Usable photos first time, a likely repair route within hours, and reused or edited images caught before payment. An engineer or handler makes every settlement and total-loss decision.
At a glance
An illustrative scenario. The numbers are typical examples, not a client case study. The cloud services and UK rules were checked on 9 November 2026.
Most motor and home claims now come with photos, and many with video. A customer sends six pictures of a dented wing, taken at night, with no wide shot and no number plate. A repairer sends an estimate with close-ups that don't show the rest of the car. A homeowner sends a video of a wet ceiling. Before anyone can decide anything, an engineer or handler has to work out what the pictures show, ask for better ones, and often book a physical inspection anyway.
In a typical insurer, a large share of motor claims wait days for an engineer's first look, and most of that wait is chasing usable images. Meanwhile the car sits in a garage or a storage yard, hire costs run on, and the customer hears nothing. Edited, reused and now AI-generated images are a growing worry too. The ABI reports that insurers detected £1.34bn of fraudulent general insurance claims in 2025, with motor the largest share, and doctored photos are an increasingly common part of them.
Getting the money right matters as much as getting it fast. The FCA's review of total-loss motor claims found some insurers offered less than a vehicle was worth, applied blanket deductions, and only raised offers when customers complained. That led to around £200m of redress for about 270,000 motorists. Home and travel claims handling is under review during 2026 following the Which? super-complaint. Any AI in this process has to make outcomes fairer and more consistent, not just faster.
Agree these, and measure the baseline, before anything is built. Example targets are typical, not promises.
How long from first notification until an engineer or handler has a usable view of the damage.
ExampleFrom 3–5 days to same day
The share of submissions that need no follow-up request for more or better images.
ExampleAbove 85%
Claims an engineer can assess from images, without a site visit, where they agree that's right.
ExampleUp from about a third to over half of motor claims
How often the engineer disagrees with the suggested route. Very high means it isn't helping; near zero can mean people have stopped checking.
ExampleStable, and reviewed weekly
Manipulated or reused images confirmed by counter-fraud before payment, and honest customers wrongly flagged.
ExampleMore caught, false flags falling
Total-loss offers against guide prices, offers raised after challenge, and outcomes by customer group.
ExampleNo group worse off
Cross-sector rules are on the UK regulation map.
There's no AI-specific rulebook. The FCA applies the Duty to AI like any other change. The photo route must not add friction or sludge: customers can always choose an inspection instead, every suggestion carries a plain-English reason, and outcomes are monitored by customer group.
Claims must be handled promptly and fairly, and not unreasonably rejected. The FCA's 2024 review of total-loss valuations found low first offers and blanket deductions unfair. So the model never sets a value or applies a deduction. An engineer or handler sets every settlement, and valuations are monitored against guide prices.
Write-offs are categorised A (scrap), B (break), S (repairable, structural) or N (repairable, non-structural), and recorded on MIAFTR, now part of the MIB's Navigate platform. The 2025 update added rules for electric and hybrid vehicles, including high-voltage battery damage. The model can flag a possible total loss. An engineer makes and records the category.
The Insurance Fraud Register holds confirmed fraud only. An image flag is a reason to look closer, not proof. Counter-fraud investigates, a person decides whether a claim is fraudulent, and only that decision can lead to a register entry or a rejection.
Photos capture far more than damage: faces, other people's number plates (personal data in the ICO's view), the inside of people's homes, and things that can reveal health or religion. That means a DPIA before the pilot, capture guidance that avoids people, automatic blurring of bystanders and third-party plates, UK processing, a clear retention period, and a contract that stops providers training on your images.
Since 5 February 2026, the strict limit on solely automated significant decisions applies mainly where special category data is involved, and home images can reveal it. Every other automated decision needs safeguards: telling people, letting them challenge it, and a person who can step in. A person on every adverse decision keeps the design clearly within the rules.
Some customers can't take good photos: after a flood, with a disability, or in shock after an accident. The guided link never pressures them, offers help by phone, and flags signs of vulnerability to a trained handler.
Showing AWSAzuresovereign / on-prem services. Switch with the toggle; every diagram and table on this page follows it.
01Channels
02Intake
03Intelligence
04Orchestration
05Systems of record
06People
| Layer | AWS | Azure | Sovereign / on-prem |
|---|---|---|---|
| Guided capture and intakeA link by text or email, not an app download. Quality checks run on the phone as each photo is taken. | A web capture page on Amplify Hosting, API Gateway, S3 | A web capture page on Static Web Apps, API Management, Blob Storage | Your web tier, Kong, Ceph or IONOS S3 Object Storage |
| Original image storeKeep the original file, untouched, with a hash. Every check and every person works from copies. | Amazon S3 with Object Lock | Blob Storage with immutable storage | Ceph with S3 Object Lock, or IONOS S3 Object Storage |
| Reading damage in images and videoAsk for structured output: parts, damage type, severity, and what can't be seen. Test on night-time, wet and partial photos, not showroom shots. | Multimodal models on Amazon Bedrock; Bedrock Data Automation for video | Multimodal models in Microsoft Foundry; Azure Content Understanding for video | Open-weight vision-language models served with vLLM on your own GPUs, or IONOS AI Model Hub |
| Custom models: damaged parts, reuse, edits, redactionTrain on your own labelled claims. Azure Custom Vision retires on 25 September 2028, so don't start new work on it. | SageMaker AI | Azure Machine Learning | MLflow and Kubeflow on Kubernetes |
| Provenance checksRecord content credentials when present. Their absence is normal and proves nothing; their presence is useful but not proof. | The open-source C2PA library, run on AWS Lambda or ECS | The open-source C2PA library, run on Azure Functions or Container Apps | The open-source C2PA library, run on Kubernetes |
| Duplicate and reuse searchImage fingerprints and embeddings compared against every past claim, and against web images where your counter-fraud team agrees. | Amazon OpenSearch Service (vector search) | Azure AI Search (vector search) | OpenSearch, or PostgreSQL with pgvector |
| Agent runtimeThe original Bedrock Agents closed to new customers on 30 July 2026; new builds use AgentCore. | Amazon Bedrock AgentCore | Foundry Agent Service | An open-source agent framework (e.g. LangGraph) on Kubernetes |
| Workflow | AWS Step Functions | Logic Apps or Durable Functions | Temporal, or n8n |
| Middleware and events | Amazon EventBridge, SQS | Event Grid, Service Bus | Apache Kafka or RabbitMQ |
| Link to claims, estimating and valuation systemsReuse the estimating and valuation tools engineers already trust. The model reads their output; it doesn't replace them. | API Gateway, plus your existing integration platform if you have one | API Management, plus your existing integration platform if you have one | Kong or your existing API gateway, plus your existing integration platform if you have one |
| Human reviewThe marked-up images and reasons appear where engineers already work, not in a new screen. | In the engineers' and handlers' existing desktop | In the engineers' and handlers' existing desktop | In the engineers' and handlers' existing desktop |
| Guardrails | Bedrock Guardrails | Foundry guardrails (Azure AI Content Safety) | Open-source guardrails (e.g. NeMo Guardrails, Llama Guard) |
| Identity, keys, data protection | IAM Identity Center, KMS, Macie | Entra ID, Key Vault, Purview | Keycloak or your existing Active Directory; OpenBao or HashiCorp Vault; an HSM |
| Monitoring and auditLog every image check, every suggestion and every human decision, with the reason and the model version. | CloudWatch, CloudTrail | Azure Monitor, Log Analytics | Prometheus, Grafana, OpenTelemetry |
| UK data residencyCheck each model and media service is offered in a UK region before you commit. IONOS AI Model Hub runs in German data centres: EU, not UK, residency. | eu-west-2 (London) | UK South / UK West | Your own UK data centre, or a UK or EU provider |
On a phone, the table shows the platform you pick. Sovereign / on-prem means open, portable components you can run in your own data centre or with a UK or European provider; IONOS is shown as one example, and others exist. What sovereign cloud means →
Many organisations keep core systems and sensitive data on their own servers. This is how the same design works across both.
Stays on-prem
The core claims system, the estimating and valuation tools, the fraud engine and the long-term image archive can all stay where they are. Many insurers keep claims images on their own storage for years, and nothing here needs them to move.
Uses cloud AI
The multimodal damage reading and the agent can run in AWS or Azure. The integrity and reuse checks are a good fit for your own estate, because they need the original files and the full archive of past claims.
How they connect
A private link (AWS Direct Connect or Azure ExpressRoute) and the API layer. Originals stay in your store; only redacted working copies and the fields the model needs are sent, and results come back as events onto your queue.
If the link drops
The capture page and image store keep working, so customers can still send photos. Engineers assess from the images directly, and the backlog is read when the link returns.
What each agent may do through its tools, what it may not, and where a person signs off. The limits are enforced by the platform, not the prompt.
Photo capture assistant
Where a person signs offThe assistant decides nothing about the claim. A customer who stops part-way is passed to a handler, not left waiting.
Damage assessment agent
Where a person signs offAn engineer or handler signs off every repair route, every total-loss decision and salvage category, and every settlement figure. Counter-fraud decides every suspected-fraud case. The agent's permissions are enforced by the platform's identity and policy controls, not by instructions in a prompt.
Images are large and arrive in bursts after storms and floods. Uploads go straight to object storage, and a small event (claim ID, file, hash) goes on the queue. The models work through the queue at their own pace, so a surge never stalls the upload page.
The claims, estimating and valuation systems are reached only through the API layer, never directly. If you already run an integration platform, the AI uses it like any other system. Every message carries an ID, so a retry never attaches the same assessment twice.
The AI can be switched off without stopping the business. If it's paused, photos still arrive and are stored, and engineers assess them exactly as they did before.
Map today's process, measure time to first assessment and inspection rates, sample last quarter's photos, and pick the first claim type (often low-speed motor damage).
Gate · must be true to passBaseline agreed, image sample labelled by engineers, DPIA started, counter-fraud and compliance involved.
Guided capture live for one claim type; the damage model and integrity checks in shadow mode, compared against engineers' decisions.
Gate · must be true to passAgreement and override rate within limits, false fraud flags acceptable, no customer group worse off, compliance sign-off.
Suggestions go live in the engineers' desktop for the first claim type, with monitoring, support and a rollback plan.
Gate · must be true to passMonitoring, support model, rollback plan and engineer training in place.
Add home escape-of-water and storm claims, then repairer-submitted images, one at a time.
Gate · must be true to passEach new claim type passes the same pilot gate, with its own capture guide.
| Activity | Claims director | Engineering lead | Solutions architectI can fill this | Technical PMI can fill this | EngineersI can fill this | Counter-fraud lead | Compliance & DPO | SecurityI can fill this |
|---|---|---|---|---|---|---|---|---|
| Business case and benefits | Accountable | Consulted | Consulted | Responsible | Informed | Consulted | Consulted | Informed |
| Target architecture | Informed | Consulted | Accountable and Responsible | Consulted | Consulted | Consulted | Consulted | Consulted |
| Capture guides and image labelling | Informed | Accountable and Responsible | Consulted | Consulted | Responsible | Consulted | Consulted | Informed |
| DPIA and Consumer Duty review | Accountable | Consulted | Consulted | Consulted | Informed | Consulted | Responsible | Consulted |
| Integrity checks and fraud referral rules | Informed | Consulted | Consulted | Consulted | Responsible | Accountable | Consulted | Consulted |
| Build and integration | Informed | Consulted | Accountable | Responsible | Responsible | Informed | Informed | Consulted |
| Accuracy testing against engineers' decisions | Informed | Accountable | Responsible | Consulted | Responsible | Consulted | Consulted | Informed |
| Go-live decision | Accountable | Responsible | Consulted | Responsible | Informed | Consulted | Consulted | Consulted |
| Run, monitor and improve | Informed | Accountable | Consulted | Consulted | Responsible | Consulted | Consulted | Consulted |
R ResponsibleA AccountableC ConsultedI Informed· not involved
Option A
If there's no architect on the programme, I'd own the design end to end: the capture flow, the image pipeline and integrity checks, the integration with your claims and estimating systems, and the gates. I'd report to your programme board.
Option B
If your partner or your own team already has an architect, I'd take on the image pipeline, the vision models and the claims-system integration under their lead, and give them a second senior opinion on the rest.
Goes wrong
The model was trained on clean, daylight photos. Real claims arrive at night, in the rain, cropped and compressed.
How the design avoids it
Pilot on last quarter's real images, labelled by your own engineers, every claim type and channel.
Goes wrong
The model's guess becomes the offer, and settlements drift low.
How the design avoids it
The model never sees or sets a value. Engineers set every figure, and valuations are monitored against guide prices.
Goes wrong
Image flags are treated as proof, and honest customers are accused.
How the design avoids it
A flag starts an investigation, never a decline. Measure false flags and review them with counter-fraud every week.
Goes wrong
The photo route becomes the only route, and customers who can't use it are stuck.
How the design avoids it
An inspection or a call back is offered at every step, and drop-outs go to a handler.
Goes wrong
Images are kept forever and shared too widely, including faces and the inside of homes.
How the design avoids it
Redact bystanders and third-party plates, limit who sees originals, and set a retention period in the DPIA.
Goes wrong
Content credentials are read as a seal of truth.
How the design avoids it
Treat them as one signal among several. Missing credentials are normal; present ones can still be wrong.
Changes already announced or under way that affect this design. More in the UK horizon scan.
Keep image-integrity evidence in a structured, explainable form, so your counter-fraud team can share confirmed cases through whatever route the IFB provides.
Expect sharper expectations on delays, communication and claim acceptance in home claims. Report outcomes for photo-assessed home claims separately from the start.
Regulations in force since May 2026 require the ICO to produce a statutory code. Keep the decision log, the explanations and the human review records in a form you could show the ICO.
Read and store credentials whenever they're present, and design the checks so they get stronger as signing spreads, without penalising customers whose phones don't sign.
If you already run damage or parts classifiers on Custom Vision, plan the move to Azure Machine Learning or Foundry models now, with your labelled images exported.
Good answers are specific. Vague answers to any of these are worth following up.
Whether you need someone to lead the design or to support the architect you already have, let's talk it through.
