Insurance · After Handoff
After Handoff: A Normal Tuesday
We don’t measure success by the launch date. We measure it by what happens six months later, when a production AI system is still running and we’re not in the room.
Here’s a normal Tuesday for the claims intake AI automation we built for a mid-market insurance carrier—not a pilot that died in a slide.
The constraint did not expire at go-live. Wrong routing still delays a payout. A licensed person still owns anything the model cannot defend. We had refused auto-adjudication, a chatbot on the desk, and a retainer that would have kept us in the loop. What transferred was the queue, the thresholds, the eval set, and the weekly review—not a vendor login.
8:14 AM
A claim comes in through the shared inbox, the way it always has. The system reads it, pulls the policy number, classifies the claim type, and extracts the fields an adjuster needs: date of loss, parties involved, estimated damage. No one touched it.
8:15 AM
The claim is routine — matches a pattern the model has seen thousands of times, confidence is high. It’s routed straight into the queue an adjuster works from, fields pre-filled.
11:40 AM
A different claim comes in. It’s messy: two policyholders, conflicting dates, a scanned PDF with handwriting in the margins. The model isn’t confident. Instead of guessing, it flags the claim and routes it to a human review queue with a note explaining what it couldn’t resolve. This is the part we care about most — the system is built to know what it doesn’t know.
2:00 PM
The client’s ops lead notices claims from one regional office are getting flagged more often than they should. She opens the rules panel we left her — no code, no call to us — and adjusts a routing threshold herself. It’s her system now.
End of day
62% fewer claims required manual review to get to the right desk. The 38% that did were flagged correctly, not missed.
We didn’t build a black box. We built something an adjuster trusts enough to act on, and an ops lead can steer without needing us on retainer.
That’s the actual deliverable: not the model, the judgment built into how it hands off.
The result, six months on: 62% of incoming claims still skip a human sort. The rest arrive with the evidence attached. The intake lead still owns the threshold. We still do not have a retainer. That is what transferred: a production claims intake system the carrier runs without us.
Another engagement
Tell us the job that already hurts.
We’ll tell you honestly if we can help and what it would take.