Forward-Deployed Engineering · Insurance · Motor claims intake and triage
Case study: motor claims, from first notice of loss to a complete file
Illustrative engagement — not a client record. The insurer, people, volumes and results below are a representative composite, written to show how a forward-deployed engagement runs end to end. The workflow, the architecture, the controls and the method are real and technically valid.
Offering usedPilot-to-ProductionA stalled pilot hardened, integrated and put in front of real users.
Illustrative results, first 90 days after go-live, in-scope claims only:
1.1 days
Median time from intimation to a complete file with a handler
Before 3.8 days
12%
Claims needing more than one request to the customer for documents
Before 41%
99.2%
Field accuracy of data written to the claims system, weekly audit sample
Before not measured
46%
Fraud referrals the investigation team accepted for enquiry
Before 18%
Explainer09 entries
The whole story in about a minute
Nine short scenes, from claims arriving in pieces to the results. Press play, or pick a scene.
Entry 01 / 09
BeforeSummary
Claims arrived in pieces, over days.
Figure
3.8
days, the median to a complete file with a handler
Description
About 32,000 motor claims a month came by phone, app, email and garage, often with no claim number. A complete file took a median 3.8 days to reach a handler.
Particulars
- 01About 32,000 claims a month
- 02Four channels, often no claim number
- 0341% asked for documents more than once
Filed · Claims in pieces
Read chapter 1About 32,000 claims a month
Intimation to handler
3.8 days
median, for a complete file
41%
of claims needed more than one document request
Pieces of one claim
Clerks, then handlers
At a glance
- Client
- A mid-sized Indian general insurer. Motor is its largest line: private cars and two-wheelers, around 32,000 own-damage claims a month, four claims hubs and a national network of garages
- Workflow
- First notice of loss to a complete, checked claim file: read and classify what arrives, check completeness for the claim type, write fields into the claims system, flag inconsistencies to investigators, route to a handler with a summary
- Engagement
- Pilot-to-Production (eight to twelve weeks); this one ran ten weeks and four days
- Team
- One named senior forward-deployed engineer, with a second engineer for the integration weeks. Client side: the Head of Motor Claims (owner), a claims systems analyst, a platform engineer, the head of the Fraud Monitoring Unit, a security reviewer, the compliance officer for policyholder protection
- Where it runs
- The insurer's own cloud account, Indian region, with the model reached through a private endpoint that retains nothing
- Handover
- Go-live criteria ticked by the client; runbook executed by their engineers in the last fortnight; our access revoked at the gate
| Measure | Before | After |
|---|---|---|
| Median time from intimation to a complete file with a handler | 3.8 days | 1.1 days |
| Claims needing more than one request to the customer for documents | 41% | 12% |
| Field accuracy of data written to the claims system, weekly audit sample | not measured | 99.2% |
| Fraud referrals the investigation team accepted for enquiry | 18% | 46% |
| Claims denied, reduced or closed by the system | not applicable | 0, by design |
FDE / 01Case study
01 / 12
In shortClaims arrived in pieces through four channels, and a complete file took days to reach a handler.
A motor claim starts as a phone call, a few taps in the app, an email, or a photograph sent by a garage. What follows is paperwork: the claim form, the driving licence, the registration certificate, the policy copy, photographs of the vehicle, a repair estimate, a police report where one is needed, bank details for payment, and identity documents above a payout threshold.
The insurer received about 32,000 own-damage claims a month. Roughly 40% began at the call centre, a quarter in the app, 15% by email to a shared mailbox, and a fifth from network garages when a damaged vehicle was driven in. A claim arrived in pieces, over days, through different channels, often with no claim number on the attachment.
Intake was done by a team of registration clerks and then by the handlers themselves. The consequences were ordinary and expensive:
Cycle time. The median time from intimation to a file complete enough for a handler to work was 3.8 days. The regulator's own clocks — a surveyor appointed within a day of intimation, the surveyor's report inside a fortnight, a decision within days of that report — were being met by handlers working around an intake process, not helped by one.
Repeat requests. 41% of claims produced more than one request to the customer for documents. Each round cost days and produced complaints, and the 2024 general insurance reforms had made it plain that insurers should ask once, ask only for what settling the claim requires, and never ask for what underwriting already holds.
Fraud found late. Inconsistencies — a photograph taken before the reported date of loss, a licence not valid for the vehicle class, the same dented bonnet in two claims — were caught by experienced handlers, or not at all. Referrals to the investigation team were noisy; most were returned.
Routing. Claims reached whoever was next in the queue. Total loss and theft files, which need different handling, arrived with everyone.
FDE / 02Case study
02 / 12
In shortA photo pilot tried to price and approve claims, was never connected, and showed a fraud score with no reasons.
Eight months earlier the digital team had run a pilot with an outside vendor on photograph-based damage estimation. It showed well: upload four photographs, receive a repair estimate, approve small claims automatically. It never reached production.
It aimed at the wrong decision. The pilot's purpose was to approve and price claims. The Head of Motor Claims and the compliance officer would not accept a model setting or reducing what a customer was paid.
Photographs cannot see the whole loss. On the insurer's own claims, the estimates were reasonable for bumpers and panels and wrong where it mattered: a wheel sitting out of line, a bent member, a cracked mounting. Published experience elsewhere says the same. Hidden damage is why surveyors exist.
It was never connected. The pilot read an export of photographs and wrote nothing back. Every field still had to be typed.
It leaked a judgement to the wrong place. The pilot showed handlers a single fraud score with no reasons. In one call a handler told a customer their claim had been "flagged by the system". That incident, more than any technical finding, is why the pilot was stopped.
Nothing scored it. Accuracy was the vendor's number on the vendor's sample.
The pilot was not worthless. Its document classifier and much of its field extraction were sound, and the code was the insurer's. What it lacked was the work between a demonstration and a claims operation: integration, security, evaluation, monitoring, and a clear line about which decisions are a person's.
FDE / 03Case study
03 / 12
In shortWeek one found the real first problem, matching documents to claims; the rules for going live were signed in week two.
Week one produced a gap report: what the pilot had, what production needed, and what would be thrown away. Photograph-based estimation left the decision path entirely. It survives only as two safe checks — whether the photograph set is complete for the claim type and legible, and a list of visibly affected panels for the surveyor to confirm or correct. The system never prices a repair.
What the week also found:
FIG. 3.1- A false assumption. Everyone believed claims reported by telephone arrived without documents. In fact the call centre sent a link, and the customer's photographs and licence arrived hours or days later, in a different channel, usually without the claim number. Matching those fragments to the right claim was the real first problem, and nobody had named it.
- The claims system had a supported interface for creating and updating claims, already used by the app. Write-back could use the same path and the same permissions a clerk has.
- The insurer was entitled, under existing arrangements, to verify registration and licence details against the national vehicle and driving licence databases, and to look up total-loss and theft history in the industry data repository. Nobody in claims was using them at intake.
- The Fraud Monitoring Unit had a board-approved anti-fraud policy and a list of red flag indicators, written for the framework that takes effect in April 2026. They were in a document, not in a system.
The go-live criteria, drafted in week one, signed by the owner in week two:
- WorkflowOwn-damage motor claims for private cars and two-wheelers, all four channels, from intimation to a complete, checked file routed to a handler
- Out of scopeThird-party injury and death claims, commercial vehicles, the surveyor's assessment, the settlement decision, payment, repudiation
- MetricMedian time from intimation to a complete file with a handler; share of claims needing more than one document request; 99% field accuracy on the golden set
- OwnerHead of Motor Claims; fraud indicators owned by the head of the Fraud Monitoring Unit
- GuardrailsThe system never denies, reduces, delays or prices a claim. Fraud indicators go only to the Fraud Monitoring Unit, with reasons, and are never visible to the customer, the garage or the intermediary. Surveyor appointment stays in the existing process
- Exit gateGolden-set score above threshold, security review signed, monitoring live, rollback rehearsed, one named owner trained, two weeks of shadow-run parity
Timeline09 Stages
The engagement, stage by stage
Nine stages on one clock, from claims arriving in pieces to ninety days after go-live. Press play, or pick a stage.
Note
Claims arrived in pieces, over days.
Detail
About 32,000 motor claims a month came by phone, app, email and garage, often with no claim number. A complete file took a median 3.8 days to reach a handler.
Particulars
- 01About 32,000 claims a month
- 02Four channels, often no claim number
- 0341% asked for documents more than once
Filed · Claims in piecesBefore
Read chapter 1FDE / 04Case study
04 / 12
In shortTen weeks and four days: connect it, secure it, test it, watch it, switch it on.
four days
W1–W2Assess
What happened
The pilot read line by line; photograph estimation removed from the decision path; go-live criteria signed
What existed at the end
A gap report and signed criteria
What happened
The pilot read line by line; photograph estimation removed from the decision path; go-live criteria signed
What existed at the end
A gap report and signed criteria
What happened
Intake from the call centre's records, the app, the shared mailbox and the garage portal; the fragment-matching service; write-back through the claims system's interface
What existed at the end
Claims created and updated by the service, under a service account with a clerk's permissions
What happened
Threat model, private model endpoint inside the insurer's account, secrets in their vault, egress controlled, log retention set to the longer of the periods their rules require
What existed at the end
A security review signed by the insurer's team
What happened
Golden set built from past claims; red-team slice; harness in the insurer's pipeline. Access to the licence and registration verification services needed a legal addendum, which took six working days, so the clock paused, in writing
What existed at the end
An evaluation suite that gates every release
What happened
Dashboards, thresholds, paging, rollback rehearsed on the production path
What existed at the end
Monitoring live and a rehearsed rollback
What happened
10% of intimations, then half, then all in scope, each step gated on the shadow-run comparison
What existed at the end
Real claims routed, criteria ticked, runbook signed
four days
Calendar time was ten weeks and four days. The four days were the paused clock in week four, recorded the day it stopped.
FDE / 05Case study
05 / 12
In shortA system that assembles and checks the claim file, routes it, and leaves every decision to people.
Try a claim05 / 07
Pick what arrives. Watch where the claim goes.
The same pipeline reads everything that arrives. What it finds decides whether the file goes to a handler or to a person first.
Routing slipStep 07 / 07
Claim log01
- 01IntakeLands in one place in the insurer's account, hashed on arrival
- 02Match to claimLinked to the claim by registration and policy number
- 03ClassifyLabelled: claim form, licence, registration certificate, photographs
- 04ExtractFields filled; each one shows the page it came from
- 05Rules checkComplete for the claim type; licence valid on the date of loss
- 06RouteBy claim type, estimate band, location, skill and load
- 07Handler queueFile, narrative, checklist result and evidence
Claim status
- Claims it denied
- 0
- Runs in
- Insurer's cloud
Intake. Every channel lands in one place inside the insurer's account: the call record and its transcript, app uploads, mailbox attachments, garage uploads. Each item is hashed on arrival, which catches the same photograph sent twice and the same photograph sent for two claims.
Match. The fragment matcher links loose documents to the right claim by registration number, policy number, mobile number, claim number where present, and the time window. Its confidence is scored. Below the threshold the item goes to a clerk's queue rather than being attached to a claim on a guess.
Classify. A document model labels each item: claim form, driving licence, registration certificate, policy copy, police report, repair estimate, bank details, identity document, vehicle photograph, or something else.
Extract. A layout-aware document model reads the page; a language model, reached through a private endpoint in the insurer's region with no retention, fills a strict schema — registration and chassis numbers, licence number and validity, class of vehicle, date, time and place of loss, police report number and date, estimate lines and totals. Each field carries the document and the region of the page it came from. A speech-recognition model hosted in the insurer's region transcribes the intimation call, and the language model writes a short narrative of the loss for the handler, with each statement linked to the part of the transcript it came from.
Check. Deterministic rules, not a model:
Completeness by claim type. A theft or total loss file needs documents an ordinary dent does not. The checklist is the insurer's own, and it is the claim type that selects it.
Consistency. Registration and chassis numbers against the registration record; licence valid on the date of loss and for the class of vehicle; policy in force on the date of loss; police report date against the reported loss date; estimate arithmetic; bank details against the policy holder's name.
Red flag indicators, exactly as the Fraud Monitoring Unit wrote them: intimation soon after a policy was issued or revived, photograph metadata inconsistent with the reported time or place, the same damage appearing in another claim, a vehicle with prior total-loss history, repeated combinations of garage and claimant. Each indicator is a named rule with a written reason.
Route. Rules route: claim type, estimate band, location, handler skill and current load. Theft and total loss go to the specialists. The handler receives the file, the narrative, the checklist result and the evidence behind each field.
Fraud indicators go to investigators only. A triggered indicator produces an item in the Fraud Monitoring Unit's queue with the rule, the evidence and the documents. There is no score on the handler's screen, no flag in any customer-facing message, and nothing the garage or intermediary can see. The investigation team decides what to enquire into. A claim under enquiry follows the insurer's existing process and its published timelines.
Nothing is decided. The system does not deny, reduce, price or close a claim, and it does not appoint the surveyor. It prepares a file and says what is missing. Handlers, surveyors and investigators do the rest.
One composite request. Where something is missing, the system drafts a single request listing everything needed for that claim type, for the handler to approve and send. Documents already held from underwriting are never asked for again.
System map18 objects
How the pieces connect
Channels, rules, people and the claims system, lit one scene at a time. Press play, or pick an event.
Timeline
OntologyMotor claims intakeClaims in pieces
In this sceneNot in this sceneZoom 6×4
ObjectSourceBefore
Four channels
32,000 a month
01Properties
Claims arrived in pieces, over days.
3.8Metric · days, the median to a complete file with a handler
- About 32,000 claims a month
- Four channels, often no claim number
- 41% asked for documents more than once
02Description
About 32,000 motor claims a month came by phone, app, email and garage, often with no claim number. A complete file took a median 3.8 days to reach a handler.
03Linked objects3 Links
- ClerksQueueIn this scene
- HandlerPersonIn this scene
- IntakeServiceNot in this scene
Graph18objects3selected0events
Read chapter 1FDE / 06Case study
06 / 12
In shortA test of 2,000 past claims, plus fraud and red-team slices, that every release must pass.
2,000 past claims sampled across channels, claim types, vehicle types, states and seasons, including monsoon months when photographs are worst.
Answers taken from the closed file: the fields as finally corrected in the claims system, the documents that were actually needed, and the claim type as finally classified.
A fraud slice, chosen with the Fraud Monitoring Unit: 120 past claims later confirmed as fraudulent, and 400 that were referred and cleared. The cleared ones mattered most; the harness measures both what the rules catch and how many honest customers they would have disturbed.
A red-team slice: a licence expired three days before the loss, a police report for a different vehicle, an estimate with lines that do not total, a photograph set reused from an earlier claim, two claims for one dent.
Field-level scoring with a 99% threshold on critical fields, recall measured on missing-document detection, and a ceiling on the false-referral rate that the investigation team set themselves.
The suite runs in the insurer's pipeline. In week seven it stopped a release: a change that improved reading of one state's registration certificates started swapping the chassis and engine numbers on another state's older format. Nothing reached a claim file. The failing layout joined the golden set.
FDE / 07Case study
07 / 12
In shortEverything runs in the insurer's own cloud, and the system can do only what a registration clerk can.
Perimeter. Everything runs in the insurer's own cloud account in an Indian region. The model is reached through a private endpoint that retains nothing. Egress is blocked at the edge and the block was tested.
Identity. The service account can create and update claims and attach documents, exactly as a registration clerk can, and nothing else. It cannot approve, pay or close. Our engineer's access ran through the insurer's identity provider and was revoked at the gate.
Can
- create and update claims
- attach documents
Cannot
- approve
- pay
- close
Logs. Every read, extraction, rule result, route and write is attributed and timestamped in the insurer's log store, kept for the longer of the periods their information-security and data-protection rules require. The cyber incident reporting steps are in the runbook.
Autonomy is bounded. A model reads, transcribes, classifies and drafts. Deterministic rules check and route. People decide: the handler on the claim, the surveyor on the damage, the investigator on a red flag, the authorised officer on anything declined.
Fair treatment. Nothing the system produces is disclosed to a customer as a reason. Requests to customers are written in plain language and sent once. The compliance officer reviewed every customer-facing template before go-live.
FDE / 08Case study
08 / 12
In shortA fortnight side by side with the clerks, then switched on in three steps, with one flag to roll back.
The shadow run was the gate, and it ran a fortnight. The system prepared every in-scope claim while clerks and handlers worked as before. Its files were compared with theirs, and its red flag indicators with what the investigation team had actually pursued. Disagreements sorted into three piles: the system missed something, the person missed something, and the checklist was wrong. The third pile went back to the Head of Motor Claims; two document requirements were dropped as things the insurer already held.
it ran a fortnight
the system missed something
the person missed something
the checklist was wrong
went back to the Head of Motor Claims
Go-live went by share of intimations, not by date: 10%, then half, then all in scope. Each step needed a clean day of audit samples and a clean day of referral quality. Rollback was one flag that sent every claim to the clerks' queue. It was rehearsed in week eight on the production path.
0110%
02half
03all in scope
EACH STEP · a clean day of audit samples and a clean day of referral quality
FDE / 09Case study
09 / 12
In shortThe insurer's engineers proved they could release, roll back and add a red flag before we left.
In the dry run the insurer's engineers released, rolled back and added a red flag indicator written by the Fraud Monitoring Unit that morning, scoring it against the fraud slice before it went live. Our engineer watched.
Key numbers09 Items
The story in nine numbers
One number for each scene, from 3.8 days to a complete file down to 1.1. Press play, or pick a number.
Items 01 / 09
Endorsement 01
Before3.8
days, the median to a complete file with a handler
Particulars
Claims arrived in pieces, over days.
About 32,000 motor claims a month came by phone, app, email and garage, often with no claim number. A complete file took a median 3.8 days to reach a handler.
- About 32,000 claims a month
- Four channels, often no claim number
- 41% asked for documents more than once
No claim number
FiledClaims in pieces
Read chapter 1FDE / 10Case study
10 / 12
In shortFiles reach handlers sooner, fewer customers are asked twice, and no claim is decided by the system.
Figures are illustrative, measured on in-scope claims over the first 90 days after go-live.
A complete file reached a handler in a median of 1.1 days, down from 3.8. Most of the gain came from fragment matching, not from extraction.
Claims needing more than one document request fell from 41% to 12%. The single composite request did that.
BEFORE41%AFTER12%99.2% field accuracy on the weekly audit of data written to the claims system, against a 99% threshold.
AFTER99.2%Referrals the investigation team accepted rose from 18% to 46%, on a smaller number of referrals. Each one arrives with the rule and the evidence, so it can be judged in minutes.
BEFORE18%AFTER46%No claim was denied, reduced or closed by the system, and no fraud indicator was shown outside the Fraud Monitoring Unit, because the design does not allow either.
No handler role was cut. Two registration clerks moved to the exceptions queue, where fragment matching is deliberately conservative.
What did not improve, and was never promised:
FIG. 10.2Surveyor report times did not change. They are the surveyors' to meet.
Garage estimates are as rough as they always were, and the system does not price repairs.
Night-time two-wheeler photographs are still read poorly. They go to a person.
Third-party claims were untouched. They were out of scope and the scope page said so.
The overall claim settlement ratio did not move. Files arrive sooner and cleaner; what is payable is unchanged.
FDE / 11Case study
11 / 12
In shortSix lessons, starting with: automate the intake, not the judgement.
- Automate the intake, not the judgement.
Everything difficult about this workflow is assembly. Everything dangerous about it is decision.
- Find the real first problem.
It was not reading documents. It was knowing which claim a photograph belonged to.
- Fraud flags are investigators' property.
A score on a handler's screen ends up in a sentence to a customer. Reasons, to a named team, only.
- Measure the false referrals, not only the catches.
The investigation team should set that ceiling themselves.
- Keep the pilot's good half.
The classifier was sound. The decision it was pointed at was not.
- Photographs do not show a bent member.
Anything that prices repairs from pictures will be wrong where money is.
FDE / 12Case study
12 / 12
In shortThe insurer kept the system and extended it to commercial vehicles on its own.
The insurer kept the system and its team. Two months after the gate, their engineers extended intake to commercial vehicles using the same checklists mechanism. A second engagement was scoped for the reimbursement side of their health line, where the same assembly problem exists with different documents.