The AI Reality Check

Mostly to read, extract and summarise, so that people reach their decisions faster. In underwriting, AI reads broker submissions, schedules of values and loss runs, pre-fills the risk record and triages what arrives against appetite; Zurich cut manual triage time from 75 minutes to 15 per commercial submission, in figures released with its supplier. In claims, it takes first notification of loss, routes claims to the right handler, summarises calls and files and flags fraud; Aviva runs more than 80 models across its claims journey. In service, it answers routine questions and drafts documents, with the clearest results where a route to a person stays open.

Pricing is where AI has been in production longest, particularly in personal lines, while AI that makes the underwriting or claims decision itself is still rare outside simple, standardised products. All six case studies in the eBook kept a person at the decision and measured the gain in the time it took that person to reach it. For a commercial insurer or MGA, the practical order is AI that prepares the work, then AI that informs the decision and only then AI that makes it. Each step depends on the data the firm holds, which is why the core policy administration system matters as much as the model.

It depends on who is counted. In the UK, 95% of the insurers in the Bank of England and FCA's 2024 survey used AI, the highest rate of any sector, though only 19 of the 118 firms that responded were insurers and most of them were large. The survey has not been repeated, so it is still the most recent official measure. In South Africa, the FSCA and Prudential Authority found 8% of insurers using AI in 2025, against 52% of banks, from a base that includes thousands of small financial services providers.

Using AI is not the same as scaling it. BCG found that 7% of insurers had scaled AI and about two thirds were still piloting. EIOPA's survey of 347 European insurers found nearly two thirds using generative AI, most of them at proof of concept. Return is where the evidence is thinnest: 11% of insurance leaders in KPMG's 2026 survey described their AI return as very clear, from a sample of 44. At the top of the market, Evident found 20 of the 30 largest insurers in North America and Europe reporting at least one AI use case with measured outcomes. The eBook does not use the widely repeated claim that 95% of organisations get zero return from generative AI, and explains why.

No regulator in either market has written AI-specific rules for insurers. The FCA has said it does not plan to. Instead, AI use is governed by the existing regime: the Consumer Duty, which requires firms to deliver and evidence good customer outcomes, the Senior Managers and Certification Regime, which places accountability for a system's behaviour with a named individual. Data protection law applies too. Since 5 February 2026, Articles 22A to 22D of the UK GDPR allow significant decisions made solely by automated means on a wider range of lawful bases, provided the firm tells the individual, offers human intervention and lets the decision be contested.

In South Africa, the FSCA is developing an AI framework but had issued no guidance by October 2026. The binding constraint is section 71 of POPIA, which restricts decisions with legal consequences based solely on automated processing. The EU AI Act treats AI used to price life and health cover for individuals as high-risk, with obligations now due from 2 December 2027, while commercial property and casualty lines are outside that category. Every one of these duties is met from records: which model produced an output, what data it used and who approved it. A firm whose core systems keep that record, as claims management software with a full audit trail does, can defend each new use of AI.

Agentic AI carries out a sequence of steps towards a goal using tools and data, rather than answering a single prompt. On a claim, an agent might read the notification, check the policy wording, pull weather data, query a fraud model, calculate a settlement and draft a decision for a person to approve. The difference from earlier automation is that the system decides the order of steps and which tools to use, which makes it more flexible and harder to test. Our glossary entry on agentic AI explains the term in more detail.

The published examples in insurance are few and deliberately narrow. Allianz's Project Nemo uses seven agents on food spoilage claims under AUD 500, cut processing and settlement time by 80% and leaves every payout with a person. Evident found that 8% of the AI use cases disclosed by the largest insurers pointed towards agentic reasoning, and Gartner expects more than 40% of agentic projects to be cancelled by the end of 2027. The deployments that work share four traits: high volume and low value per item, well-defined steps, every action logged and a person approving anything that commits money or cover. Submission intake, document chasing, bordereaux validation and renewal preparation fit that profile. A complex coverage decision or a large-loss settlement does not.

Pick one high-volume process, measure it, make sure its data is captured in structured form at source and use AI that comes inside platforms with open APIs and a full audit trail. The programmes with published results each started with one process and one number, such as triage minutes per submission at Zurich or liability assessment days at Aviva. The best first candidates in commercial lines are document-heavy tasks where a person still makes the decision: submission intake, bordereaux validation, first notification of loss, document generation and broker servicing queries. Each has a clear measure, a low regulatory profile and a payback of two years or less in the eBook's rating.

A firm writing £20 million to £250 million of premium will not employ a data science team, and does not need to. Most of its AI will arrive inside the policy administration, claims and document systems it buys, so the selection questions move from features to architecture: whether the platform exposes its data through APIs, records which model made which decision and lets the business change the rules around a model through configuration. Governance can be light if it covers an inventory of AI systems, a named owner for each, bias testing, outcome tracking and supplier questions. For MGAs, our page on MGA software shows how Genasys connects the AI services a firm chooses to its data.