AI for Insurance in the UK: Where It Pays in 2026

UK insurance is running two problems at once: claims cost is rising in a motor market that barely breaks even, and under Consumer Duty the regulator now expects firms to evidence good outcomes, not just report them. UK motor insurers are set to only break even in 2025 at a net combined ratio of 101%, and EY forecasts that ratio rising to 111% in 2026 as claims costs outpace premiums. Insurers paid out £11.9 billion across 2.5 million motor claims in 2025, according to the Association of British Insurers, so every pound of claims-handling cost now comes off a shrinking margin.

That is the context a Claims Director or COO is buying against, and it is why "where does AI pay" is a sharper question in insurance than a capability tour can answer. This guide maps where AI earns its place across UK claims, fraud, underwriting and conduct operations in 2026, which regulator duty sits behind each one, and where each DILR line fits. It is a hub: the deep mechanisms live in dedicated posts, and this page points at them rather than repeating them.

This guide is shipped by the team behind Dilr Voice, a multi-agent voice AI platform that runs FNOL and claims intake, renewals and proactive outbound updates on the phone. Or see DATS, the senior-led consulting system that places AI inside an insurer's own claims and underwriting operations.

Where does AI pay first in UK insurance?

AI pays first in UK insurance where volume, cost and a regulator's attention meet, which in 2026 means the claims journey and the fraud queue. Claims-handling cost weighs directly on margin, sharply so in motor, complaint volumes rise when nobody calls the claimant first, and fraud detection is climbing faster than investigation capacity. Underwriting submission handling and Consumer Duty evidence follow close behind. Each of those is a measurable operational number, not an AI project.

The rest of this guide takes those in turn. The pattern across all of them is the same: a repetitive, high-volume task that a regulator now watches, where a governed AI system removes cost or produces evidence without adding headcount. Insurance is a strong first market for this precisely because the numbers are public and the duties are written down, so the business case can be built from sourced figures rather than a vendor's promise. It also means the wrong first move is easy to spot: any AI project that cannot name the operational number it will move, or the regulator duty it will help evidence, is a project looking for a problem rather than solving one. Our industry hub for voice AI sets out the same test across sectors, and the sibling guides for banking and fintech apply it to adjacent regulated markets.

What is a loss-making motor market doing to the case for AI in claims?

A loss-making motor market turns every avoidable claims-handling cost into a direct margin problem, which is what makes AI in claims a board-level decision. When the net combined ratio sits above 100, meaning claims and costs exceed premium income, the money to fund better service has to come from lower handling cost rather than higher premiums. That is why automating repetitive contact and triage carries a real, measurable return for a UK motor insurer.

The numbers make the point plainly. UK motor insurers are set to only break even in 2025 at a net combined ratio of 101%, after a profitable 97% in 2024, and EY expects losses in 2026 at a forecast 111%. That trend is the single clearest financial signal in the sector, and it is worth seeing as a curve rather than a snapshot: a ratio moving from below 100 to 111 in three years is a structural squeeze, not a bad quarter, and it changes which cost lines an insurer can afford to leave manual.

The cost pressure is not confined to motor. Property insurance payouts reached £6.1 billion in 2025, the highest annual total on record, the ABI reports, with £1.5 billion in the final quarter alone as adverse weather drove a surge. A weather-driven surge is exactly the moment a fixed contact-centre headcount cannot keep up with proactive updates, and it is where a voice platform that scales with call volume, rather than a fixed rota, absorbs the peak. The deep version of that argument sits in the voice AI claims intake playbook; here it is enough to say the cost problem is broad and cyclical.

What did the FCA's claims-handling review tell insurers to fix?

The FCA's claims-handling review told insurers to fix how they oversee outsourced services and how quickly they settle, and it placed the duty on the insurer that owns the policy rather than on a distributing broker. Under Consumer Duty, an insurer now has to be able to show what its outsourced handlers are doing and that claimants are being kept informed, not merely report that outcomes are good. Both are places where a governed AI system helps directly.

The regulator, following its 2024 review of the motor market and the home and travel sector, described poor claims-handling practices that included delays in settling claims and high complaint volumes, and it noted that in its sample only 32% of storm damage claims in 2024 resulted in a payment. In the review's own words, the FCA found "Lack of oversight of outsourced services, resulting in poor customer outcomes, delays in settling claims and high complaint volumes". A voice agent can make the proactive status call that stops a claimant chasing, and a document intelligence layer, delivered through our AI execution office, can surface the management information the review said firms lacked. The Financial Ombudsman Service is the backstop when that communication fails, and reducing avoidable complaints is the cheapest way to stay out of it.

How do insurers evidence Consumer Duty in their board reports?

Insurers evidence Consumer Duty in board reports by showing that they monitor outcomes across customer groups, and the FCA's review of those reports named this a common area for improvement. The duty binds the insurer's board and senior management, not a software supplier, so the evidence must be produced as a by-product of the operation rather than assembled by hand before a deadline. This is where AI earns its place on the evidence side rather than the contact side.

The regulator named better data quality as a specific area for improvement, noting that some firms lacked sufficient data quality to justify their conclusions or to give their governing body adequate assurance. A document and outcome-analysis layer that reads complaints, claims decisions and communications can generate that monitoring evidence continuously, so a board report rests on live data rather than a fortnightly scramble. The same evidence problem shows up in banking, where conduct data quality is the recurring FCA finding, which is why the evidence layer belongs inside the operation and with the insurer.

Where does voice AI fit in claims intake and renewals?

Voice AI fits in insurance wherever a phone call is repetitive, time-critical and measured, which in claims means first notification of loss and status updates, and in servicing means renewals and broker queries. Dilr Voice is an enterprise voice AI platform from DILR.AI that chains specialised agents into a single call, so a claims line can greet the caller, capture structured FNOL data and trigger the next system action without a human picking up.

It runs an inbound front desk with after-hours routing and a warm transfer to a person when a call needs one, and it handles proactive outbound status updates on a schedule so a claimant hears from the insurer before they call to chase. The detailed mechanics live in their own guides rather than here. The FNOL and claims intake playbook covers how a 24-hour line keeps data completeness high and complaints low at once, and the broker renewals guide covers renewal conversion and query handling. At map level, what matters is that Voice answers the FCA's communication finding and the property-surge staffing problem directly, ships per-country compliance rules by default such as recording consent and full call audit trails, and does it without expanding the contact centre. For the full picture across sectors, the voice AI enterprise pillar sets out how the platform is built and governed.

The same diagnostic logic underpins our AI operating model consulting, which sets the governance and human-escalation rules for a claims line before a single call touches a live policyholder.

When does fraud triage outgrow the special investigation unit?

Fraud triage outgrows the special investigation unit when detected volume rises faster than the headcount available to investigate it, which is the pattern in the 2024 fraud figures. Flat investigation capacity against a rising detection curve means either leakage into paid claims or slower settlement for honest claimants, and both cost money when general-insurance margins are already thin. This is where triage automation pays, by putting investigator time on the likeliest cases.

Insurers uncovered over 98,400 fraud-related claims in 2024, up 12% from 88,100 in 2023 and worth £1.16 billion, the ABI reports, and prevented an estimated 684,800 fraudulent applications, up 7.4%. AI pays here by triaging referrals so investigators focus on the strongest cases, and by assembling the evidence pack before a human opens the file. A well-triaged queue also protects the honest claimant, because time not spent chasing weak referrals is time available to settle genuine claims quickly, which is the same complaint and delay problem the FCA raised on the claims-handling side. This is a DATS delivery rather than a voice one, and the deep treatment of how referrals change when they arrive pre-assembled belongs in a dedicated post rather than this hub, which maps the shape of the pressure rather than the mechanism.

What can a private model safely extract from claims files?

A private model can safely extract structured data from the medical documents in a claims file, such as scans, lab reports and correspondence, when it runs inside the insurer's own environment so no policyholder data leaves the estate. Dilr Mira is a class of private clinical small language models from DILR.AI that turn documents like these into schema-valid JSON on the customer's own hardware. Data residency, not model novelty, is what decides it for a CISO.

The reason this matters on a value map is that document handling is the quiet cost in claims, where medical reports and correspondence are re-keyed by hand before anyone can assess a bodily-injury or health claim. Mira keeps that extraction inside the estate, which is exactly what matters when the file is a medical record, and DILR is researching the same schema-agnostic approach for wider insurance claims. On the underwriting side the same re-keying tax appears where commercial submissions arrive as unstructured emails and PDFs, and that document-intelligence work is a DATS delivery rather than a model one. The full mechanism, including how extraction accuracy is evidenced on medical files, sits in Mira's own material rather than here; the clinical document extraction guide is the place to go deeper. On this page Mira is one node in the map, the private-data corner of the claims operation.

When does a governed agent desk belong in claims operations?

A governed agent desk belongs in claims operations once several AI tasks run at once and a regulator will ask who approved what, the point where oversight outgrows a spreadsheet. Cognibl, from DILR.AI, is a governed work-management platform where people and AI agents share one board and no task reaches a done status without attached proof, on an append-only record attributed to the agent by name. For a claims operation it is a candidate control layer, not a point tool.

The case for it grows with the number of automated steps. One voice line and one extraction pipeline can be supervised informally; a claims operation running intake, triage, chasing and evidence generation together needs a single place that records every action, the evidence behind it and its outcome. That auditable record is what turns a set of AI tasks into something a board and the FCA can inspect, and it speaks directly to the review's finding on insufficient management information. You can see how the governed agent model works in practice on the Cognibl product page. On the insurance map, it is the node that appears once the others are already earning their keep, and it is what would keep an expanding set of agents inside the insurer's own governance rather than outside it.

Where each DILR line pays in UK insurance

Insurance is a sector where five of the six DILR lines have a genuine role and one does not, so the honest map names all six. The ordering below follows the value: Dilr Voice leads, DATS, Dilr Mira and Dilr Academy are secondary, Cognibl is a candidate once the operation scales, and DILR Studio does not apply. It also reads as a sequence rather than a shopping list: the order that works is to start with the voice contact problem, add document intelligence and conduct evidence through DATS, embed a private extraction model where the data is most sensitive, and only then formalise oversight and team-wide conduct literacy. That order matters, because placing the control layer before there is anything to control is how an AI programme stalls.

• Dilr Voice leads. FNOL and claims intake, policy servicing and renewals, claims status and proactive outbound updates, and broker query handling. It answers the claims-cost and complaint-delay problem the FCA named, and the voice AI enterprise pillar covers how the platform is built.
• DATS is secondary. The senior-led consulting system places AI for claims document intelligence, underwriting submission extraction and Consumer Duty evidence inside the insurer's perimeter, through the enterprise AI consulting guide rather than a platform sale.
• Dilr Mira is secondary. Private extraction from the medical documents inside bodily-injury and health-insurance claims, kept on the insurer's own hardware, as set out in the clinical document extraction guide.
• Dilr Academy is secondary. On-demand interactive courses for claims and underwriting teams, including conduct and Consumer Duty literacy, built the way the AI teacher buyers guide sets out for schools, universities and learning and development teams.
• Cognibl is a candidate. The governed agent desk that could run and record a claims-operations agent team once several tasks run at once.
• DILR Studio does not apply, with no self-serve use named in the insurance source materials for this industry.

What is the best way to place AI in a UK insurer in 2026?

The best way to place AI in a UK insurer in 2026 is to start where a public number and a regulator duty already point, prove it against that number, then expand, not by buying a platform and hunting for a use case. The clearest first places are claims contact and fraud triage, because both carry a measurable cost and a written FCA expectation. The best tool fits the operation and the data-residency posture, not the loudest demo.

On the voice side, buyers evaluating Dilr Voice will also look at PolyAI, Sierra and Decagon, and lighter builder platforms such as Vapi and Retell AI; a team that only wants a self-serve builder with no consulting layer may be better served by one of those. On the consulting and document side, DATS competes with Accenture, Deloitte, PwC and Faculty, and for a very large multi-year transformation a global systems integrator may win on scale and existing framework agreements. DILR's case is narrower and, for a mid-market UK insurer, sharper: senior people who diagnose where AI belongs in a four to six week engagement, place a governed system against a named number, then run it, with the DATS consulting system as the entry point rather than a licence. If you want to weigh the voice options yourself, the best AI voice agent guide for 2026 is the neutral starting point.

Want to see this in production? Try Dilr Voice for insurance, weigh the market with our best AI voice agent guide for 2026, book an AI placement diagnostic, or read about our approach to placing AI inside regulated claims operations.

The whole map holds together as a sequence rather than a menu, which is easiest to see as a flow.

Is UK insurance still a strong AI investment case in 2026?

Yes, because the cost pressure and the regulatory expectation are rising together, which favours AI that removes cost while producing evidence. A motor market set to only break even at 101% in 2025, and forecast into losses at 111% in 2026, cannot fund service improvement from premium. An insurer that places governed AI against a named claims or fraud number, and can evidence the outcome, has turned a margin problem into a place where AI clearly pays.

Does Consumer Duty bind the insurer or the broker?

Consumer Duty binds the party responsible for the outcome, which for a manufactured insurance product is the insurer. The FCA's board-reports review puts that duty on the insurer's board and senior management and looks for monitoring of outcomes across the distribution chain, and its claims-handling review treats oversight of an outsourced claims provider as the insurer's own responsibility. A broker can be separately in scope for its own role, but the manufacturer's outcome duty does not transfer.

Written by the Dilr.ai engineering team, practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.