AI-conducted exit interview

An AI-conducted exit interview is an optional, adaptive chat with a leaving employee. The bot asks follow-up questions. HR owns the conversation: what was asked, what was skipped, what is stored, who may read it, and whether anyone follows up.

Do not treat the thread as a decision. A fluent chat is still a source file. It is not a regrettable-attrition label, not a manager briefing, and not a root-cause report. If a later regrettable-attrition classifier runs, it runs on fields HR approved after review. It does not run inside this conversation, and tone in the chat is not a class.

Workday, Culture Amp, Qualtrics, and Lattice sit in the same class of systems many companies already use for employee records, surveys, and listening programs. This page does not rank them or assume a named module. Store the invite, the consent flag, and the sealed transcript in the system of record you already trust. The bot is the interviewer. The HRBP or people-ops lead remains the owner.

Two failure modes show up before anyone has finished reading. A summary quote with no pointer to a transcript turn is not evidence; drop it. A model-written sentence in quotation marks, as if the leaver said it, is fabrication. Delete it. Do not replace it with a sharper complaint that merely sounds like them.

Send the invite from HR, not from the line manager, and not from the bot speaking as if it were a colleague. Say that a conversation agent will ask follow-ups. Say they can decline any question. Say declined items stay blank. Say HR will read the transcript before any follow-up. Say nothing is auto-shared with their manager.

Keep the invite short enough to finish in one screen. Purpose: understand the leave so the company can fix what is actually fixable. Boundary: this is not a performance discussion and not a negotiation. Skip path: they can ignore the link; offboarding continues either way.

Send it after the resignation is accepted, not in the meeting where they resign, and not in the last hour of the last day. They need a private window. A same-day ping from a bot, with no HR name on the invite, feels like a trap.

Do not bolt this invite onto a systems dump or a manager stay conversation. Those are different jobs with different consent. If you need owners, runbooks, and in-flight work captured, run a knowledge capture agent as a separate pass.

If they never open the chat, there is no interview. Do not invent a narrative that they seemed disengaged to fill the hole. Do not copy themes from the last three exits on that team into this person's file.

A follow-up fires only when it can quote or tightly paraphrase a specific prior turn and point at that turn. If the probe cannot show its cite, it does not send. Store the cite on the probe record as a parent turn id. If the id is missing, the probe is invalid even when the wording sounds related.

Here is one illustrative path, not a measured case. A senior IC writes that they are leaving because "roadmap calls kept landing in Slack the night before a launch." A valid probe names that sentence and asks one thing: what would have made those calls land earlier, and who they expected to make them. An invalid probe asks whether they stopped trusting leadership, without citing the Slack line. An invented probe is worse: it claims "You said engineering leadership ignored you" when they never said that.

If they answer the cite-backed probe, store the new turn with a parent pointer to the cited turn. If they decline, skip, or write that they would rather not discuss it, the follow-up field stays empty. Empty stays empty. Do not backfill from the offer letter, from a compensation band, or from what other leavers in that org said.

Compensation, manager conflict, harassment, and anything that could become a legal claim are decline-friendly by default. The bot may ask once, and only with a cite if they already raised the topic. It does not re-ask to complete the record. Completeness is not a reason to overwrite a blank.

Free-text survey boxes cannot ask the second question while the first answer is still in the room. Adaptive chat is an improvement only when the second question is visibly grounded in the first. Ungrounded curiosity reads as interrogation, and people stop answering.

The HRBP or a designated people-ops reviewer reads the thread before any excerpt, theme tag, ticket, or forward exists. Review is a checklist, not a gut read.

Confirm every probe cites a prior turn. Confirm blanks are still blank where they declined. Confirm every sentence in a summary is copy-pasteable from a cited turn, or is clearly labeled as HR paraphrase with the raw turn attached. Confirm nothing has been routed to the manager.

If a summary uses quotation marks and the words are not in the transcript, delete the quote. Do not polish their wording into a sharper complaint. Polishing is how invented quotes sneak in while still feeling true.

Do not auto-share with the manager. If the leaver asked that their manager not see the detail, that constraint survives review. If they did not object, HR still decides whether a theme (not a dump of the chat) is useful, and HR is the one who sends it. The bot does not CC anyone.

Do not treat the chat as classified regrettable. Three paragraphs and a tired tone are not a regrettable flag. High performance on a prior cycle is a different data join under a different control. Mixing the two in this workflow trains the company to treat every articulate exit as a special class of loss.

If the thread suggests harm, discrimination, or a legal hold, stop the happy-path review and follow your existing investigation path. The bot does not open that case.

When the file is clean, you may hand structured, cited excerpts to exit interview thematic analysis. Themes operate on approved text. They do not get to invent a reason the person must have meant.

This conversation produces a transcript, a map of cited probes, and a set of blanks. It does not produce a regrettable score, a cause breakdown, or a message to the skip-level.

attrition root-cause attribution is a later step. It needs more than one exit and a method that can survive missing fields. A single adaptive chat with three blanks is a weak input. Do not let a model fill those blanks so the attribution job looks complete.

The same rule applies to manager inboxes. An automatic awareness forward is why people type thin survey answers: they assume the person they are leaving will read a warmed-over version of the chat. If HR wants the manager to hear something, HR writes that something, without orphan quotes.

In the first month, red-team the files. Pick a summary. Click every quote. Confirm the cite. Confirm the blanks. Confirm no manager share and no regrettable label. If any check fails, the workflow is not ready, even if the chat felt fluent.