AI strategy · Governance
GenAI, control and humans in motor claims
Twelve processes settle a motor claim, and each is carried by some mix of a generative model, a deterministic control, and a person. Start with the exhibit.
Where each process runs today, and its headroom
Each corner is one of the three carriers, each ball one of the twelve processes. Select one to isolate its position and its headroom. How to read this →
The GenAI corner is empty. No process today has generative AI carrying even a third of the work, and the cluster runs along the edge between control and humans. Motor claims is still a rules-and-judgment business with generative AI at the periphery. Process 10 ends exactly where process 1 sits today.
Written from administering motor claims for insurers. Each of the twelve processes was assessed on one question: of the work this process takes, how much is carried by generative AI, how much by deterministic control, and how much by a person. The three shares are allocated across a fixed total, so a gain for one carrier is always a loss for another. Published research is cited where it reaches these processes. Method and limits sets out what the allocations are and are not.
In this paper
What this paper finds
A motor claim is settled through twelve separate decisions, and each one is carried by some mix of three things: a generative AI model, a deterministic control, or a person. Axxion took each of the twelve in turn and asked how much of the work each of the three carries in the industry today, and how much each could carry with the technology that already exists, properly governed.
Three findings came out of the exercise. Generative AI runs at under half of its potential, and the constraint is the control layer and not the models. The control layer itself does not shrink as generative AI grows: it holds at the same share of the work in both states, which is the clearest single test of whether an operation is deploying AI correctly. The third finding carries the consequences. Human oversight falls sharply while concentrating into the four processes that carry the most money, the most dispute and the least measurement.
- Generative AI is good at one specific thing, and a claim file is full of it. Turning unstructured input into structured output is the core competence. A claim arrives as photographs, police reports, handwritten estimates, invoices and voice notes, which is the least structured input in financial services.
- The control layer does not shrink. Generative AI's share more than doubles while the deterministic layer holds steady, and control stays the largest of the three in both states. An operation whose control layer thins as its models grow has removed the thing that made the models safe to run.
- Human oversight drops by about 60%. Liability, subrogation, settlement approval and repair method account for 40% of all human judgment in a claim today, and half of it once generative AI is deployed. People move out of the routine work and are left holding the contested work.
- The person is the least inspectable part of the system. A model can be tested against held-out data, versioned and monitored for drift. A control can be read by anyone and audited against the files it governed. Neither applies to a person, and adjuster-to-adjuster variance on comparable motor files has never been published by any regulator, actuarial body or journal that an August 2026 search reached.
- The end state is not an extreme. Generative AI fails by opacity and control answers it. Control fails by rigidity and generative AI answers it. People fail by inconsistency and the other two answer it. Each carrier covers another's blind spot, which is why all three survive.
1. The three carriers
Every step in a motor claim is carried by a mix of three things. Defining them is important, because most writing on this subject uses "AI" to mean all three at once and then argues about the result.
Generative AI
Probabilistic models: large language models, vision models, and the prediction models alongside them. What they are good at is turning unstructured input into structured output. A photograph of a police report becomes fields. A set of damage images becomes a parts and labor list. What they return is a confidence, never a certainty.
Fails by opacity. A confident wrong answer is indistinguishable from a confident right one at the moment it is produced.
Control
Deterministic systems. Validation rules, written thresholds, benchmark tables, authority matrices, policy wording, and the governance layer around all of it: explainability, reason codes, audit trails, model inventories, version history. A control returns the same answer to the same input every time, and anyone can read why.
Fails by rigidity. A total loss threshold set against one salvage market keeps answering after that market has moved.
Humans
Co-human experts, not a fallback and not a rubber stamp. The human takes the contested call, the unusual claim, and the exception the other two surfaced and could not settle between them.
Fails by inconsistency. Two competent adjusters reach different answers on the same file, and this failure mode has no instrument pointed at it in any claims operation the published record describes.
The design rule that follows. Generative AI proposes. Control decides whether the proposal is admissible and records why. Humans settle what the first two cannot. A process that skips the middle carrier has a model writing straight into a system of record. A process that skips the third has no path for the claim the model has never seen.
2. Reading the triangle
Each process carries a triangle. The corners are the three carriers. The ball shows how much of that process each corner is doing. The white ball is where the motor insurance industry runs today; the navy ball is where the process can run with generative AI deployed and properly controlled. The vector between them is available headroom, and it carries no date.
That is what Exhibit 1 draws. Open it ↑ and take one process at a time: the balls carrying more than one number sit at identical allocations, and the arrow is the distance between the two states rather than a schedule for crossing it.
Capgemini reached a similar picture by survey in 2026, surveying 344 executives and 809 employees including 200 claim adjusters: coverage determination, liability assessment and settlement "still rely heavily on individual adjuster judgment, applied inconsistently across teams, regions, and complexity levels," while AI has been deployed "only at the edges of the claims process." Two different methods, pointing the same way.
3. The twelve processes
Each entry gives where the industry sits, how far generative AI moves it, and what each of the three carriers does in the end state. Where Axxion's own operation illustrates a point, it is marked.
01 · FNOL document extraction
A police report, a driving license, an Emirates ID and a workshop invoice arrive as photographs taken at the roadside, at night, at an angle.
Today
Most operations put a person on every document. That is a trained adjuster performing optical character recognition by hand, and it is the single largest waste of skilled time in the claim.
GenAI in the end state
Reads the images, identifies each document type, and returns structured fields: plate, chassis, date and time of loss, location, parties, police report number, and the officer's account of what happened. This is the competence at its purest.
Control in the end state
Nothing the model produces writes to the claims management system directly. Each field passes a deterministic gate: format validation on chassis and plate, policy match on vehicle and date of loss, a cover check, and a confidence threshold that routes rather than guesses. An audit entry records what was read, from which document, at what confidence.
Humans in the end state
Where model and controls together surface something the system cannot settle: a document that fails validation, a mismatch between police report and policy, or an early leakage or fraud indicator. In a properly built flow that is no more than 10% of notifications received.
02 · Notification valid and complete
Whether what has arrived carries enough to open a claim under this policy.
Today
A checklist confirms fields are populated and a person judges the awkward ones, which means completeness is defined as presence rather than sufficiency.
GenAI in the end state
Reads for sufficiency. A police report saying "minor contact, no injuries" and one saying "vehicle overturned, occupants transported" populate the same field and mean different things for what the file needs next. The model reads the difference and requests the specific missing item.
Control in the end state
Everything with a written answer stays written, and control gives up nothing here. Whether the policy was in force on the date of loss is not a probabilistic question. Cover scope, excess, endorsements, notification deadlines and the required document set stay deterministic, and the model works inside them rather than over them.
Humans in the end state
Late notification, a cover question with a real argument on both sides, and claims where the policy language does not clearly reach the circumstances.
03 · Claim assignment
Which handler, which track, which authority level, which service level.
Today
A routing table, which is why this process sits closest to the control corner of anything in the set. Routing tables sort on what is known at intake and cannot sort on what the claim is going to become.
GenAI in the end state
Predicts complexity and ultimate cost from the notification content before anyone has read the file. A rear-end impact on a three-year-old sedan with no injuries and a cooperative counterparty is a different claim from a multi-vehicle incident with a disputed police report, and that difference is legible in the FNOL text long before it is legible in a routing table.
Control in the end state
Holds the authority levels. Who may approve what is a governance question with a written answer, and a prediction never sets it. Control also holds the segregation rules: which claims must reach a named handler, which must never enter an automated track, which require a licensed assessor.
Humans in the end state
Rarely at assignment. The meaningful oversight is retrospective, a supervisor reading the rate at which claims change track after assignment, broken down by original track, which is the only way mis-routing becomes visible at all.
04 · Fraud referral
Whether a file goes to investigation.
Today
Indicator lists and a handler's instinct, which catches the obvious and misses the organized.
GenAI in the end state
Detects the signal no handler can see, because the signal only exists across claims: one phone number behind four unrelated notifications, a workshop whose supplement rate sits three standard deviations above its peers, a damage pattern inconsistent with the described impact, a loss reported eleven days after inception. One file at a time none of that is visible. Across a portfolio it is obvious.
Control in the end state
Records a reason on referral and, harder and rarer, on non-referral. Tracks referral precision instead of referral volume, which is what stops a model drifting into flagging everything. Holds the escalation path and the evidentiary standard the investigation must meet.
Humans in the end state
On every referral. This is an accusatory act against a named customer, and it keeps a human owner even where the model is right. That is a governance choice rather than a capability limit, and it is why human oversight is never designed out of this process.
05 · Damage assessment
What is damaged, how badly, and what that implies for parts and labor.
Today
The model proposes and a person concludes, almost everywhere.
GenAI in the end state
Identifies damaged panels and components from photographs, classifies severity, separates repair from replace, and infers the hidden damage that follows from a visible impact. This is the highest generative AI position in the whole set, because vehicle damage is the most mature computer vision application in insurance and the images are standardized by the process that captures them.
Control in the end state
Less than anywhere else in the claim, and saying so matters. There is a confidence gate, a rule that certain damage types always require physical inspection, and the audit trail. No table tells anyone whether a quarter panel is repairable.
Humans in the end state
Structural damage, suspected water or fire involvement, anything below the confidence gate, and a blind re-inspection sample where a second assessor sees the images without the model output. That sample is the only control that detects automation bias, and it is the one most operations skip.
06 · Repair method, OEM against aftermarket
Which parts go on the vehicle.
Today
Policy wording and a parts schedule, negotiated against what the workshop can actually source.
GenAI in the end state
Builds the availability and price picture across suppliers, matches part numbers across catalogs that disagree with each other, and flags when the schedule has gone stale against the vehicles actually arriving. It does not choose.
Control in the end state
Everything. Policy wording, parts schedule, warranty position, and the age and value bands that determine what is permissible are all written, and they are what the insurer is contractually bound by. This is a control-corner process and it should stay one.
Humans in the end state
The negotiation with the workshop, and the exceptions where schedule and vehicle disagree. That includes a live and growing category in the UAE: electric and Chinese-brand vehicles entering the fleet faster than parts schedules are revised, where the schedule has no entry and a person has to decide.
07 · Repair estimate challenge
Whether the workshop's estimate is right, and which lines to contest.
Today
Benchmark tables and a handler's experience, which means the challenge is only ever as good as the individual running it. This is one of the largest sources of unmeasured variance in a claims operation.
GenAI in the end state
Reads the estimate line by line, including the ones that arrive as a photograph of a printout, normalizes line descriptions against a standard taxonomy, and compares each against the benchmark distribution for that vehicle and that repair. It returns the contested lines with a reason attached: labor hours above the ninetieth percentile, a part billed as new that the damage assessment never called for, a paint time inconsistent with the panel count.
Control in the end state
Holds the benchmark itself, which is a deterministic artifact built from the operation's own history. Holds the tolerance bands that decide what is worth contesting. Records the sustained rate on challenges by reason and by workshop, which is the only way anyone knows whether a challenge is right or merely aggressive.
Humans in the end state
The conversation with the workshop, which is a negotiation between two people who will speak again next week. The model produces the position; a person holds it.
08 · Total loss determination
Repair or write off.
Today
A threshold rule comparing repair cost against vehicle value, with a person handling the files that land near the line.
GenAI in the end state
Improves both inputs and takes none of the decision. It produces a repair cost earlier and more accurately from the damage assessment, and a pre-accident value from the comparables.
Control in the end state
Makes the call, and keeps everything it has. A threshold comparing cost against value is written, auditable and applied identically to every file, which is exactly what a decision of this kind should be. Control also owns the review cadence on the threshold, tested against realized salvage recovery rather than against policy convention.
Humans in the end state
Files landing in a narrow band either side of the threshold, and vehicles where salvage value is genuinely uncertain.
09 · Total loss valuation
What the vehicle was worth the moment before the loss.
Today
A comparables model proposes, adjustment rules modify, and a person can override. This is the most disputed number in motor claims.
GenAI in the end state
Widens the comparable set well beyond what a person can assemble, reads condition and specification out of listing text and photographs, and normalizes across listings that describe the same trim four different ways. In a market like the UAE, where specification varies by import route, that normalization is most of the work.
Control in the end state
Gives up nothing, because this is the figure the owner will contest. The adjustment logic is written. The comparable set is disclosed. Every condition deduction is documented against evidence. The audit trail is the entire defense, and a valuation a claimant cannot interrogate is a complaint already in motion.
Humans in the end state
Disputes, modified or unusual vehicles, and any file where the comparable set is thin enough that the distribution is doing more work than the data.
10 · Liability apportionment
Who was at fault, and in what proportion.
Today
A police report, a legal framework, and a judgment between two accounts of the same event. The most human process in the claim.
GenAI in the end state
Reconstructs the incident from the report, the images and the parties' accounts. Compresses forty pages of correspondence to the two points still in dispute. Retrieves how comparable circumstances were apportioned before. Everything the person reads becomes machine-produced, and the decision itself barely moves.
Control in the end state
Holds whatever apportionment framework applies, and the escalation path where the parties disagree.
Humans in the end state
Always, and this process has the least headroom of anything in the set.
11 · Subrogation viability
Whether recovery is worth pursuing, and against whom.
Today
Rules of thumb about value and counterparty, plus a handler's read of the file.
GenAI in the end state
Reads liability and the counterparty position out of the file, then predicts recoverability and expected net recovery against the cost of pursuing it. This is a clean supervised learning problem, because the label is unambiguous: either the money came back or it did not, and every operation is sitting on years of it.
Control in the end state
Holds limitation periods, pursue thresholds, and reason codes on files not pursued.
Humans in the end state
Recoveries above a written value, and counterparties where the commercial relationship is a factor no model can see.
12 · Settlement approval
Whether this amount may be paid, and by whom.
Today
An authority matrix and a person's signature, split roughly evenly.
GenAI in the end state
Almost nothing. Generative AI approves no payments, and the only thing it adds is assembling the approval pack so the approver sees the complete position in one place.
Control in the end state
Nearly everything, and more of it than today. When the figures underneath become reliable, the authority matrix can be widened and written down with confidence, so approvals that used to need a person's judgment become approvals that satisfy a documented rule. The matrix is versioned, and the rate at which approvals are sought outside it is monitored.
Humans in the end state
Above the authority threshold, on any file with a coverage question, and on anything a regulator or a court might later read.
4. What the twelve add up to
Industry today
With GenAI, properly controlled
Control does not move. It is the constant. Generative AI's share of the work more than doubles. Human oversight drops by about 60%. The deterministic layer sits at the same share in both states, unchanged.
That was not designed into the exercise; it fell out of assessing twelve processes one at a time, and it is the most useful thing in this paper. Generative AI displaces human effort, and it does not displace control. An operation adding generative AI while holding its deterministic layer flat is building this correctly. An operation whose control layer thins as its models grow has removed the thing that made the models safe to run.
The second finding is commercial. The industry runs generative AI at well under half of what existing technology already supports. The constraint is the control layer, which has to be built first, and building one is slower and less interesting than buying a model.
5. What happens to the humans
Human oversight drops by about 60%, which reads as generative AI reducing the human role. The distribution says something more specific.
Four processes hold 40% of all human judgment in a claim today: liability apportionment, subrogation viability, settlement approval and repair method. In the end state the same four hold 50%. People move out of the routine work and are left holding the contested work, and the inputs to all four become machine-produced along the way.
Which makes the human the largest uncontrolled variable in the system. A model can be tested against data it never saw in training, versioned, monitored for drift and inventoried. A control can be read by anyone and audited against the files it governed. Neither applies to a person, and the published record is close to empty: adjuster-to-adjuster variance on comparable motor files has never been measured by any regulator, actuarial body or journal that an August 2026 search reached.
The one occasion anyone measured expert divergence on identical insurance files is a noise audit reported by Kahneman, Sibony and Sunstein, in which adjusters independently assessed written descriptions of the same cases. The median ratio between two adjusters estimating the same claim's ultimate cost was 43%. Most of that insurer's executives had guessed 10% or less, and when 828 CEOs and senior executives across industries were asked the same question, 10% was both the median and the most common answer.
There is a second problem, and it has no entry in any claim file. On processes where generative AI supplies the inputs and a person concludes, liability apportionment above all, the record says a human decided. Nothing records that the human began from a machine's reconstruction.
The research on that condition is consistent. In a controlled reader study, very experienced radiologists scored 82.3% accuracy when the AI suggestion shown to them was correct and 45.5% when it was wrong, with the wrong suggestions inserted by the researchers. In a process-control study reported by Parasuraman and Manzey, roughly 43% of operators missed a false automated diagnosis the first time one appeared, and about half of those had already checked every parameter needed to disprove it.
The governance literature arrives at the same place. Ben Green surveyed 41 policies requiring human oversight of government algorithms and found that any nominal human involvement satisfies the requirement, which creates an incentive to introduce oversight that looks like a control and functions as a signature. Riikka Koulu describes human oversight as easily observable and in danger of becoming "an empty procedural shell used as a stand-in justification for algorithmization."
A signature is not oversight. Oversight is the override rate, the reason code, the share of overrides later reversed, and the gap between what the machine proposed and what the person concluded. All four are countable, and none of them is counted in most claims operations.
6. The end state keeps all three
The triangle has three corners and none of them is a destination. A claims operation works because each carrier's failure is answered by one of the other two.
| Failure | Answered by | How, in practice |
|---|---|---|
| GenAI fails by opacityA confident wrong answer looks exactly like a right one. | Control | A confidence threshold that routes instead of guessing. A hard authority ceiling above which no model output takes effect. A reason code on every machine output, from a closed list. An audit trail that survives the model being retrained. |
| Control fails by rigidityA threshold set against one market keeps answering after the market moved. | GenAI | The model supplies the evidence that the rule has gone stale: realized salvage against the total loss threshold, live parts availability against the schedule, benchmark drift against the estimate tolerance. |
| Humans fail by inconsistencyTwo competent adjusters reach different answers on the same file. | Both | The model narrows the range before the person sees it. The control bounds what the person may conclude without a second signature. The person is spent where judgment is required and nowhere else. |
That last row is the operating case for generative AI in claims, and it inverts the usual one. Generative AI is worth deploying because it makes human judgment more consistent, not because it removes human judgment. A handler working from a normalized estimate, a disclosed comparable set and a bounded authority matrix produces a narrower spread of outcomes than the same handler working from a blank file and thirty years of instinct.
The condition is that the human control has to be designed rather than assumed. A review step that nobody measures produces a signature. What turns it into a control is measuring it.
7. What to build, and UAE compliance
Five things, all answerable from an existing claims operation and none of them waiting on a regulator.
- Measure divergence on the four that carry the money.Liability, settlement approval, subrogation and repair method. For each, how often two handlers reach a different answer on comparable files. No published figure exists for motor, so the first insurer to hold one holds something nobody can contest.
- Separate what the machine proposed from what the person concluded.Two fields, every file, on every process where generative AI supplies the input and a person signs. Without that split there is no way to detect anchoring, and no way to answer a supervisor asking who decided.
- Give every override a reason code and a number.The reason code comes from a closed list. Override rate by reason, and the share later reversed, are the two figures that show whether human review is doing work.
- Write the authority boundary before the model reaches it.For each process, the money figure above which no automated output takes effect without a person. Setting it after deployment means setting it around what the system already does.
- Put the same fields in the vendor contract.Most of this will be bought rather than built: 80% of settlement amount models and 72% of image evaluation models in the NAIC survey came from third parties. An insurer that cannot obtain these fields from a vendor is holding a responsibility it cannot evidence.
What the CBUAE guidance note adds
For UAE insurers, February put a supervisory expectation behind that list. On 11 February 2026 the Central Bank of the UAE issued a guidance note on the responsible adoption of artificial intelligence by licensed financial institutions, with insurance providers in scope. Its legal status is principles-based supervisory expectation: the only "shall" sits in the definitions and the rest runs on "should."
Those definitions do something no other published instrument does. A high-impact decision is defined as any determination using AI that materially affects a customer's access to financial products or services, and an insurance claim is one of the two named examples. The regulator has named the insurance claim itself.
What the note expects against that definition is close to the list above: an inventory of every AI model with purpose and risk rating, audit trails and clear data provenance, explainability with greater force on high-impact decisions, meaningful human oversight with human-out-of-the-loop reserved for low-risk processes, and consumers able to request human review or an explanation. Institutions should not employ models they have no control over, and they remain responsible for AI functions they outsource.
Third-party administrators, claims platforms and AI vendors are not named in the note and are not direct addressees. They are captured through its outsourcing section, which leaves the licensed insurer to demonstrate oversight. The binding version sits in the Risk Management and Internal Controls Regulation for Insurance Companies, which requires outsourced material business activity to be overseen in the equivalent manner to work kept in-house and makes the company fully responsible for the risks arising from any process it outsources.
8. Method and limits
Each of the twelve processes was assessed on one question: of the work this process takes, how much is carried by generative AI, how much by deterministic control, and how much by a person. The three shares are allocated across a fixed total, so a gain for one carrier is always a loss for another and no process can improve on every axis at once. Every position in this paper is one of those allocations, and the triangles are that allocation drawn.
The today position is Axxion's read of industry practice. The end state is what the process can run at with technology that exists now, properly controlled, and it carries no date.
Four of the twelve are informed by published data: claim assignment, damage assessment, total loss valuation and fraud referral. The first three come from the automation and augmentation split in the NAIC private passenger auto survey, calculated across all three of its categories. Figures in wider circulation drop the survey's support column from the denominator and run high as a result. That data fixes the ordering of those processes against each other and does not set the allocation.
The remaining eight carry no external reference, because none was found. The NAIC survey excluded linear regression, generalized linear and generalized additive models, and anything a company could realistically have used in the year 2000 or earlier, so it understates machine involvement. It counts models in use or under construction, so none of its figures describes live deployment alone.
The allocations are judgment, not measurement. The concentration finding in section 5 survives a one-step change to the allocation on any single process. The direction holds and the decimal does not, which is why it is stated as a direction and a rough share rather than to the percentage point.
Jurisdiction. The CBUAE guidance note and the outsourcing regulation apply to UAE licensed institutions. None of the other instruments or datasets binds a UAE or GCC insurer. The NAIC survey is American, EIOPA covers 25 EU and EEA countries, and the research on human oversight of machine output comes from radiology, process control and government algorithm policy. They are included as evidence of direction and because they are the only published material of their kind. No equivalent GCC dataset was found.
Sources
- CBUAE, Guidance Note on the Consumer Protection and Responsible Adoption and Use of Artificial Intelligence and Machine Learning by Licensed Financial Institutions in the U.A.E., issued 11 February 2026.
- CBUAE, Risk Management and Internal Controls Regulation for Insurance Companies, Circular No. 25/2022.
- NAIC, Private Passenger Auto Artificial Intelligence/Machine Learning Survey Results, 8 December 2022, 193 responses across nine states.
- EIOPA, Generative AI Market Survey: Outlook, Use Cases and Risk Management, 2 February 2026, 347 undertakings across 25 countries.
- Capgemini Research Institute, World Property and Casualty Insurance Report 2026, 5 May 2026, 344 executives and 809 employees across 16 markets.
- Kahneman, Sibony and Sunstein, Noise: A Flaw in Human Judgment, 2021, chapter 2.
- Dratsch et al., Radiology 307(4), 2023.
- Manzey, Reichenbach et al. 2008, reported in Parasuraman and Manzey, Human Factors 52(3), 2010.
- Green, Computer Law & Security Review 45, 2022.
- Koulu, Maastricht Journal of European and Comparative Law 27(6), 2020.
Appendix: where the value sits, and how an insurer would measure it
Twelve processes scored against five benefit dimensions, with the measure that carries each one. The value is spread across the whole claim, and no single process carries the case on its own. The map reads in two directions. Across a row: what one process does for the insurer. Down a column: which processes have to move for one benefit to arrive.
The codes inside each cell name the measure. The letter is the benefit family and the number identifies the specific KPI within it: T for turnaround, D for data quality, S for satisfaction, L for leakage and fraud, C for cost. All 22 are defined in the KPI reference below. The row total counts effects, not money: two points for a primary effect, one for a contributing effect, zero for no effect or a trade-off. A process scoring 7 is not worth more than one scoring 4; it is broader, not bigger.
| Turnaround | Data quality | Satisfaction | Leakage & fraud | Repair cost | Total |
|---|
No single process carries the full effect. Twelve processes score between four and seven out of ten, and the five dimensions land between 11 and 16. Value is spread across the whole claim. There is no silver-bullet process, and buying one AI product for one step returns a fraction of what is available. The benefit is a property of the operation rather than of any component in it, which is exactly why point solutions disappoint and why this has to be done end to end.
Two trade-offs
Fraud referral costs time
A referral stops a claim while an investigation runs, so turnaround gets worse on every referred file, and a referral made against an honest claimant is the most damaging single event in a claim. Both costs are real and neither is a reason not to do it. They are a reason to govern it on precision rather than on volume, because a model rewarded for flagging will flag, and the cost lands on customers who did nothing wrong.
The evidence position is honest and thin: Axxion's exception gate has stopped claims pre-authorization on coverage and documentation grounds, which is measured.
Estimate challenge costs time
Contesting an estimate means a conversation with a workshop, and that conversation takes time the vehicle spends off the road. An operation that deploys challenge without watching turnaround will book the saving and lose the service level, and the customer will experience the saving as a delay.
The measure that resolves it is the sustained rate: a challenge that does not hold has cost days and returned nothing. A 47% reduction against the initial workshop estimate is measured at claim level across a live pilot. Sustained rate by reason and by workshop is defined but not yet reported, which is the gap to close before this is presented as a repeatable capability.
The 22 measures
Twenty-two measures carry the five dimensions across all twelve processes. The reference code is the benefit family plus a sequence number, so any cell in the value map can be traced straight to a definition here.
| Ref | KPI | Definition |
|---|---|---|
| T · Turnaround | ||
| T1 | Notification to first touch | Hours from receipt to first substantive action. Median and 90th percentile. |
| T2 | Receipt to survey | Working days, acceptance to inspection complete. |
| T3 | Survey to authorization | Working days, assessment complete to repair authorized. |
| T4 | Authorization to delivery | Working days to handback, banded by damage class. |
| T5 | End-to-end cycle time | Calendar days, notification to handback. Median, never mean. |
| T6 | Touch time against elapsed time | Hours worked over days elapsed. Separates waiting from working. |
| D · Data quality | ||
| D1 | Field completeness at intake | Share of the required field set populated at acceptance, weighted by downstream dependency. |
| D2 | Structured capture rate | Share of claim facts held as fields rather than free text. |
| D3 | Extraction accuracy | Field-level agreement between machine read and source document, on a sampled re-read. |
| D4 | Downstream analyzability | Binary, per analysis: can severity, frequency, cycle time and recovery rate be computed without a fresh data request. |
| D5 | Linkage integrity | Share of payment transactions linkable to a claim, a policy and a vehicle. |
| S · Satisfaction | ||
| S1 | Post-delivery satisfaction | Score at handback within 48 hours, always reported with response rate. |
| S2 | Update promise adherence | Share of open claims receiving a proactive update inside the promised interval. |
| S3 | Complaint and dispute rate | Per 100 claims, split by cause: cost, delay, quality, communication. |
| L · Leakage and fraud | ||
| L1 | Recovery capture rate | Recovered against recoverable identified, with collection lag in days. |
| L2 | Exception catch rate | Claims stopped pre-authorization by reason: coverage, duplicate, documentation, damage inconsistency. |
| L3 | Referral precision | Share of fraud referrals substantiated. Guards against a model rewarded for volume. |
| C · Cost | ||
| C1 | Cost against opening estimate | Reduction from first workshop quote to authorization. Directional; overstates saving. |
| C2 | Cost against matched baseline | Reduction against comparable claims in a control period. The only defensible savings measure. |
| C3 | Cost per claim by severity band | Banded, so mix shift cannot be mistaken for saving. |
| C4 | Challenge sustained rate | Share of contested lines that hold after workshop response, by reason and by workshop. |
| C5 | Rework rate | Same vehicle returning inside 90 days. A saving that produces rework is not a saving. |
The data-quality finding
The same pattern shows up in almost every dataset that reaches Axxion's operation: the fields the analysis needs were rarely captured in the first place. The gap sits primarily at intake, not anywhere downstream.
What that looks like in practice is consistent. Payment transactions arrive without a claim identifier, so several payments cannot be tied to one claim and per-claim severity cannot be produced at all. Exposure is missing, so frequency and burning cost cannot be calculated. Premium is missing, so loss ratio cannot be calculated. Reserves and claim status are missing, so ultimate loss cannot be estimated and averages end up mixing claims at different maturities. Vehicle detail stops at make, so vehicle-age effects cannot be drawn. Often there is no date on the money, which means collection speed cannot be computed even when that is the exact question being asked.
Every field named above is one that generative AI can populate at first notification, from documents that already arrive with the claim, provided a deterministic layer validates it and commits it.
The industry treats data quality as a reporting problem to be solved at the warehouse. It is an intake problem, and it is solved at first notification or not at all.
Testing this on a live book
Everything in this paper is checkable. The allocation across the twelve processes, the finding that control does not shrink as generative AI grows, the concentration of human judgment into four decisions: none of it has to be taken on trust. It can be tested against an insurer's own claims inside a single quarter.
That is what a pilot is for. A defined slice of claims runs through the governed process described here, alongside the existing operation rather than in place of it. The insurer keeps assessment, approval and settlement authority throughout. Nothing is migrated, no core system is touched, and the arrangement can be stopped at short notice.
What the insurer holds at the end.
- A cost reduction with a method attached. Measured on the insurer's own claims, line by line, against its own historical baseline, so the figure can be handed to an actuary rather than to a marketing department.
- Cycle time as a distribution instead of an impression. Stage by stage, banded by damage class, with waiting separated from working. Most operations have never seen their own claims this way.
- The first measurement of its own handler variance. How far two competent people diverge on comparable files. No insurer has published this figure, in any market.
- A governance record that survives a supervisory question. Every machine output recorded as a proposal, every commit logged, every override reasoned and counted.
What has to be in place for the result to mean anything. A pilot without a baseline can only report what happened, not what changed. These are agreed before the first claim is routed, not fitted around the results afterwards: 18 to 24 months of historical repair data banded by severity; a written definition of the measures that count; what the machine proposed and what the person concluded, as two separate fields; override reason codes from a closed list, reported with the override rate and the share later reversed; satisfaction captured at handback within 48 hours with the response rate beside the score; authorization-to-delivery timing banded by damage class; and recovery capture rate with collection lag. Fraud referral sits outside a first pilot until referral precision is tracked.