Motor claims
£11.7bn
Paid out by UK motor insurers in 2024 across 2.4 million claims, with the average claim reaching £4,900. Property and health sit on similar trajectories.
Association of British Insurers, February 2025. [1]
Insurance
AI agents built for UK insurance operations.
Insurance is a high-scrutiny environment. Every output has to be defensible: to regulators, to the ombudsman, to a policyholder who reads every line. We design, build, and run AI agents so your underwriting and claims teams can get back to the judgement calls only they can make.
The Pressure
Volume is up. Margin is tight. And the work still has to stand up.
The work isn't getting smaller and the headcount to absorb it isn't getting larger. Quality is the harder problem. AI alone doesn't solve it.
Claims leakage
11%
Average claims leakage found across UK claims operations in forensic audits, with leakage present in nearly one in three claims.
Claims & Withheld Money Consultants (CWMC). [2]
Lloyd's market
14%
Of Lloyd's market firms that had deployed or tested generative or agentic AI in underwriting as of April 2026. That reflects governance-led caution, not lack of interest.
Lloyd's Market Association / Barnett Waddingham, April 2026. [3]
The Gap
Buying a tool is easy. Trusting its output is the hard part.
Trusting it enough to put it in front of a policyholder, an ombudsman, or a regulator is what has held the market back. That gap is where we build.
Designed around your firm's process, not from a template.
Below are examples of workflow areas we can build agents for. The exact steps and outputs are designed around your firm's specific process during a Discovery session.
What We'd Build
Four areas where an AI workhorse earns its place.
FNOL intake and triage
This could look like an agent that takes a first notification of loss through to a triaged, routed claim.
- Notifications ingested from email, broker portal, web form or phone-note transcription
- Cause of loss, policyholder details and incident facts extracted; the policy on file checked
- Potential coverage issues flagged, fraud indicators run against historic patterns
- Straightforward claims fast-tracked; complex or contentious ones routed to a senior handler with the full context already assembled
The Output
A triaged claim with all relevant fields populated in your claims system, the routing decision logged with reasoning, and supporting documents indexed for the handler.
Broker submission processing
We might build this as an agent that ingests a broker submission and produces a structured risk profile for the underwriter.
- PDFs, MRC slips, statements of fact and free-text emails read and extracted
- Risk data normalised against your schema, missing information flagged
- The submission run against your appetite criteria, exclusions flagged
- Precedent risks surfaced from your own historic book, a preliminary risk note drafted
The Output
A structured submission summary, appetite match flagged with reasoning, missing data listed, and historic comparable risks surfaced for the underwriter's review.
Claims document review
An agent at this stage could read the supporting evidence and produce a structured summary for the case handler.
- Medical reports, surveyor assessments, loss adjuster reports, repair estimates and claims correspondence read in full
- Relevant facts extracted and cross-referenced against the policy wording
- Inconsistencies between documents identified
- The agent's reasoning shown for every flagged finding, with the underlying source alongside
The Output
A concise summary of the supporting evidence, structured for the handler's settlement decision, with the source documents linked and any contradictions flagged.
Complaint handling and FOS response drafting
Or an agent that assembles the file when a complaint is opened and drafts a structured response in your house style.
- The full claims file, correspondence record and original policy wording pulled together
- Precedent decisions retrieved from your own settled cases
- Each complaint point mapped to the evidence, weak points flagged for your senior handler
- The Consumer Duty timeline held, because the file assembly is already done
The Output
A draft FOS or internal complaint response, evidence assembled and cross-referenced, ready for your senior complaint handler to review and finalise.
Design, Build, Run
We design it. We build it. We run it.
Most providers stop at the build. We operate the agent in production, retrain it as your processes evolve, and stand behind the output.
Free 30 minutes
Design
A free 30-minute Discovery session to map the process, agree the agent's scope, and decide whether it's a good fit. If it isn't, we say so.
Fixed scope
Build
Senior engineers build the agent against your specification. Prompts, evaluation criteria, and output formatting are tested against your own examples before anything goes live.
Live operation
Run
We monitor the agent in production, retrain it as your process evolves, and stand behind the output quality. Your team manages the work, not the model.
Skin In The Game
You only pay when the agent delivers.
Insurance buyers have been pitched plenty of AI tools sold on promise. Our model is built the other way around.
There's a refundable commitment fee at the start of each agent, which goes towards the work and is returned in full if we don't deliver to the agreed specification. Once the agent is live, you're billed per output. A failed output isn't billed. There are no per-seat licences and no infrastructure fees on top.
The model exists because insurance work has to stand up. If our output doesn't, our commercial position doesn't either.
The Regulators
Ready for whoever your work answers to.
UK insurance operates under principles-based AI oversight rather than prescriptive AI rules, but the obligations are real and the direction of travel is clear.
Consumer Duty requires firms to demonstrate good outcomes for customers. The FCA is actively scrutinising claims handling, and AI-assisted processes are a specific area of focus.
SMCR means a named senior manager carries accountability for AI-mediated decisions. The agent is a tool; the responsibility stays with a person.
PRA SS1/23 sets model risk management principles that explicitly cover AI and machine learning. Validation, documentation, and ongoing monitoring are all in scope.
ICO guidance on automated decision-making requires documented oversight and meaningful human review at consequential decision points.
The LMA's 2026 AI adoption toolkit sets clear expectations for Lloyd's market participants on governance, testing, and accountability.
All of it points the same way: AI in insurance is acceptable where governance is documented and human oversight is meaningful. We design every agent with that frame in mind. Every output is auditable, every decision is logged with reasoning, and human review sits at the points where it matters.
Our infrastructure runs on isolated Azure tenancy through our parent company Synextra, which has spent many years running governance-led cloud environments for UK businesses.
- FCA · Financial Conduct Authority
- PRA · Prudential Regulation Authority
- ICO · Information Commissioner's Office
- LMA · Lloyd's Market Association
- FOS · Financial Ombudsman Service
- ABI · Association of British Insurers
Common Questions
Questions we hear from insurance firms scoping their first agent.
These are the questions most likely to come up in scoping. The specific answers for your firm sit in a Discovery session, not on a landing page.
How do AI agents work alongside our Consumer Duty obligations?
Consumer Duty applies to outcomes, not to which tool produced the outcome. The agents we build sit upstream of human review at the points where a customer-facing decision is made. The agent assembles, summarises, and drafts. A named handler reviews and approves. Every output is logged with the agent's reasoning, which gives you the audit trail Consumer Duty reviews ask for.
Under SMCR, who is accountable for what the agent does?
Your named senior manager remains accountable for the decision. The agent is a tool that produces drafts and summaries, not a decision-maker. We document the agent's design, validation, and ongoing performance so the senior manager has what they need to discharge their SMCR responsibility. Where the FCA or PRA later clarifies guidance, the documentation is built to flex.
Our core systems are old and we run several of them. Can an agent actually integrate?
Probably, and often without a core system replacement. Most insurance agents we scope read inputs from email, document stores, and shared drives, and write outputs back into the claims or underwriting system through whatever interface that system supports. Where APIs don't exist, we work with the formats your team is already using. Integration is part of the build, not a separate project.
How do we know the output is good enough to put in front of a policyholder or an underwriter?
Before go-live, the agent is tested against your own historic examples. We measure output accuracy against the standard you'd apply to a competent human handler. We don't deploy until the agent meets that standard, and we monitor for drift in production. If the output quality slips, we retrain. Failed outputs aren't billed.
A lot of our submissions and slips arrive as PDFs and email attachments. Can an agent handle that?
Yes. Most of the document-heavy work in UK insurance still flows through unstructured PDFs, broker emails, and MRC slips. Building agents that handle that input reliably is harder than building agents that read clean structured data, and it's the part most off-the-shelf tools get wrong. It's also the part we've built carefully for.
What about firms that have already tried AI and found the output unreliable?
Most AI in the market fails a serious scrutiny test. The reason is almost always thin engineering on prompts, evaluation, and output formatting, combined with no one running the model after handover. Small changes in prompt wording can shift accuracy by a percentage point or two, which at document volume is the difference between useful and unusable. Getting that right before go-live, and keeping it right afterwards, is the job we do.
Start Here Tell us what's slowing
Tell us what's slowing
your operation down.
We'll tell you honestly whether an agent is a good fit for the process you have in mind. If it isn't, we'll say so.
30 minutes. No obligation. No sales pitch.
Sources
Stats cited on this page.
- Association of British Insurers, *UK motor insurers paid record £11.7bn in claims in 2024*, February 2025.
- Claims & Withheld Money Consultants (CWMC), *Claims Leakage*, analysis of forensic audits across UK claims operations.
- Lloyd's Market Association / Barnett Waddingham, *AI adoption more than doubles across the Lloyd's market in 12 months*, April 2026 survey of 39 managing agents representing over 60% of Lloyd's stamp capacity.