THE MAP HAS THE WRONG SCALE

Why insurers are navigating an exponential world with linear thinking

One European insurance group spent eight years and more than $500 million on a cross-country platforming programme. They never finished it. 

In contrast, another completed its claims platforming project. But 500% over budget. 

These are not outliers. Boston Consulting Group’s insurance practice documented them as representative of a wider pattern: programmes conceived for one technology environment, executed across years in which that environment transformed repeatedly, and delivered - if delivered at all - into a world the original brief could not have anticipated.

So let’s be clear. The people who commissioned those programmes were not making reckless decisions. They were applying a planning logic that had served the industry well for decades: identify the transformation needed, estimate the investment required, build a business case, execute over three to five years. 

That logic assumed the environment would be broadly stable across the programme timeline. Occasionally it would not be, and programmes would need to adjust. But the fundamental shape of the world - the technology available, the cost of capability, the competitive landscape - would be recognisable from beginning to end.

Yet now? That assumption is broken.

The problem with straight lines

In February 2019, OpenAI trained GPT-2 on 32 specialised chips running for 168 hours at a cost of approximately $43,000. Sam Altman’s ‘not for profit’ considered the results significant enough that they initially withheld the full model, citing safety concerns.

By July 2024, engineer Andrej Karpathy reproduced the same model in 24 hours on a single eight-chip node for $672. By January 2026, using his nanochat framework with improved algorithms and data quality, he matched GPT-2's performance in just over three hours for $73 - a 600-fold reduction in seven years. Karpathy's own estimate is that training costs are falling roughly 2.5 times per year. And he suspects that figure understates the real trajectory.

The training cost story is echoed in inference. Stanford's Human-Centered AI Institute found the cost of querying an AI model performing at GPT-3.5 level fell more than 280-fold in the eighteen months between November 2022 and October 2024 alone. Epoch AI's analysis across six performance benchmarks found inference costs declining anywhere between nine and 900 times per year, with a median of 50 times per year.

Human intuition is not built for this. We reason in straight lines. We can follow a compound interest table intellectually, but we do not feel it. 

Which means that, since the arrival of AI, any plan that extends three to five years forward is making assumptions about the cost and capability of AI that will be wrong, probably within twelve months, certainly within three years. Not directionally wrong. Categorically wrong.

The experimenter who cannot scale

Insurance has not ignored AI. The data suggests the opposite. BCG's 2025 study found 67% of insurers were testing generative AI programmes. McKinsey's most recent global survey found insurance respondents are now just as likely as those in the technology sector to report AI use - a dramatic shift from even two years ago. Ninety percent of insurance executives identify AI as a top strategic priority for 2025.

Yet only 7% have achieved scaled deployment.

That gap - between the 67% experimenting and the 7% scaling - is the clearest evidence that the industry is responding to an exponential environment with a linear playbook. Pilot programmes are scoped, assessed and handed to transformation offices. Transformation offices build business cases. Business cases go to investment committees. Investment committees apply the same three-to-five-year logic that has always served them. By the time approval comes, the pilot is obsolete.

McKinsey puts the competitive consequence in stark terms. The small cohort of AI-leading insurers created 6.1 times the total shareholder return of laggards over five years. That spread is wider than in most other sectors. The cost of waiting is not linear either.

The commodity clock

In 2016, Lemonade deployed an AI claims bot called AI Jim. It processed roughly 30% of claims without human intervention and settled one claim in a record three seconds. At the time, this was a genuine competitive differentiator. Industry analysts wrote about it for years.

By 2024, it was table stakes.

Geico now assesses vehicle damage with AI within seconds, replacing 30-minute manual reviews. Allstate automated nearly all claims-related emails across 50,000 daily communications. Aviva deployed more than 80 AI models, cut liability assessment times by 23 days, improved routing accuracy by 30%, reduced complaints by 65% and saved £60 million in a single year.

What Lemonade built as a differentiator, incumbents replicated in under a decade. The next cycle will be faster. The cost collapse documented by Stanford is not just an abstract number - it is the mechanism that drives commoditisation. When the cost of a capability collapses by two orders of magnitude, every competitor in the market gains access to it. Advantages built on access to expensive technology dissolve.

The efficiency story compounds this further. Stanford HAI found that the model size required to achieve equivalent benchmark performance shrank from 540 billion parameters to 3.8 billion - a 142-fold reduction. More intelligence, from a smaller, cheaper model. The frontier does not just get cheaper. It gets accessible.

The question is no longer whether to invest in AI. It is where on the commoditisation curve you are investing. A capability that feels cutting-edge today is a checkbox feature in three years. The organisations pulling ahead are the ones who have worked this out.

The data readiness myth

Ask almost any insurance CIO why AI programmes are not scaling and you will hear a variation of the same answer: the data is not ready.

It is worth understanding where this belief came from, because it is not entirely wrong. It is just expired.

In 2018, that advice was correct. Traditional machine learning models required clean, structured, labelled data.Building a model on messy legacy data produced messy results. The logical response was to clean the data first. Systems integrators and consulting firms built entire practices around this principle, and they charged accordingly.

The technology has since moved on. Modern AI systems are specifically designed to work with messy, unstructured data - to find signal in noise rather than requiring noise to be removed first. The call recording that cannot be transcribed by your legacy system is exactly the kind of input that today's AI handles natively. The unstructured complaint that your data warehouse cannot categorise is precisely where large language models add value.

Multiple industry surveys confirm the persistence of the myth. LIMRA and Deloitte both found 78% of insurers citing data quality and legacy systems as the primary barrier to AI adoption. McKinsey's European insurer survey found 60% describing their traditional data as "evolving" - the diplomatic version of "not ready."

The irony is that some of the richest, most actionable data in an insurance operation is the kind that was never going to sit in a clean relational database. Every broker submission full of handwritten notes and inconsistent formatting. Every claims call where the customer explains what actually happened in their own words. Every complaint letter that contains, buried in the frustration, a precise description of where the process failed. None of that belongs in a data warehouse. All of it is exactly what modern AI handles well.

The data readiness problem has not been solved. Perfect data has never existed in insurance-or any industry-and never will. What has changed is that imperfect data is no longer a roadblock. We can now process it effectively and, crucially, use AI to actively structure and improve that data as we go.

The expanding regulatory surface

Regulation is moving in the same direction.

The Central Bank of Ireland's revised Consumer Protection Code takes effect this month, introducing the "Securing Customers' Interests" standard - an absolute obligation that cannot be defended by showing reasonable measures were taken. Complaints must be resolved within 40 working days. Claimants must be notified of developments within 10 business days. The standard must be embedded in commercial objectives, culture, strategy, business model, and operations.

The EU AI Act, in force since August 2024, classifies AI systems used for risk assessment and pricing in life and health insurance as high-risk. Penalties for high-risk system violations reach €15 million or 3% of global annual turnover. EIOPA published its Opinion on AI Governance and Risk Management in August 2025, establishing supervisory expectations across fairness, ethics, data governance, transparency, and human oversight. DORA added ICT risk management and incident notification obligations from January 2025.

PwC and TheCityUK estimate annual regulatory compliance costs across UK financial services now exceed £33.9 billion - more than 13% of operating costs - and are rising. BIBA and London Economics found regulation costs amount to 5.2% of insurance premiums collected.

Most firms treat this as a cost of doing business. It does not have to be. Every unresolved customer interaction, every complaint that misses a deadline, every claim that takes longer than it should - these are now regulatory exposures as well as operational failures. The AI capabilities that reduce handling time and improve resolution rates shrink that exposure. The compliance cost and the AI investment case are the same conversation.

The intelligence opportunity

Most of the conversation about AI in insurance focuses on the customer-facing layer: chatbots, digital self-service, automated claims triage. Those capabilities matter. They are also well on their way to commodity.

The less-discussed opportunity is the intelligence layer - understanding, at an operational level, what is actually happening in customer interactions and why.

McKinsey's analysis of millions of contact centre interactions finds that 50-60% of customer enquiries remain transactional despite years of digital channel investment. Those interactions are happening on the phone. They are being handled by agents working from fragmented knowledge bases, searching for information across too many SharePoint sites, too many Teams channels, too many sources of truth. McKinsey found that 30-40% of call time in insurance claims handling is silent. Not talking. Searching.

This is not a technology problem specific to one firm. It is structural across the industry.

The NBER's landmark study found AI conversational assistance increased issues resolved per hour by 14% on average and by 34% for newer agents. MetLife saw first-call resolution improve 3.5% and customer satisfaction rise 13% after deploying real-time AI coaching across ten call centres.

These are not headline-grabbing numbers. They are compounding ones.

The firms that will own the next five years in insurance are not the ones that deployed the most AI tools. They are the ones that built the intelligence layer: systems that identify patterns in what customers cannot get answered, surface emerging problems before they become complaint spikes, and close the loop between customer interaction and operational knowledge.

Unlike a transformation programme, this kind of investment does not need three years to show results. It starts in one team, proves itself, and expands.

Recalibrating the map

BCG's 70% failure rate for digital transformations is not mainly a story about poor execution. It is a story about programmes conceived for one environment and delivered into another. Technology evolution outran the timeline. Given the pace at which things were moving even then, it was probably always going to.

Consider what that pace has done to expert forecasts. In November 2025, McKinsey revised its estimate of U.S. work hours automatable with current technology from 30% to 57% - nearly doubling its own projection in under two years. This is McKinsey: the firm that advises the boards commissioning those five-year programmes. If the world's most sophisticated strategy consultancy could not see two years forward, the assumption that a transformation mandate can reliably span five is difficult to defend.

What is different now is not just the speed but the character of change. Cost collapse, efficiency breakthroughs, benchmark records falling within months of being set, coding automation reaching a billion dollars in run-rate revenue in six months - these are not separate waves arriving in sequence. They have converged. There is one wall of water now, and it is already overhead.

A five-year transformation mandate written in 2024 was built on assumptions about what AI could and could not do that are already obsolete. That is not a failure of vision. It is what happens when a linear planning instrument meets an exponential environment.

The map is not wrong. It is drawn to the wrong scale.

Ninety percent of insurance executives say AI is a top priority. The question worth sitting with is not whether to invest. It is whether the planning logic being applied to that investment is calibrated for the world it will be executed in.

That recalibration is the work.

Brightbeam works with insurance organisations to implement AI that delivers impact quickly and compounds over time. If this piece raised questions worth exploring, we would be glad to talk.