Compare
Can't we just do this with ChatGPT?
Last updated
Partly, yes. For one conversation, with a good interviewer and a manager who reads every line, recording the call and asking a chat assistant to summarise it gets you most of the way.
It gets harder with many interviews, a reviewer who wasn't in the call, an agent that has to say "I don't know", or a works council asking how the person's words were kept apart from the model's. Keeplore is knowledge capture for AI agents. This page runs one answer through both routes and says when the do-it-yourself route is enough.
What does the do-it-yourself route look like?
You record the exit interview on a video call, paste the transcript into any general chat assistant (ChatGPT, Claude, Gemini), and ask for the rules this person follows. You get a tidy summary back.
It is a reasonable approach, and the summary is often useful. The differences are in the process around the model, not in the model.
What do both routes do with the same answer?
Here is the excerpt from the Acme sample on How Keeplore works. Jonas Berg works in support. Maya Chen, the billing lead, reviews his rules.
Interviewer: Annual plans have a 30-day refund window. What happens after it?
Jonas: Officially it's no refunds on annual, but if they're past the window we do a prorated credit for the months left. Anything over five hundred I check with Dana in finance first.
Interviewer: Can you tell me about the last time that happened?
Jonas: Last month. A customer asked in week six and we credited the unused months.Jonas said "officially", so the interviewer asked for a real case instead of moving on. In a recorded call, that follow-up depends on who is asking.
- Annual plans: no refunds after the 30-day window.
- Exception: offer a prorated credit for the unused months.
- Credits over €500: check with Dana in finance first.That is a correct summary, and many teams would be happy with it. You could also prompt for verbatim quotes and approve the rules in a spreadsheet. The difference is that Keeplore checks each quote against the transcript in code and drops the rule if the quote isn't there, rather than trusting the model to follow the prompt.
| Field | Value |
|---|---|
| When | A customer on an annual plan asks for a refund after the 30-day window |
| Do | Offer a prorated credit for the months left. If the credit is over €500, hand over to a person for finance approval |
| Why | Not stated. Keeplore leaves the reason empty if the person did not give one |
| Quote | "Officially it's no refunds on annual, but if they're past the window we do a prorated credit for the months left. Anything over five hundred I check with Dana in finance first." |
| Source | Jonas Berg, Support Specialist |
The 30-day window comes from the interviewer's question. The quote is Jonas's own words. Before any agent can use the rule, three things happen. Keeplore checks that the quote appears, as whole words, inside one of Jonas's answers. Jonas sees the rule under his name after the interview and can flag it. Maya, who wasn't in the conversation, sees the quote and the question it answered, then approves, edits or rejects the rule.
How do the two routes compare?
The middle column describes a typical do-it-yourself setup, not any one tool.
| Recording + chat assistant | Keeplore | |
|---|---|---|
| The interview | A person runs a call you schedule. Follow-ups depend on them | A typed chat, one question at a time, about 30 to 45 minutes. It asks follow-ups on "usually", "it depends" and "except" across six areas. The person can pause and come back |
| Quotes | As good as your prompt and your own check | Checked in code. Each quote must appear in one of the person's own answers, not in the interviewer's words and not stitched across two answers, or the rule is dropped. A quote of only "yes" or "sure" is not kept |
| Guesses and plans | Mixed in with the rest unless you separate them | The model labels each rule as practice, plan or guess. Guesses are dropped. Plans are capped at 50% confidence (a score the model gives, not a measurement), so they are marked "Check this" and approved one by one |
| Review | Your own process | A manager approves, edits or rejects each rule. A new rule that contradicts an approved one is flagged. Keeplore doesn't decide which is right |
| The person's say | Up to you | After they press Finish, they see every rule written under their name and can flag any that's wrong before a manager approves it |
| When nothing applies | Up to the agent the summary is pasted into | POST /api/v1/query returns an empty answer and no rules, so your agent can hand over to a person |
| Data | Depends on your plan. Business plans can include a data processing agreement, and some offer EU data residency | App and database in Frankfurt. Interview text is processed by Anthropic in the US. A DPA is offered. The person agrees before the interview starts, and you can delete an interview and its rules in one click |
Why a typed chat and not a call?
A recorded call is easier for the person on the day. A typed chat means nobody has to find an hour that suits both of you, run the interview well, or transcribe it, and the person can stop and come back between other work. Whether your people will finish one is worth testing on a single person first.
What changes when there is more than one interview?
A reviewer who wasn't in the call. Reading a summary, Maya can't easily tell what Jonas said from what the model added. With his exact words and the question they answered, she can.
An agent that has to say "I don't know". When nothing in a pasted summary matches, what happens is up to that agent. When Keeplore's query finds no approved rule, the answer is empty, and your agent can hand over to a person. When a rule applies, the rules used come back with the answer and who they came from. The answer text itself is model-written and is not checked against the rules.
A works council. They will ask how the person's words were kept apart from the model's. Every rule carries the person's exact words, the person can flag their own rules, and only rules a manager has approved reach an agent. Keeplore does not yet record who approved or edited each rule.
What about privacy?
A recording of an employee is employee data, whichever route you take. If your company's chat assistant is on a business plan with a data processing agreement, and perhaps EU data residency, the do-it-yourself route may keep the data closer to home than Keeplore does, because Keeplore's interview text is processed by Anthropic in the US. On a personal or free account there is usually no agreement at all.
Either way, decide where the recording goes and tell the person before you start. Keeplore's setup is on the security page.
When is the do-it-yourself route enough?
- One person, one conversation. You don't need a system for a single handover.
- You have a skilled interviewer. If someone on your team reliably gets to the exception, you have the hardest part.
- You will read every line yourself. You are the review.
- No agent will use the result. If a person reads the summary, nothing needs to answer "I don't know".
- Your data terms already cover it. A business plan with a data processing agreement, and the person told where it goes.
Related
01Can we use a chat assistant and Keeplore together?
Yes. A chat assistant is useful for drafting interview questions or tidying your own notes. Keeplore's rules come only from the person's answers in its own interview.
02How do Keeplore's rules get into Fin or another agent?
Your agent can call POST /api/v1/query, or you can export every approved rule as JSON or Markdown with GET /api/v1/export and load it into your agent's guidance or knowledge. There is no pre-built integration and the export is manual. The empty answer and approved-only rules apply when your agent calls the query API.
03Does Keeplore check that a rule is correct?
No. It checks that the quote is in the person's own words. Whether the rule is right is the manager's call, which is why each one is approved before an agent can use it.
Try it on one exception.
Pick one exception your agent gets wrong. Interview the person who handles it with one of your 2 free interviews, and compare the rule with your own summary.