There is currently a lot of enthusiasm in the insurance industry about AI, and that enthusiasm is largely justified. The technology is capable and evolving rapidly, with significant opportunities for its application in underwriting, distribution, customer service and claims.
While all this is true, there is a gap between what AI can do and what many insurers will actually get from it and closing that gap is not primarily an AI problem but one rooted in the foundation of the underlying technology.
The layer problem
Every AI agent making an underwriting decision draws on product logic, appetite rules, and underwriting guidelines that exist within the system the business relies on. In fast-moving markets, rules can evolve over time. If they are not consistently updated and controlled, the agent may be drawing on information that no longer fully reflects the business’s current position.
And, as it is designed to do, AI will apply that information consistently and at scale. If the underlying rules are not current or well controlled, that scale can amplify outcomes that do not fully reflect the business’s intent. This is not a limitation of AI itself, but a reflection of the importance of having accurate, business-controlled foundations in place.
The insurance industry has heavily invested in adding front-end capabilities, including submission ingestion, automated quoting, and agent workflows. These are real advances, but in many cases, they are writing decisions to policy administration systems that have not kept pace with the rate of development.
What this looks like in practice
A policy would be written under one set of underwriting rules: the risk is calculated, the quote is generated, and the policy is finalised. Months later, the customer makes a mid-term adjustment, but the appetite and market conditions have changed in the intervening period, and the system has not caught up. The adjustment will then be processed against different rules than the original policy. And then at renewal, a third set of rules will be in play. When the fact that these processes have fallen out of step surfaces in a claim or a regulatory review, the explanation that an AI agent handled the decision will not carry much weight.
What customers really need
From the outset, INSTANDA was designed to give insurers full control, which puts it in a strong position as AI becomes table stakes. But rather than assume how AI should be applied, we worked closely with our clients to understand where it could add the most value.
We discovered that the biggest demand was not for automation of high-volume, predictable decisions, as deterministic logic already handles those actions well, and in many cases more cheaply than AI would. What customers wanted was support for the people who come back to the system periodically, such as brokers handling renewals, operations leads making mid-term adjustments, and product managers revisiting rating factors.
In these moments, decisions are more complex, contextual information is important, and the risk of, error is higher. AI can add significant value in this area, but only if the system those people are working in reflects the rules the business is operating on today.
This demand shaped the products we built for our clients, with each one assuming a current, business-controlled foundation, not dependent on IT queues or months behind the decisions already being made in the market.
The governance question
There is a version of the AI conversation that focuses almost entirely on the performance of the model: accuracy rates, processing speed, and the volume of decisions handled per day. These are not unimportant, but they are limiting indicators for the most pressing questions in AI.
The question that matters more is whether the business can stand behind every decision made by the AI; whether the rules it acted on were the rules the business intended, and whether, if a decision is challenged, there is a clear, explainable answer.
When the foundation is right, these questions can be answered, but when it is not, the AI layer creates a problem that is more opaque and more difficult to manage than the one it was meant to solve.
A practical starting point
Insurers do not need to resolve every technology challenge before they can use AI well, but they do need to know what their AI is acting on, and they need to be confident that those foundations are under the control of the business.
That means understanding where product logic lives and who can change it, knowing whether underwriting rules are consistent across the policy lifecycle, and having governance in place before decisions are being made at scale.
The insurers that benefit most from AI will not necessarily be those that deployed it earliest. They will be the ones who made sure it had something solid to work on.
