Use case

Keeplore for customer support teams running an AI agent

Last updated 24 October 2026

If your AI support agent gets refund exceptions wrong, the answer is usually in the head of one of your senior people, not in the help centre. Keeplore interviews that person, turns what they say into rules a manager approves, and gives your agent each rule with the name and role of the person it came from. The reviewer sees the exact quote behind every rule.

What is the gap between the help centre and the case in front of the agent?

Your agent answers from the help centre. Your best people handle the cases that don't fit it, and they do it from memory.

Most tickets match an article. Some don't. A customer cancels mid-term. An enterprise account has its own terms. A credit is over the amount a team lead can approve. The article states the rule. The exception is in someone's head. The agent quotes the article or escalates.

What do support experts know that the help centre doesn't?

Common categories:

• Refund exceptions
• Enterprise contract terms that override the standard policy
• Who approves credits above a limit
• Customers who get different handling, and why
• Which step in an internal tool goes wrong and how to work around it

Exceptions like these:

• "Past the 30-day window on annual plans, Acme offers a prorated credit for the months left."
• "Credits over €500 need sign-off from the finance lead."
• "The billing tool shows the wrong renewal date for accounts migrated from the old system. Check the contract."

Each of these is the kind of thing a Keeplore interview asks about.

Who is this for?

Teams that run an AI support agent, such as Intercom Fin, Zendesk AI agents, Sierra, Decagon or Ada, or one they built themselves.

Keeplore is not affiliated with these vendors and has no integration with them today. Your agent can call Keeplore's API, or you can export the rules and add them to the guidance or knowledge feature your agent tool offers.

It works for people who are leaving, retiring or changing role. It also works for people who are staying: the colleague everyone asks when a case doesn't fit the policy.

How does it fit into a support team?

1. Pick two people. One who knows the exceptions. One who is leaving, changing role, or is just the person everyone asks.
2. Send each a private link. They chat with Keeplore's AI interviewer in plain text for 30 to 45 minutes. They can pause and come back. They give consent before it starts.
3. Review the rules. Each rule has a condition (when), an action (do), a reason (why) and the exact quote. A manager approves, edits or rejects each one.
4. Send the approved rules to your agent. Your agent calls POST /api/v1/query and gets an answer plus the rules it used. If no approved rule applies, the answer is empty and your agent can treat that as "hand over to a person". Every plan has a monthly AI budget: when it is reached, the query API returns HTTP 402 ("This workspace has reached its monthly AI limit"). Each API key is limited to 60 queries a minute. Or export every approved rule with GET /api/v1/export as JSON or Markdown and load it into your agent's guidance or knowledge feature. Check your agent tool's documentation for the format it accepts.

For how each step works, see How Keeplore works. For how this sits alongside retrieval over your help centre, see Keeplore vs RAG.

Why does a sourced rule matter in support?

Because someone has to be able to check it.

• Every rule keeps the person and the quote. The reviewer sees who said it and in what words. Your agent gets the person's name and role with each rule.
• Only approved rules reach the agent. A manager reads each one first.
• If no approved rule applies, the answer is empty. When a rule does apply, a model writes a short answer from the approved rules, and the rules it used come back with it so your agent, or you, can check. The answer text is model-written and is not checked against the rules.

Keeplore checks that each quote appears in what the employee typed. Very short quotes prove little, and the check does not show the rule is right. That is why a manager approves each rule. How the quote check works.

How do we start?

Start with the free trial: 2 interviews, work email, no card. Pick the person your team asks when a case doesn't fit the policy. Read the rules that come back.

If you want to go further, a pilot is 4 weeks, €1,500 flat, for up to 5 interviews. If you continue, the €1,500 is credited in full. After that it is a yearly plan sized to your company: €6,000 for up to 300 people with 15 interviews, €12,000 for 300 to 1,000 people with 40, or €24,000 above 1,000 with 100. Extra interviews are €300 each. See pricing.

How do you judge the pilot?

Here is a method. It makes no promise about results.

1. List your 10 most-escalated questions.
2. Add 5 questions whose right answer is "we don't know".
3. Before the interviews, run all 15 through the query API. Count how many get an answer and how many come back empty.
4. Run the interviews and approve the rules. Run the same 15 again.
5. Count again. Read each answer against what your best person would say. The 5 "we don't know" questions should still come back empty.

For your works council

Facts you can check on /security:

• The employee agrees before the interview starts.
• Managers at the employee's company see the answers; the employee is told this before the interview starts.
• The query API returns the employee's name and role with each rule.
• There are no email, Slack or calendar integrations.

What should IT and security know?

• The app and database run in Frankfurt. Interview text is processed by Anthropic in the United States, sign-in data by Clerk, and invitation emails by Resend, both also in the United States.
• Keeplore offers a data processing agreement. Employees give consent before every interview. You can delete an interview and its rules in one click.
• Keeplore has no SOC 2 or ISO certification today.

Related

• How Keeplore works
• Keeplore vs RAG
• Pricing
• Security
• Knowledge transfer questions