Skip to content
be-human.ai
← Back to the journal

Context & memory

Capturing Staff Know-How for AI Agents Before a Departure

A step-by-step routine for interviewing a departing expert, turning their procedures into reviewed records and loading them into company memory for an AI agent.

Be Human typographic cover with the title: Capturing Staff Know-How for AI Agents Before a Departure
Conceptual illustration created with AI · Be Human

To capture staff know-how for AI agents before a departure, interview the expert while they still work with you, turn each procedure they run into a short written record with its trigger, steps, exceptions and approval rule, and load only the records a named reviewer has accepted into the agent's company memory. The aim is not a transcript of everything the person knows. It is a set of procedures the agent can follow, cite and have corrected after the expert has gone.

Key takeaways

  • Start the capture as soon as notice is given, and schedule the sessions well before the final two weeks.
  • Interview the expert around real cases from the recent past rather than a generic job description.
  • Write each procedure as a record with a trigger, ordered steps, known exceptions, an owner and an approval boundary.
  • Route every captured fact through a reviewer before any agent can use it.
  • Test the agent against the cases the expert handled, and assign an owner to every record that passes review.

Why a departure leaves a gap the agent cannot see

When a specialist leaves, the company loses more than a person. Much of what makes a process work lives in habits: which customer always needs a phone call before a refund, which invoice format the finance team quietly rejects, which exception the former manager approved by word of mouth. None of this appears in the procedure document, because the document was written for the situation the process was designed for, not for the cases that actually happen.

An AI agent with access to shared files answers from whatever it can find. If the easiest thing to find is an outdated policy, it will confidently repeat that policy. The gap is not that the agent knows nothing. The agent knows a version of the process that the expert had already stopped following. The broader problem of tacit knowledge is covered in how to capture tribal knowledge for AI.

Plan the capture while the expert is still here

Decide the scope before the first conversation. Name the role, the decisions the agent will eventually support or prepare, and the records that already exist. Then work through this sequence:

  • Name the role and the two or three recurring tasks the agent would support.
  • List the systems, mailboxes, folders and channels where the expert's work actually lives.
  • Ask the expert's manager which decisions must stay under human approval.
  • Book two or three sessions of about an hour each, with at least a week between the first and the last.
  • Agree in writing who will review the output and by what date.

Spacing the sessions matters. The expert often notices a missing exception only after reading a draft record, so the second session should correct the first. Keep the notes from each session so the later conversations build on earlier ones instead of starting over.

Run the interview around real cases

Abstract questions produce polished answers that describe the official process. Real cases produce the working one. Ask the expert to pick the most recent cases that were hard, unusual or escalated, then walk through each one from first contact to close. Use the same questions for every case:

  • What triggered this case, and in what words did the customer or colleague describe it?
  • Which information did you check first, and where did it live?
  • What did you do that is not written anywhere, and why?
  • What would make you stop and escalate?
  • Who else needed to know, and what would they need to see?
  • Read back a one-paragraph summary of the case and ask what is wrong with it.

Record the session only if the expert agrees and company policy allows it, and treat the recording as material for drafting rather than as memory itself. Remove customer identifiers and personal remarks before anything is stored.

Turn each answer into a reviewable record

A transcript is hard to correct and easy to misread. Convert each procedure into a short record with fixed fields, so a reviewer can check one thing at a time. A useful record contains:

  • Trigger: the situation that starts the procedure, described in the words staff use.
  • Steps: the ordered actions, each with the system or source it touches.
  • Exceptions: the known cases where the usual path does not apply, with the rule for each.
  • Sources: the documents and versions the steps depend on.
  • Approval boundary: what the agent may do alone, what it may only prepare, and who must approve the rest.
  • Owner: the person accountable for keeping the record accurate after the expert leaves.

Keep the expert's exact wording where it carries a rule, such as a phrase that must be used with customers, and paraphrase everything else. Mark anything the expert was unsure about as unverified so the reviewer knows where to look first. Deciding who may save a fact into shared memory in the first place is a separate design question, explained in memory write gates for superintelligent agents.

Who approves and maintains the records

Records should not enter shared memory on the expert's word alone. Assign a reviewer from the team that will use the agent, and ideally a second person who can check the approval boundaries against company policy. The reviewer accepts, edits or rejects each record. Rejected items go back to the expert or are dropped. Accepted records carry the reviewer's name, a date and the version of the source they describe.

In a managed engagement, Be Human defines the record format for the role, agrees the approval boundaries with the company, and loads only accepted records into the agent's memory while its identity, tools and permissions stay under the company's control. Monitoring and permission review continue after the handover, so records that fall out of date are surfaced for review rather than trusted by default.

Test the captured know-how against past cases

Before the agent handles live work, replay the cases the expert handled. Give the agent the original intake for each case, not the expert's answer, and compare its proposed steps with what the expert actually did. Differences usually fall into three groups: the record is missing a step, the record is wrong, or the agent chose a path the approval boundary should have blocked. Each group needs a different fix, and the case should be replayed after the fix.

Illustrative scenario: a support lead who is leaving is asked to describe how a delayed order is handled for a customer on a long-term contract. The interview captures a waiver the lead applied from memory. The first replay shows the agent's draft charging the standard rate, because the waiver never made it into the record. The reviewer adds the exception with its condition and source, the approval boundary keeps waivers for human sign-off, and the same case passes on replay. Test design for recall and forgetting is covered in memory test cases for superintelligent agents.

Keep the records current after the handover

A captured procedure starts aging the day it is written. Schedule a short review of each record after the first real cases pass through it, and again whenever the underlying tool, policy or approval rule changes. Put the owner's name on every record so that a question from staff reaches someone who can answer it. Retire records that no longer match practice instead of leaving them beside the new version, because two competing procedures in memory are harder to correct than one. A routine for this upkeep is described in AI knowledge base maintenance: owners and a review routine.

If you are not yet sure which procedure from this team should be captured first, take the free Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

How long does a departure capture take? Most roles need two or three interview sessions spread over at least a week, followed by time for the expert to review drafts. Expect the review work to take longer than the interviews, because the reviewer needs to check each record against the real systems. Start as soon as notice is given, since a compressed schedule usually produces records with missing exceptions.

Can an agent learn the same things from the expert's old emails and chat logs? Old messages show what happened, not what was intended, and they often contain customer details and personal remarks that should never reach memory. Use them to find recent cases and to check what the expert says in interviews. Store only the procedures a reviewer has accepted.

Who should conduct the interview? Ideally someone who understands the role but was not the expert's daily manager, since that makes it easier to ask why a step exists. The interviewer should not also be the person who approves the records, so that the review stays independent. Hold the session in whichever language the expert uses most comfortably, whether English or Portuguese.

What if the expert leaves on short notice or will not take part? Use the artifacts they left behind, such as ticket histories, saved templates and approval threads, and interview a colleague who handled the same cases. Mark every record built this way as inferred so the reviewer treats it as a draft. The best protection is earlier: ask every team to keep its procedures in reviewable form while people are still in the role.

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.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →