Skip to main content
DoneThat

AI Adoption GuideHRExit

Knowledge capture agent

Summarizes a leaver's docs, decisions, and key contacts into a structured handover for the successor.

HR processPlanSourceSelectHireOnboardDevelopRewardExit

By Don, DoneThat’s AI coach · updated

A handover line is complete only when the leaver can open the cite

A capture draft is useful when each line points at a licensed document, decision record, or contact the leaver can still open. That is the quality bar. If the artifact is missing, the line stays empty. Empty is a valid result. A fluent sentence with no cite is not.

The manager still owns the handover the successor will use. The leaver still reviews. The agent drafts. It does not auto-publish, and it does not scrape unauthorized inboxes. Licensed wiki, tickets, and repo only.

Read a line the way the successor will. If they cannot click through to a wiki page, a ticket, a recorded decision, or a named contact that already exists in those systems, they cannot verify the claim on day one.

A line with no artifact cite is the first failure mode. Treat it as unfinished. Delete it or blank it. Do not keep it because it "sounds like what they did."

Load licensed wiki, tickets, and repo, nothing else

Load artifacts from systems the company already licenses and that the leaver already may access. In many organizations that class includes a wiki such as Confluence, a tracker such as Jira, a source host such as GitHub, and an HR system of record such as Workday for last day, role, and successor assignment. Those names are examples of a class, not a ranked list. Use the licensed equivalents you already run.

Collect pages the leaver owns or recently edited, tickets where they are assignee or reporter, pull requests and CODEOWNERS files in repos they maintain, and decision records already linked from those tickets or spaces. Bound the pull to that license. Do not attach a personal mailbox, a private chat history, or a drive the security team has not approved, even when those stores would make the pack look complete.

Licensed means the leaver's work account can open the object today in the company tenant. If they cannot open it, you cannot ask them to confirm it, and you should not cite it.

This job is not a skills inventory. skills inference from work artifacts asks what the person could do. Capture asks what the successor must pick up next week. Keep the pulls from colliding so a handover line does not quietly become a talent claim.

If Workday has no successor yet, still load the leaver's licensed corpus. The draft can wait for a named recipient. It should not wait to start citing material, because licenses often end on the last day.

Stop if the only remaining sources are unsanctioned. A thinner pack from Confluence, Jira, and GitHub is safer than a rich pack built from mail the leaver never agreed to process.

Draft cited lines and keep missing artifacts blank

Write each line as a successor-facing claim plus a cite. The claim is the work to continue. The cite is the wiki page, ticket, commit, or recorded contact that supports it.

One illustration: a platform engineer leaves on Friday. Do not write "Alex handles overnight jobs." Write that overnight batch recovery is documented in the Confluence runbook titled "Payments overnight job," which the leaver can still open in page history; that open cutover work sits on Jira epic PAY-4412; that GitHub CODEOWNERS for services/payments lists the two remaining reviewers. If the on-call escalation name exists only in a personal message, leave the contact line blank. Do not invent a person to complete the section. Inventing a contact is a failure mode. The successor will page the wrong desk, or nobody.

When a wiki page exists but records no decision, cite the page and leave the decision field empty. When a Jira ticket exists but has no linked decision record, cite the ticket and leave the decision field empty. When GitHub shows the leaver as the only CODEOWNER and names no deputy, cite that file and leave the deputy empty. The blank is the manager's queue, not a defect to hide.

Do not fill blanks from standup memory, from a laptop spreadsheet, or from "everyone knows." If it is not in the licensed corpus, it is not in the draft.

Group lines by successor job, not by vendor. The successor needs "how we recover the overnight job," not a Confluence export, a Jira export, and a GitHub export stacked in three dumps. Each group still cites back to those systems.

Run leaver review before the manager accepts

Give the draft to the leaver first, while their licenses still work. They are the only reviewer who can confirm that the cited page is the right page, that the ticket is still the open work, and that the named contact is still valid.

Ask them to mark each line confirmed, corrected, or still blank. Confirmed means they opened the cite. Corrected means they pointed at a different licensed artifact. Still blank means the corpus has nothing, and they agree it has nothing. A correction must not be a new undocumented claim.

If they say "use Maria instead" and Maria does not appear in CODEOWNERS, the wiki, or the ticket, keep the contact blank and flag the disagreement for the manager. Do not silently insert Maria.

A cite the leaver cannot open is not a cite they can confirm. If access already dropped, restore read access for the review window or drop the line.

Do not treat this pass as an ai-conducted exit interview. The interview holds reasons for leaving and what people will say in conversation. Capture is documentary. You can run both. Do not pour interview names into handover blanks. A spoken owner that never appears in wiki, tickets, or repo stays out of the pack unless the manager later adds a human-owned line.

Keep the manager as owner and the draft unpublished

The output is a draft until the manager accepts it. Treating the draft as published is the failure mode that turns a review artifact into the official handover. If the agent writes a Confluence page the team already watches, files a Jira ticket marked complete, or opens a GitHub issue the successor will see, you have published without an owner.

The manager accepts or rejects each line. They may add a line with no system cite. Label that line manager-authored so nobody reads it as artifact-backed. They may also ship blanks. A blank the manager has seen is an owned gap. A blank the agent filled with a guess is a defect.

Hold the draft in a review surface shared by leaver and manager, not in the space the successor already reads. The manager sends the pack. The agent does not.

A manager briefing generator can put status in front of the manager. It does not replace acceptance. Briefing is attention. Ownership is which lines the successor may trust.

If the seat will be backfilled later, do not feed this draft into an internal opportunity matcher. Matching remaining people to open work is a different workflow. Mixing it with exit capture pushes the draft to over-claim what the leaver really owned.

Stop conditions that protect the successor

Refuse a line with no artifact cite. Blank it or delete it. Prose without a page, ticket, decision record, or licensed contact is not a handover line.

Refuse auto-publish. The successor should not find the pack as a surprise wiki page.

Refuse invented contacts, invented decisions, and invented owners. If GitHub, Jira, Confluence, and Workday do not show the person, the field stays empty.

Refuse unauthorized stores. Licensed wiki, tickets, and repo only.

What the successor then receives is auditable. Each line either opens a real artifact the leaver already reviewed, or it is empty and the manager already knows the work was never written down. That is the quality bar on a last day. Anything smoother is usually a line you cannot defend.

Is this worth automating for you?

Whether this pays back depends on how much time it takes your team today. Most teams estimate that from memory, and the estimate is usually wrong in one direction or the other. This one is rated high effort to implement, so the baseline matters more than usual.

DoneThat reconstructs where the time actually went, with no timers to forget, so you can measure the baseline before committing to a project and check the gain afterward.

Measure the baseline first