AI for Insurance: A Practical Guide for UK Insurers and Brokers

Insurance has quietly become one of the most AI-active corners of UK financial services. The Bank of England and FCA's 2024 survey of UK financial services firms found the insurance sector reporting the highest AI adoption of any sector surveyed, at 95 per cent, against 75 per cent across financial services as a whole. So the question for most insurers and brokers is no longer whether to use AI, but where it genuinely pays off, where it creates regulatory and reputational risk, and how to deploy it without falling foul of the Consumer Duty or the Equality Act.

This guide is written for decision-makers in UK insurers, brokers and MGAs: people who own claims, underwriting, pricing, distribution or compliance, not the engineers who build the systems. It covers where AI earns its keep, where it should stop, the rules that apply in this sector and how to start sensibly. For the broader technical grounding, our pillar on Claude AI agents for business is a good companion read.

Where AI earns its keep in insurance

The strongest use cases share a pattern: AI does the heavy reading, sorting and drafting, while a person keeps the judgement on anything that affects a customer outcome. The areas where insurers and brokers are seeing real value include:

• Claims triage and straight-through processing. AI classifies incoming claims at first notification of loss, routes them by complexity, auto-settles simple low-value claims and flags complex ones for adjusters. The business value is faster settlement, lower handling cost per claim and freed adjuster capacity for the cases that genuinely need human judgement.
• Fraud detection at claim and application stage. Machine-learning models score claims against historical fraud patterns, network analysis spots organised rings such as crash-for-cash, and newer tools detect AI-generated or manipulated documents. The prize is large: Aviva reported blocking more than 6,000 fraudulent claims and preventing over £60 million in losses in the first half of 2025, and the Association of British Insurers put total detected UK insurance fraud at £1.16 billion in 2024. Stopping fraud ultimately reduces pressure on premiums.
• Underwriting knowledge retrieval. A generative-AI assistant can ingest lengthy underwriting guides and answer underwriters' questions in seconds rather than them reading hundreds of pages. Allianz UK reported that its underwriting assistant, BRIAN, handled around 13,000 queries and saved an estimated 65,000 minutes of information-gathering, roughly 135 working days, in its first year after a January 2025 rollout. The human underwriter keeps the decision.
• Pricing and risk-model support. AI can enrich and clean rating data, surface risk signals and support actuarial and pricing models, sharpening risk selection and speeding repricing. It is decision support rather than autonomous pricing, because the final pricing logic has to be explainable and demonstrably fair.
• Customer service automation. Chatbots and voice assistants handle routine policy questions, mid-term adjustments, renewals and document requests, with handover to a person for anything material. This gives customers 24/7 self-service and shorter waits while keeping advised and regulated interactions human-led.
• Document processing and data extraction. AI reads and structures policy documents, schedules, medical reports, surveys, broker submissions and emails, combining OCR with large language models. It cuts manual rekeying, speeds new-business and claims intake, and improves the data feeding downstream systems.
• Broker submission and quote handling. For commercial brokers and MGAs, AI parses inbound submissions, extracts the risk data, matches it to appetite and pre-populates quote requests, giving faster turnaround and better triage of which risks to pursue.
• Communications and complaints summarisation. AI drafts and standardises customer letters and summarises long case files or call transcripts, improving consistency and audit trails, with a person checking anything customer-affecting.

Where AI stops, and why that matters in insurance

The honest limitations matter more in insurance than in most sectors, because the regulator expects you to be able to explain and defend customer-affecting decisions.

• The explainability gap. Many high-performing and generative models are hard to fully interpret. The Bank of England and FCA 2024 survey found only 34 per cent of firms claimed complete understanding of the AI they used, with 46 per cent reporting only partial understanding. That is a serious issue when the FCA expects firms to explain decisions that affect customers.
• Hallucination and accuracy risk. Large language models can produce confident but wrong answers, fabricate references or misread documents. In underwriting guidance, claims decisions or customer communications that can cause real harm, so outputs need verification and the technology is best treated as assistance, not autonomous decision-making.
• Bias and proxy discrimination. Models trained on historical or geographic data can encode unfair patterns even without using protected characteristics directly. Citizens Advice estimated in 2024 that people of colour may pay around £250 more per year for car insurance than White customers, the so-called ethnicity penalty, where postcode and area data act as proxies. Detecting and stripping that bias is difficult and is an active area of regulatory scrutiny.
• Fraudsters use AI too. Aviva has reported a rise in AI-generated fake documents and images supporting fraudulent claims, so detection models have to keep pace with increasingly sophisticated, AI-assisted fraud.
• Data quality and legacy systems. Insurers and brokers often hold fragmented data across legacy policy and claims systems. AI value is capped by messy, siloed or poorly governed data, and getting models reliably integrated into core workflows is much harder than running a pilot.

The rules: insurance-specific obligations

This is where insurance AI projects most often come unstuck. The UK has chosen not to write a bespoke AI rulebook, so existing frameworks do the work. You need to understand which ones apply.

The FCA's approach is technology-neutral. The FCA has deliberately decided not to introduce AI-specific rules. Its approach, set out in "AI and the FCA: our approach" published on 9 September 2025, is principles-based and outcomes-focused, relying on existing frameworks rather than a separate AI regime. That means the standard you are held to is the standard you already know.

The Consumer Duty is the central lever. Where AI touches retail customers, the Consumer Duty requires firms to deliver good outcomes and avoid foreseeable harm. AI used in pricing, underwriting, claims or service has to be shown not to produce poor or unfair outcomes, and you must be able to evidence that. Closely tied to this are the Fair Value rules and the General Insurance Pricing Practices reforms, in force since January 2022, which include the ban on price walking at renewal. The FCA is evaluating those remedies in its Evaluation Paper 25/2 and has signalled it will act where pricing or AI models create unfair differential outcomes.

The Equality Act 2010 and proxy discrimination. The Act prohibits discrimination on protected characteristics. In insurance the live concern is indirect discrimination through proxy variables, the ethnicity penalty raised by Citizens Advice, where area data can lead to higher prices for communities of colour even though insurers do not collect ethnicity. The FCA has indicated it will scrutinise and can take enforcement action where biased data drives unfair pricing.

Accountability sits with a named person under SM&CR. Under the Senior Managers and Certification Regime, accountability for AI decisions cannot be outsourced to a model. A named senior manager remains accountable, which is why governance and oversight of AI have to sit with identifiable individuals.

Human oversight of regulated decisions. Solely automated decisions with legal or similarly significant effects, such as declining a claim or refusing and pricing cover, are restricted under UK GDPR. Following the Data (Use and Access) Act 2025, the relevant provisions moved to new Articles 22A to 22D, a more permissive but safeguard-led regime that requires genuine, active human involvement rather than a token rubber-stamp, plus transparency, the ability to contest a decision and the right to obtain human intervention.

UK GDPR and the ICO. Lawful basis, fairness, transparency and data minimisation all apply. The ICO's guidance on AI and data protection, and its automated decision-making and profiling guidance updated for the 2025 Act with a 2026 consultation on draft ADM guidance, set expectations on bias mitigation, data protection impact assessments, record-keeping and explainability to data subjects. Firms must also give customers meaningful information about how AI affects them, for example a premium decision; the Bank of England and FCA survey flagged transparency and explainability as a top adoption constraint.

Model risk and professional standards. For relevant firms, the PRA's Supervisory Statement SS1/23 on model risk management applies and is being stretched to cover generative and agentic AI, and the Bank of England frames widespread AI as a financial-stability consideration. Alongside the regulators, the Chartered Insurance Institute's Code of Ethics expects insurance professionals to take personal accountability for outcomes created by AI; the CII argued to the Treasury Select Committee in March 2025 that institutions must remain accountable for AI decision-making.

One useful route for higher-risk deployments: rather than new rules, the FCA created its AI Lab in October 2024, a Supercharged Sandbox, and AI Live Testing, with live cohorts from October 2025 and a second cohort in April 2026, so firms can develop and test AI alongside the regulator.

Off-the-shelf AI or a custom agent?

Off-the-shelf tools, such as vendor fraud-scoring platforms, document-extraction services, embedded software features and general LLM assistants, are quick to adopt and good for common, lower-risk tasks like document processing, first-line customer service and standard fraud signals. Their limit in UK insurance is control and evidence: you often cannot fully see or explain the model, tune it to your Fair Value and Consumer Duty obligations, keep customer data where you need it, or produce the audit trail the FCA and ICO expect.

A custom agent, built on your own data with your guardrails, human-in-the-loop checkpoints, logging and explainability baked in, makes more sense for anything touching regulated decisions, such as pricing support, underwriting referral or claims triage and decisioning, where accountability under SM&CR and the Consumer Duty rests with the firm. The sensible pattern for most insurers and brokers is to buy for commodity tasks and build, or heavily configure, for the regulated, customer-affecting core, always keeping a person accountable for the final decision. We explore the same trade-off across the wider industry in our guide to AI for financial services.

How to start

The firms that get value from AI tend to start narrow and govern hard, rather than launching a broad programme. A workable sequence:

• Pick one task with a clear owner. Choose something with high volume and contained risk, such as document extraction at new-business intake or summarising claims case files, rather than starting with pricing.
• Set the governance before the build. Name the accountable senior manager, decide where the human checkpoint sits, and run a data protection impact assessment if personal data is involved.
• Pilot on real cases in a controlled way. Run the AI alongside existing process so you can compare outputs and catch errors before anything reaches a customer.
• Measure outcomes, not just speed. Track accuracy, customer impact and fairness, not only handling time, so you can evidence good outcomes to the regulator.
• Keep a meaningful human in the loop. For any decision with a significant effect on a customer, ensure the human involvement is genuine and documented, not a rubber-stamp.

What it costs

SpotDev works to fixed packages so cost is predictable from the outset. Most engagements begin with an AI and Data Readiness Assessment at £5,000, which establishes where AI can safely add value across your claims, underwriting, pricing or service operations and what your data and governance gaps are. Delivery of a custom Claude agent or rollout then runs from £8,000 to £45,000 depending on scope, with a first rollout typically live in two to three weeks. We have delivered 300+ technology projects, our engineers are in-house and nothing is subcontracted. If you want to scope a specific use case, talk to a Claude-specialist engineer.

Frequently asked questions

Does the FCA have specific rules for AI in insurance?

No. The FCA has deliberately chosen not to introduce AI-specific rules. Its approach, set out in "AI and the FCA: our approach" in September 2025, is technology-neutral and outcomes-focused, so AI is held to existing frameworks such as the Consumer Duty, the Fair Value rules and SM&CR rather than a separate AI rulebook.

Can we use AI to decline claims or set prices automatically?

Solely automated decisions with a legal or similarly significant effect, such as declining a claim or pricing cover, are restricted under UK GDPR. Following the Data (Use and Access) Act 2025, Articles 22A to 22D require genuine, active human involvement, transparency, and the right for a customer to contest the decision and obtain human intervention. In practice this means AI supports the decision and a person remains accountable for it.

How do we avoid AI creating unfair or discriminatory pricing?

The risk is indirect discrimination through proxy variables, such as area data standing in for ethnicity, which Citizens Advice highlighted as the ethnicity penalty. You need bias testing, clear documentation of what data drives pricing, and the ability to explain outcomes, because the FCA has indicated it can take enforcement action where biased data drives unfair pricing under the Equality Act and the Fair Value rules.

Should we buy an off-the-shelf tool or build a custom agent?

Buy for commodity, lower-risk tasks like document processing and first-line customer service. Build or heavily configure a custom agent for regulated, customer-affecting work such as pricing support, underwriting referral and claims decisioning, where you need to explain the model, meet Fair Value and Consumer Duty obligations, control your data and produce an audit trail, and where accountability under SM&CR rests with your firm.

Work with a Claude specialist

SpotDev designs, builds and deploys custom Claude agents and enterprise Claude rollouts for UK businesses, with fixed packages from £8,000 to £45,000 and a first rollout live in two to three weeks. Explore our Claude implementation packages or talk to one of our engineers.