In July 2012, a European composite insurance group running several distribution labels closed out the pilot phase of a digital fitness assessment. Inside the method sits a written specification of what a motor claim was supposed to feel like in 2020, step by step, with the capabilities underneath it derived from that description rather than from an inventory of what the group already ran.
The target year has been and gone, and the technology the authors could not name in 2012 is on the table in 2026. That makes the document markable. Every sentence in the 2020 claim was a bet about what would turn out to be hard, and the years since settle most of those bets.
The instrument was built backwards
Most capability models are inventories dressed as strategy: someone lists what the estate contains, colours each item, and calls the reds a plan. This one was built the other way around. The group worked inside an industry consortium with outside partners to share digital practice across insurers, and the consortium's first deliverable was the assessment instrument itself.
Its construction rule was strict. Model the customer journeys you intend to serve in 2020, then derive the capabilities those journeys require. Four journeys were modelled: claims, sales including campaign management, value-chain integration, and product development. Motor claims was the worked case, chosen because it covers most of what a typical claims process does.
The design principles for the 2020 version of that journey were written out before a single capability was named: decisions taken on abundant and accurate data; the insurer managing the customer's mobility rather than merely covering the damage, which pulls it into the event far earlier than a first notification of loss; the insurer's own craft contributing to the experience; common tasks processed instantly and automatically; and claims prevented and fraud detected by learning from what had already happened.
The second of those is still the strategic argument in claims. Redefining the insurer's job from reimbursing a dent to keeping someone moving pulls the operation upstream of the loss, a proposition change rather than an automation project, which is why so much of what followed was organisational.
The derivation produced 22 capabilities, sorted into five groups spanning the insurance value chain and scored on a five-step maturity scale, from a business view and a technology view separately. Traceability is what makes it hold up: no capability made the list unless a modelled journey step needed it, so the pilot could say which gap blocked which step.
The claim as it was written for 2020
The claims journey was laid out in five stages: damage occurs, needs are determined, a solution is found, the event is learned from, and the claim is settled. The narrative underneath was written before the vocabulary that describes it existed.
An accident on an icy road badly damages a car. The vehicle's own system reports the crash, and the insurer calls the driver rather than waiting to be called. The loss details are verified without human keying and cross-checked against outside sources such as road cameras. The system works out from what it already knows that a trip is planned and a replacement car will be needed for it, then books the repair by taking the slot and the location from the driver's own calendar. The claim form arrives already filled in, with nothing left to do but approve it, and money is settled automatically between every party involved. Cars approaching the same stretch are warned about the ice through their navigation systems, and the insurer reads a public post afterwards to judge how it felt.
Three commercial offers fall out of that single claim, and none is a campaign: a larger courtesy car for the trip at an extra fee, extra cover for the trip itself, and a re-spray prompted by paint chips the inspection noticed. Each comes from something observed while handling the loss.
Two things traced to that journey still read as unusually mature for a 2012 document. The in-car notification was written as something to build jointly with vehicle manufacturers, an engineering partnership rather than a procurement line. And the use of personal and social data was specified as staying inside the limits the customer had set, which put privacy in as an enabling capability rather than a constraint bolted on at the end.
The line at the foot of the slide
Across the bottom of the journey slide, set larger than the narrative it qualifies, runs one caveat. The whole journey assumes a back end that is already standardised, automated and integrated.
The authors did not bury it. They stated it plainly, as a precondition of everything above it. But a precondition is not a scored capability. The parts of the model that spoke to that plumbing sat in the technology group, and neither was among the eight capabilities traced onto the claims journey itself. The narrative got the credit; the integration got a caption, and that caption is what the years since turned out to be about.
What the pilot actually concluded
None of the above describes an estate that existed. The pilot's job was to measure a real organisation against its own 2020 description, and it produced a set of capability gaps and the conclusion that a group-wide roadmap was needed to close them. The upper steps of the scale say where the group intended to get to, not where it was.
Every AI maturity assessment on the market sets the same trap. A target-state description scores badly on first contact by design, and the instrument only earns its fee if the gap is treated as work to fund rather than as a verdict on the company.
What made this one better than most is that the moves came attached. Each step up the scale was written together with the initiatives that would produce it, so the answer to what raises this by one sat on the page rather than up for debate.
Grading it in 2026
Mark the 2020 claim step by step and the pattern is sharp. The perception and reasoning steps have collapsed in cost. Assessing damage and working out what has to happen to resolve the claim is ordinary applied AI in 2026 rather than a stretch. Cross-checking the loss against outside sources is retrieval and verification, the same class of work as reading paint chips out of an image and turning them into an offer. Pre-filling a claim so the customer only approves it is the default shape of a well-built claims agent, and reading the experience from a public post rather than a survey nobody answers is solved too. Working out that someone needs a car for a planned trip, once a profiling exercise requiring a modelled segment, is a reasoning step over context already attached to the claim.
Nearly all of the intelligence in that journey is now available. What has not moved is the sentence at the foot of the slide.
Routing every party who has to touch the claim is not an intelligence problem. It is the question of whether the tow truck, the repairer, another label's policy administration and the group's settlement engine can be driven through one process against one version of the claim. Settling money between all of them automatically is the same problem wearing a finance hat. An agent that reasons beautifully and then writes into four systems that disagree about the claim produces fast, confident, inconsistent work, worse than a slow human who checks.
That is the honest grade. The 2012 authors made the plumbing a precondition because the reasoning looked like the hard part. The ranking has inverted. Agentic AI closes the steps the document treated as visionary and leaves untouched the integration it assumed.
The second inversion concerns privacy. In 2012 it was one capability among 22. Give an agent the ability to act across a claim and privacy stops being a scoring line and becomes the operating boundary: what the agent may read, what it may do without a human approving it, and whether every action can be reconstructed afterwards.
The third is a warning for anyone about to buy an AI maturity assessment. The instrument did not confine itself to technology: a full group of its capabilities covered how the organisation works and decides, and not one of those appears among the eight traced onto the claims journey. The journey slide had no use for them, and no model, vendor or agent raises them for you. Yet whether a claims handler may accept an agent's recommendation without a second signature is exactly that kind of question. An operation that scores well there gets value from an agent. One that does not gets a faster version of the process it already has, which was never the point.
The claim as a loop, and the harness around it
Read the five stages again as a control loop rather than a storyboard and the build gets much smaller. Damage occurs, needs are determined, a solution is found, the event is learned from, the claim is settled: that is an observe step, a decide step, an act step and a learn step, and each pass either closes the claim or hands back the reason it could not. Written that way the 2012 narrative stops being a description of a future and becomes something an engineering team can implement against, because every stage says what the loop must observe, what it may decide, and what it is allowed to do next.
Computer vision does the observing wherever the work arrives as a document or an image, and in motor claims most of it does. The inspection that noticed the paint chips, the claim form that arrives pre-filled for approval and the settlement paperwork passing between every party involved are all image and document work, and turning them into structured claim facts is ordinary applied AI now rather than a research programme. The journey asked for loss details verified without human keying and cross-checked against outside sources such as road cameras. That is a vision and retrieval pipeline, and it is buildable today.
What decides how fast the rest goes is the harness around the loop rather than the model inside it. The harness is the unglamorous scaffolding an agent runs in: the tools it is allowed to call, the schema of the claim it reads and writes, the checks that run before an action commits, the retries and escalations when a step fails, the evaluation set that says whether a change made the loop better or worse, and the trace that lets an auditor replay any decision after the fact. Build the harness first and the second journey costs a fraction of the first, because sales, value-chain integration and product development reuse the same tools, schema, checks and traces. Build a clever model first and every journey is a rebuild.
Autonomy then becomes a dial set per step rather than a decision taken once. The same loop can run with a handler approving every action, with approval required only above a threshold, or with the routine claim settled end to end and everything unusual escalated. Where the dial sits is an operating choice and belongs on the page next to the step, in the same way the assessment put the initiative next to the score. The organisational capabilities nobody traced onto the claims journey are what let you move the dial at all.
And the group-scale version only exists when the loops are wired to each other. A single claims agent is a demonstration. A claims loop that hands the repair booking to a scheduling loop, the payment to a settlement loop, the fraud signal to an underwriting loop and the observed paint chips to an offer loop, each with its own harness and all writing against one version of the claim, is agents carrying a real share of how the group runs. That is the version worth funding, and the thing between here and there is still the sentence at the foot of the slide. Loops cannot hand work to each other across a back end that is not standardised, automated and integrated.
What to take from it
The method is worth copying: write the journey you intend to run, decompose it into the capabilities it requires, define the steps behaviourally, and name in advance the initiative that moves each score up one. Do that for an agentic target state and the plan stops being a wish list.
The finding is worth acting on. In claims, the intelligence is now the affordable half, and most of what remains is loop and harness engineering, which is scheduled software work rather than a research bet and moves a great deal faster than the roadmap language around it suggests. Integration, a shared version of the truth across labels and third parties, and the organisational bar at the top of any maturity scale are the half that decides whether an agent leaves the pilot. That is what the 2012 authors compressed into one assumed line, and it has aged better than anything else in the document.
- 22
- Capabilities derived from the 2020 journeys
- 5
- Steps on the maturity scale
- 4
- Customer journeys modelled for 2020
- 5
- Stages in the specified 2020 claim
In 2012 the reasoning looked like the hard part and the plumbing went in as a precondition. In 2026 the ranking has inverted, and the precondition is the programme.
“The reasoning steps in that 2020 claim are the cheap part now. The precondition written across the foot of the slide, a back end already standardised, automated and integrated, is still the whole job.”
Get in touch
Put RealAI’s applied-AI team on your hardest data problem.
We help enterprises move from pilots to production: sovereign models, governed data, and agents you can audit. Start with a value-first assessment.