maxtrifonov.dev
DE
Email me

Case study

AI conversation analysis

How a company reviews its customer calls continuously, with every result traceable.

The starting point

Advisors follow a fixed conversation guide in every call. Whether they actually stuck to it could only be checked by spot checks: someone had to listen back to individual calls.

The brief

Check every call against the guide automatically, give a reason for every assessment, and show team leads how their team is doing without them having to listen back themselves.

The approach

An AI checks every point of the guide and cites the passage in the conversation it relies on. A second, independent model double-checks every result. Where there is doubt, a person decides, and the review learns from each of those decisions.

  1. Transcriptthe text of the conversation, the basis for the review
  2. AI reviewevery point of the conversation guide, with the supporting passage
  3. Independent reviewera second model double-checks every result
  4. Clearresult with a reason
    Borderlinea person decides. The review learns from it, so more cases become clear.
  5. Overview for team leadsresults by team and advisor

An example

A made-up customer service call: a customer is moving house. Tap a point to highlight the passage.

Conversation guide
  1. AgentGood afternoon, you're through to customer service, my name is Jana Keller.
  2. CustomerHi, I'm moving in May and I'd like to take my connection with me.
  3. AgentOf course. So you're moving to Berlin on May 1st and want the connection to move with you, right?
  4. CustomerExactly.
  5. AgentFiber is available at the new address, and we can set it up on May 2nd.
  6. CustomerDoes that cost extra?
  7. AgentThere's a moving fee, you'll see it on your bill.
  8. CustomerOkay. So what do I need to do now?
  9. AgentI'll email you the confirmation today, with a link to book the appointment.
  10. CustomerGreat, thanks.

Line 1. Names the service and gives their own name.

Line 3. Plays back the date, destination and what the customer wants.

Borderline, line 7. The fee is mentioned, but not the amount. Whether that is enough is for a person to decide, and their decision feeds back into the review.

Line 9. Confirmation by email with a booking link promised.

The outcome

  • 20,000+ analyses, 2,000+ transcripts a month, ongoing
  • Every assessment comes with a reason and the supporting passage
  • Team leads see results by team and advisor
  • The share of borderline cases drops with every human decision

Stack

Gemini, Claude, Node.js, Supabase, Google Cloud Run

What I took away

An AI assessment is only useful if you can check it. A reason and the supporting passage matter more than a bare number.

Borderline cases require human judgment, and every one of those decisions improves the system.

Want to review conversations in your company?

Email me All projects