AI is expected to take cost out of insurance, but the industry is still finding it hard to prove value

Across insurance, AI investment is accelerating – along with the pressure to justify it. The industry is reported to be¹ pouring millions into AI programmes and initiatives but getting “almost nothing back”. At the same time, the external signals around whether AI spend is delivering what was promised are becoming harder to ignore: Uber’s CTO² said in April that the company burned through its entire 2026 AI coding tools budget in just four months. Amazon³ recently disclosed examples of "catastrophically expensive" AI cost overruns, including one project that exceeded its budget by more than 800 per cent before the problem was detected months later.

For insurers, the timing matters. Costs are rising, capital is tighter, and a wave of recent acquisitions across the sector have all been underpinned by a mix of efficiency and delivering hundreds of millions in synergies. The message behind deals involving Aviva, Standard Life and Zurich, is remarkably similar; the search for efficiency, scale and stronger returns from existing assets. Yet despite the scale of the savings being targeted, many of the industry's largest transformation programmes still imply relatively modest workforce reductions, proving how difficult it remains to convert productivity gains into visible cost savings. Even Aviva’s integration of Direct Line, which is targeting more than £225m of annual cost synergies, is expected to reduce combined headcount by only around 5-7% over three years.

In that environment, every major investment is facing greater scrutiny, yet many teams are still working out where AI fits and how to prove its value in a way that stands up to board-level scrutiny across the insurance sector.
 

The problem isn’t capability; it’s proving where the value goes

This is not an industry standing still with AI. The FCA’s Mills⁴ Review points to a future where increasingly autonomous AI use cases are embedded across core functions, from servicing and claims to underwriting and compliance by 2030. The direction of travel is clear, the challenge is turning that activity into measurable value, particularly in large insurers already grappling with regulatory change, cost reduction programmes and complex integration work.

Much of today’s AI investment is aimed at operational efficiency, but efficiency doesn’t automatically translate into a saving. At the same time, whilst executives increasingly see AI as a route to improved risk selection, lower leakage, increased underwriter capacity and accelerated claims decisions, the harder task is organising the operating model in a way that allows those benefits to realise at scale.

The missing link is often capacity. Workers may be completing tasks faster, generating more insight or handling greater volumes of work, but unless that capacity is deliberately redirected, much of the benefit simply gets absorbed back into the system. Proving value therefore requires a broader scorecard, one that looks beyond cost take-out to outcomes such as underwriting performance, claims leakage, customer experience and productivity, and a clearer view of how released capacity can be redeployed to create value elsewhere.

Those that solve that challenge are likely to pull further ahead. Larger insurers are increasingly pursuing end-to-end transformation programmes, while smaller players are often focused on individual use cases and local productivity improvements. Over time, that gap could become a competitive fault line: some insurers will redesign how work gets done, while others simply make existing processes marginally more efficient.
 

The risk of AI costs rising faster than value in insurance

While value remains difficult to pin down, cost is becoming increasingly visible.

Complex insurance tasks are context heavy. Underwriters, claims handlers and brokers often need to compare policy wordings, submissions, claims histories and renewal information simultaneously which requires models to retain and process large amounts of context in a single interaction. The result is that cost is driven not just by usage, but by the complexity of the task itself. And there’s a trade-off: Larger models are often needed to reason across that volume of context but cost significantly more per query. Smaller models are cheaper but may not deliver usable outcomes for complex insurance tasks.

The larger the context window, the greater the computing requirement, and the more quickly token consumption starts to rise. Add agentic AI into the mix and consumption can escalate quickly; a single recursive agent loop or flawed logic can trigger hundreds of thousands of tokens in minutes. What seems affordable in a proof of concept can quickly become a material bottom-line issue at scale. According to one report, a major healthcare insurer⁵ watched its monthly AI token consumption go from three million to over one hundred fifty million in under a year.

There is, however, a reason for optimism. Techniques such as context engineering, memory layers and model routing are helping firms reduce unnecessary token consumption. Rather than repeatedly loading entire policy files or claims histories into a model, organisations are building architectures that retain relevant context and retrieve information only when needed. In turn, this is creating a fresh imperative to strengthen underlying data foundations. Retrieval is only as effective as the data beneath it, and insurers are increasingly using AI to improve data quality and make enterprise knowledge more accessible. Better foundations mean less information models need to process, improving both performance and cost efficiency.
 

What needs to change

The industry can see the both the benefits and costs of AI clearly enough. The harder task is making sure insurers are set up to capture and prove the value.

That starts with being far more deliberate about where value is expected to land. Point solutions can often demonstrate local productivity gains, but the bigger challenge is linking those gains to wider business outcomes. Capturing value requires aligning end-to-end journeys to specific financial or operational metrics, whether loss ratio improvement, claims leakage reduction or cost per policy, while understanding how released capacity will be redeployed elsewhere in the organisation.

It also means treating AI consumption with the same discipline applied to any other critical resource. What happens when your preferred model or API becomes materially more expensive? What if performance deteriorates, or access is interrupted entirely? These are exactly the questions insurers already ask about reinsurance capacity, IT vendors, BPO partners or claims supply chains, yet token consumption is rarely treated with the same rigour.

Finally, it means moving beyond the idea of a single model strategy towards a model portfolio. Not every process warrants the use of a frontier model, just as not every risk requires the same level of capital. That means balancing capability, cost and governance while deciding where models, data and inference should run, what evidence must be retained, and when smaller or domain-adapted models are a better fit than frontier alternatives.

Footnote:

1. The industry is reported to be pouring millions into AI programmes.
2. Uber’s CTO said in April that the company burned through its entire 2026 AI coding tools budget in just four months.
3. Amazon recently disclosed examples of "catastrophically expensive" AI cost overruns.
4.The FCA’s Mills Review points to a future where increasingly autonomous AI use cases are embedded across core functions, from servicing and claims to underwriting and compliance by 2030.
5. According to one report, a major healthcare insurer watched its monthly AI token consumption go from three million to over one hundred fifty million in under a year.