AI, retrofitted is a series about the screens and services where AI has been bolted onto an existing product. Each piece takes one real, evidenced problem and one fictional brand, and the hero artefact is an AI-aware service blueprint. Northmere Insurance and Maria are invented. The problem isn’t.

The Problem

Maria is rear-ended at a red light. Nobody’s hurt; the bumper is cracked. She opens her insurer’s app, takes six guided photos, and four minutes later sees an estimate of €1,240 with a bright button: accept by Friday and get paid within 24 hours.

Her garage takes the bumper off and finds a bent bracket and a damaged rear radar, which has to be replaced and recalibrated. The real cost is €3,100. Had she accepted, the claim would have closed, and the €1,860 behind the bumper would have been hers to pay.

Two photos: the rear of the car with a cracked bumper, labelled €1,240, and the bumper removed showing a bent radar bracket, labelled €3,100
What the app saw in four minutes, and what the garage found behind the bumper.

None of this is exotic:

The model isn’t wrong about what it sees. The problem is what it doesn’t say: that a photo can’t see behind a bumper, and that accepting the money closes the claim anyway.

How might we make an automated estimate honest about what it can’t see, and put a person on the case when it matters?

How the Decision Is Made Today

Four steps, four minutes, no person involved. The model looks at six photos, spots the cracked bumper with high confidence, prices the parts, and, because it’s confident and the total is under the limit, offers instant cash.

Decision anatomy today: six photos, damage detection, a price table totalling €1,240, and a rule that triggers an instant cash offer
The decision today. The rule never asks whether the car has sensors behind the bumper.

The question it never asks is the one that matters: does this car have sensors behind the bumper? The insurer already knows. The plate gives the model and trim, and the trim says rear radar. The estimate just doesn’t use it.

What Maria Lives Through Today

The blueprint has seven lanes, from Maria to the ombudsman, with the AI as a lane of its own. Today the model’s blind spot turns into six pain points:

  1. A single number hides the uncertainty.
  2. A countdown pushes her to accept.
  3. Accepting signs a full-and-final release.
  4. The only help is a chatbot that repeats the estimate.
  5. The garage’s supplement disappears into a queue.
  6. A desk reviewer confirms the model without new evidence, which isn’t meaningful review.
Service blueprint of today's claim, seven lanes and eight stages, with six numbered pain points
Today's service. Six pain points, numbered so each one can be traced to a fix.

I didn’t draw “before” screens for a company that doesn’t exist. The pattern is common enough that a grey sketch of it does the job: one big number, one bright button, a countdown.

From Pain to Fix

Each fix keeps the number of the pain it answers:

Table mapping six pains to six fixes and the principle behind each
Six pains, six changes to the service.

The most important one is fix 3, a new decision rule. If a fitted sensor sits in the impact zone, there’s no instant payout. Instead, Maria gets an advance for what the photos do show, the claim stays open, and an inspection is recommended. The money arrives just as fast; she just doesn’t sign away what nobody has seen yet.

The Redesigned Service

Service blueprint of the redesigned claim, nine stages, with numbered fixes, decision points, a hand-off to a named assessor and three designed failure paths
The redesigned service: fixes numbered to match the pains, a hand-off with the whole case file, and three failure paths designed on purpose.

Three things change:

Six Moments

Six phone screens: estimate with boundaries, honest choice, what would change the number, a named person, case timeline, revised decision
Six screens, each answering one pain.
  1. Estimate with boundaries.
    • Maria’s own photo carries an x-ray layer: the crack is framed as seen, and the area behind the bumper is hatched, with the two radar units drawn as dashed outlines.
    • A range bar splits “seen €1,240” from “if damaged +€2,160”.
    • Tapping the hatched zone explains why it’s unknown.
  2. An honest choice. Three options with their consequences in plain words, and no countdown. For this car, keeping the claim open is the recommended one.
  3. What would change the number. The garage’s photo is read on arrival, the bent bracket is outlined, and a checklist says what evidence counts.
  4. A named person. Sam Reyes, a licensed assessor, a date, a phone number, and “Sam already has your file”. A “talk to a person” control sits in the corner of every screen.
  5. One case timeline. Every step is dated, mirrored in email, with a service clock that shows whether the next step is on time.
  6. The revised decision. €3,100, line by line, decided by Sam after seeing the car. The hatched zone from the first estimate is now solid: confirmed.

The signature runs through the whole claim. Hatching means unknown, on the photo, on the range bar, even on a report’s missing lines. It turns solid once something is confirmed.

A walkthrough of the clickable prototype, 53 seconds, no sound.

How the Decision Is Made Now

Same model, two new inputs, one new rule. The model isn’t replaced. It’s given what the insurer already knows about the car, and it’s no longer allowed to close a claim it can’t fully see.

Decision anatomy redesigned: photos plus car profile, a map of what can't be seen, a priced range, and a rule that holds back the cash
The decision now. Changes are marked in blue.

When Things Go Wrong

The happy path is the easy part. I designed three failure paths on purpose:

Three phone screens: report not itemised, visit rebooked, decision partly upheld
Three failure paths, designed on purpose.

Beyond the App

When a person takes over, the conversation leaves the app:

The assessor introduction email and the appointment text thread
The assessor email mirrors the app's timeline; the texts confirm without links.
The revised decision letter with annotations
The revised decision letter. Open the PDF.

In Context

Maria at her kitchen table with the revised decision on her phone and the printed letter beside it
The decision on her phone and in the letter: same numbers, same bar, same name.

The garage gets its own view of the same claim: the approved repair, line by line, invoiced to Northmere directly. Maria doesn’t carry paperwork between them.

Principles

Four principles carry the redesign, each from my Agent UX Framework:

PrincipleLayerIn this serviceThe test
Show your workTrustEvery number points at what it’s based onIf a figure can’t be traced to a photo, a document or a person, it isn’t shown
Communicate uncertaintyTrustSay what the system couldn’t see, and give a rangeA customer can say what the estimate doesn’t cover
Blast radius containmentGuardrailNo instant cash when hidden damage is likelyNo automated outcome is irreversible when the model knows it was partly blind
Human escalationGuardrailA named, timed person, with the context carried overThe customer never explains twice, and a person signs the outcome

Hypotheses, Not Results

This is a concept, with no user research or live data behind it. So every design bet is written as a hypothesis, with a measure and a guardrail metric that must not get worse:

We believe……will lead toMust not get worse
H1Showing a range and the blind spotsFewer disputes after teardownOffer acceptance on claims with no sensors in the impact zone
H2No instant cash with a sensor in the impact zoneFewer reopened claimsTime to final settlement
H3An advance that keeps the claim openNo one waits for money while the claim stays openAdvances above the final amount
H4A named assessor with the case fileFewer repeat contacts and escalationsAssessor hours per case
H5Telling customers before a promise slipsTrust holds when the service is lateVisits rebooked more than once

How I’d test it:

  1. Comprehension first. Do people read “€1,240 to €3,400” as a promise of €3,400?
  2. A scoped pilot. Only rear impacts on cars with rear sensors.
  3. Targets after a baseline. No promised numbers before then.

What it costs the insurer: more assessor visits and fewer instant settlements on the claims the rule catches. In return: fewer reopened claims, fewer complaints reaching an ombudsman, and a decision trail it can defend.

What I’m Not Sure About

Sources

Checked 11 October 2026. Legal points are context, not legal advice.