidate Notes
Set confidentiality rules before an AI agent reads search files, then use it for candidate notes, client feedback summaries and checkable status updates.

Superintelligent agents for executive search firms are AI agents that read candidate notes, client feedback and search status records so consultants can find context quickly and report progress accurately. They only work well when confidentiality rules, memory boundaries and approval steps are set before the agent opens any file. Be Human designs these agents around those rules first, so the firm decides what the agent may remember, what it may draft and what must always pass a partner's review.
Key takeaways
- Define confidentiality rules before an agent reads any candidate file, reference note or client document.
- Keep candidate notes, client feedback and search status in separate memory scopes with different access.
- Let the agent draft summaries and status updates, while a consultant approves anything that leaves the firm.
- Choose a success criterion you can check, such as whether each status statement traces to a source entry.
- Review permissions on a schedule, because mandates close, consultants change roles and client access shifts.
Why candidate notes are the hardest memory to hand over
Candidate notes are among the most sensitive records a search firm holds. They contain candid assessments, compensation expectations, reasons a person wants to move, and personal details that a candidate shared in confidence. The firm's value rests on judgment and discretion, so the question is not only whether an agent can read these notes, but whether it should keep what it reads, and for how long.
An agent's memory is not a neutral filing cabinet. If a remark from one mandate can surface in a conversation about a different client, the firm has a confidentiality problem even if no one intended it. Personal data rules also apply in most markets where executive search operates, including the data protection laws in Brazil and the European Union. A practical starting point is to treat candidate records as restricted by default and to open access only when a specific task requires it.
Setting confidentiality rules before any agent reads a file
The order of work matters. Rules written after the agent has already ingested a shared folder are much harder to enforce, because the data may already be copied into places nobody tracks. Before connecting any source, the firm should agree on the following points and write them down where the operations team can review them.
- Name each class of record: candidate notes, reference calls, client feedback, mandate briefs, invoices and internal coaching notes.
- For each class, decide which mandates and which roles may let the agent read it.
- Mark which details must never be stored in agent memory, such as health information, family circumstances and unverified allegations.
- Set a retention rule for each memory scope, including when it is cleared at mandate close.
- Decide which outputs may be sent to a client, which may only go to an internal partner, and which must not leave the firm at all.
- Specify who approves an outbound message and how that approval is recorded.
- Require an access log for every read and write, so a partner can answer who saw what and when.
This list is short on purpose. Its purpose is to force decisions that otherwise get made by default, usually by whoever connected the tool first.
Candidate note memory: what the agent keeps and what it lets go
The safest pattern is to avoid a single, global memory of everything the agent has ever seen. Instead, give the agent a working context for one mandate. When the mandate closes, that context is cleared or archived under the retention rule. Anything the agent needs to answer a question should be retrieved from the source record at the time of the request, not baked permanently into the agent's behavior. This matters because a retrieved answer can be traced back to its source note, and deleting the note removes the answer with it.
Each summary the agent produces should carry a reference to the note it came from, such as a record identifier and the date the note was written. A consultant reading a summary should be able to open the original in one step. If the agent cannot point to a source, the summary should be treated as unverified and not used in a client conversation. Keep the agent's answers short and specific, so that a reviewer can check them against the notes in a few minutes.
Client feedback summaries that protect candidate identity
After a shortlist interview, clients often send feedback that mixes praise, objections and questions about scope. An agent can organize that feedback into themes, such as strengths relative to the brief, concerns about experience, and open questions about timing or compensation. The important design choice is what the summary reveals. A client summary should usually describe the candidate by the reference the client already holds, and it should not pull details from other candidates' files into the same message.
Candidate files also contain text written by people outside the firm, and that text can include instructions aimed at a model, such as a line telling the reader to ignore earlier guidance. The OWASP guidance on prompt injection prevention describes why an agent should treat document contents as data to be summarized, not as commands to be followed. In practice, the agent's role and tools should be defined in the agent's setup, not in the text it reads, and any action outside the summary should require a separate approval.
Illustrative scenario: A partner asks for a feedback summary on a shortlist of four candidates for a finance leadership role. The agent groups the client's comments by criterion, cites the feedback entry for each comment, and lists two open questions the client has not answered. It does not mention compensation figures from a candidate's file, because the client did not ask about them. The partner edits one sentence and sends the summary.
Search status updates a partner can verify
Status updates are where errors become visible to clients. A useful update states the stage of each candidate, which references are outstanding, when the next client meeting is planned, and what decision the client must make. The agent can draft these from the search tracker, but the draft should be built from the tracker's current entries, not from memory of earlier conversations, because stages change quickly.
A workable success criterion is that every factual statement in an update traces to a tracker entry, and that a consultant reviewing a random sample of updates each month finds no statement that cannot be traced. This is a measurable check, and it gives the operations lead a way to decide whether the agent should expand to more mandates or be narrowed. Recurring evaluations of this kind are a standard part of how agent systems are kept reliable, as the Anthropic guidance on evals for AI agents describes in general terms.
Permissions, access and audit
Access control should follow least privilege. The agent should be able to read the one mandate folder it is working on, not the whole candidate database. Write access should be narrower still, often limited to drafts stored in a review queue rather than direct edits to records. The OWASP authorization guidance is a useful reference for checking that each permission is granted by an explicit rule rather than by convenience.
Permissions need periodic review. When a mandate closes, its scope should be removed. When a consultant changes teams, their agent access should change with them. When a new integration is added, such as a calendar or an email tool, it should be checked against the confidentiality rules before it is connected. A simple monthly review, with the access log in hand, catches most drift before it becomes an incident.
- Review which mandates each agent scope can read, and remove any that have closed.
- Confirm that outbound messages still route through a named approver.
- Check the access log for reads outside the expected mandate.
- Record any incident, including a summary that contained a detail it should not have, and update the rules.
How Be Human's managed approach handles this
Be Human begins with one useful role for the firm, such as drafting client feedback summaries for active mandates, and defines the memory scopes, tools, approval boundaries and success criterion before the agent reads anything. The managed service then covers monitoring, recurring evaluations of the summaries against their sources, permission reviews at each mandate close, integration maintenance and incident response, with controlled changes made only after the operations lead approves them. The work is delivered remotely, in English or Portuguese, and the firm keeps control of what the agent may remember and send. If you are unsure which of your candidate, client or status tasks should go first, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.
Frequently asked questions
Should an agent be allowed to read every candidate file in the firm? No, and most firms should not start there. Give the agent access to the single mandate it is working on, and keep restricted categories such as health or family information out of its memory entirely. Broader access can be considered later, once the access log and monthly review show that the narrow setup works.
Can client feedback be summarized without revealing details from other candidates? Yes, if the summary is built from the feedback for one candidate and the rules forbid pulling in other files. The summary should describe each candidate by the reference the client already has. A partner should still review the message, because the client's own wording can sometimes reveal more than intended.
What happens to the agent's memory when a search closes? The memory scope for that mandate should be cleared or archived under the retention rule the firm agreed before the work began. Answers that depend on the candidate's notes should be retrieved from the source records, so removing the source removes the answer. The closing step should also remove the agent's access to that mandate.
Who approves a status update before it reaches a client? A named consultant or partner should approve every outbound update, and the approval should be recorded. The agent can draft the update from the current tracker entries, but it should not send anything directly. This keeps accountability with people who can answer the client's questions.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


