Ask any European insurance board about artificial intelligence today, and they will answer without hesitation that they are using it as the underwriting assistant, the claims triage pilot, the document extraction tool, the customer service copilot. The investment case was made years ago and won. By any reasonable measure, the industry has embraced AI with open arms.
Then ask what has changed in the business as a result, and the room gets quieter.
Insurers have spent the past two years racing to adopt AI. Across the industry, there is no shortage of ambition or investment. But as more AI programmes move from experimentation into production, insurers are discovering that the technology itself is rarely the hardest part.
The harder part is execution.
Many AI projects can perform well in a controlled environment. The challenge comes when they have to work with the systems, processes, data, and people that run an insurance business every day.
When something works in a controlled setting and struggles in a live one, the problem is rarely the thing itself, but rather the environment it was moved into. Insurers do not necessarily have an AI capability problem. They have a context problem.
What AI is up against
Insurance is an information business that has spent decades accumulating information systems. Growth through acquisition, new products, regulatory requirements, and years of individual technology decisions have left many large insurers with multiple policy administration estates, claims platforms, finance systems, and customer databases.
It is tempting to describe this simply as a data problem, but there is an important distinction.
Data being stored across different systems is an engineering challenge. The more difficult issue is that the meaning of that data can vary between systems.
A customer can be defined differently in different applications. A policy can carry different information depending on which system holds it. Even something as fundamental as the date of loss can be interpreted differently depending on the process or application involved.
Each definition may be perfectly valid within its own system. The problem arises when AI is expected to bring those different sources together and make a decision from them.
Fragmentation of storage is an engineering inconvenience. Fragmentation of meaning is what breaks AI. This helps explain why some AI programmes look successful during a pilot but struggle when they move into production. The model may work exactly as intended, but the environment around it does not provide the consistency and context it needs to deliver the same value at scale.
There is another challenge that is easy to overlook. The process on paper is not always the process in reality. AI tends to automate the process that has been documented, rather than the process that actually happens.
This matters particularly in insurance, where experienced employees have spent years adapting processes to deal with exceptions and unusual circumstances. A claims handler may make a referral that is not recorded in a formal workflow. An underwriter may apply judgement based on information that has never been captured in a system. A team may have developed a manual check because the documented process did not account for a particular situation.
These steps are part of how the business really operates, even if they do not appear in the process documentation. If AI automates only the documented process, it can miss the knowledge that makes the process work.
This can create a particularly frustrating outcome. An AI programme is delivered on time, the technology performs as expected, but the expected return never materialises. The problem is not that the AI failed. It is that the organisation automated the wrong version of the process.
Human intervention presents a related challenge. When employees do not understand why an AI system has produced a particular recommendation, they are unlikely to rely on it completely. They may check the output, adjust it, or override it.
That is not necessarily resistance to technology. In insurance, it can be responsible behaviour.
Underwriters and claims professionals are accountable for the decisions they make. If an AI-generated recommendation cannot be explained or traced back to the information behind it, the person making the final decision still has to exercise their own judgement.
But repeated human correction is also useful information. If people consistently override an AI system, insurers should ask why. The answer may reveal missing context, an incomplete process, or important knowledge that has never been captured elsewhere.
Why this is harder in insurance
Every industry carries legacy systems. Insurance carries a few aggravating conditions on top of them.
First, decisions have long tails. A motor claim may close in months, while a liability decision made today may be scrutinised years from now. That creates a standard of traceability that many other sectors simply do not have.
It means the value of AI in insurance is not determined by accuracy alone. It also depends on whether the people accountable for the outcome can understand and defend it.
Second, judgement is part of the product.
Underwriting is not data entry with a margin attached. It is a priced opinion about an uncertain future, formed by people whose expertise often consists of context that never made it into a system of record.
AI that cannot reach that context ends up competing with the expert instead of helping them make better decisions.
Third, the operating model is distributed.
Underwriting, claims, customer service, and operations may each have their own systems, processes and data. Each function can become highly efficient at managing its own work.
But the decisions made in one part of the business can depend on information held somewhere else.
A piece of information that matters to a claims team may change how an underwriter views a risk. A customer interaction may provide context that becomes relevant to a future claim. Information held in one part of the organisation can change the meaning of information held somewhere else.
This is where disconnected AI projects begin to reach their limits.
A useful way to think about it is a building. Each team can spend time making its own room more efficient, organising the shelves, improving the furniture, and automating individual tasks. But without a floor plan showing how the rooms connect, it is difficult to understand how the building works as a whole.
The same is true of an insurer. AI can make individual processes faster. But the bigger opportunity comes when AI can understand how those processes, people, information, and decisions connect. This requires connected enterprise knowledge but also requires trust.
Insurance decisions can have long consequences. Being able to explain where an answer came from, what information informed it, and why a particular recommendation was made is not a nice-to-have.
It is part of making AI usable in the long-term.
The asset that compounds
All this points to what is genuinely scarce. It is not models – those are improving quickly and becoming increasingly accessible to every competitor. It is not compute, and increasingly it is not talent either.
The scarce asset is connected enterprise knowledge: a clear understanding of what the organisation knows, what its terms mean, how its processes actually run, and how its decisions are made and governed.
Insurers have been buying capability, when in fact, what they needed to build was context. And context has a property which capability does not. It compounds.
A model deployed against a fragmented estate can deliver a one-off, department-sized gain and leave the next project starting from zero. Knowledge encoded once, including the definitions, the lineage, the decision rules, and the governance around them, makes the second use case easier than the first, and the fifth easier again.
The first project pays for a tool. The tenth pays for a capability.
It is also the part a competitor cannot simply buy. Anyone can procure the same model another insurer is using. Nobody can simply procure an organisation’s understanding of its own customers, products, processes, and decisions.
That is what makes connected enterprise knowledge a potential competitive advantage amongst businesses. Left where much of it currently sits, in spreadsheets, undocumented workarounds, and the heads of experienced employees, it is an asset quietly depreciating.
The practical work of changing that is not particularly glamorous. It means agreeing what an entity means across the enterprise and encoding it, rather than circulating a glossary. It means capturing the reality of processes by looking at how systems and people actually work, rather than relying only on how managers say they work. It means making lineage and reasoning visible by default, so that an output arrives with its provenance attached. And it means designing governance into the point of decision, rather than adding it afterwards as a review gate.
This is the work that often loses budget arguments to another AI pilot, and it is also the work that determines whether those pilots ever become valuable.
From experimentation to outcomes
For leaders trying to convert a portfolio of pilots into results, five questions matter more than another technology assessment.
The first is simple: what business decision does this change?
Not simply which task it accelerates, but which decision it improves, who makes that decision, and how the improvement will be measured. A use case that cannot answer that is a demonstration, not a business programme.
The second is whether it understands the process as it runs, or as it is written down. If the answer is the latter, the business case may rest on a document that stopped reflecting reality some time ago.
The third is what happens when people disagree with the AI. Human oversight is important, but repeated manual correction is also a signal. If employees are constantly overriding outputs, the organisation needs to understand why for smoother operations.
The fourth is whether it can show where the answer came from. For an underwriter, claims professional, or other decision-maker to trust AI, they need to understand the information and reasoning behind its recommendations.
The fifth is whether it will make the next use case easier. If every new application requires the organisation to rebuild its data connections, definitions, and governance from scratch, it is accumulating pilots rather than building a lasting capability.
These questions shift the conversation away from how many AI projects an insurer has launched and towards what those projects have actually changed. That is the measure that matters.
None of this is an argument for less ambition, but an argument for putting ambition in the right place.
The next stage of AI adoption in insurance will not simply be about deploying more models. It will be about giving those models the context they need to work within the reality of the enterprise and changing needs.
AI needs to understand the organisation’s data, but also what that data means. It needs to understand processes as they are actually carried out, not simply as they are documented. It needs to connect knowledge across functions rather than operate within individual silos. And it needs governance and traceability built into the way decisions are made.
For insurers, this means treating enterprise knowledge as part of the foundation for AI, rather than something that can be addressed after a model has been selected. The organisations that get this right will be better placed to move from individual experiments towards AI that can support decisions across the business.
The insurers that create the greatest value from AI will not necessarily be those with the largest number of pilots or the newest model. They will be the ones that do the slower, less exciting work of making their organisations understandable to AI. Because the question is no longer whether insurers will use AI. They clearly will. The question is whether their AI understands the business well enough to make a difference.