In insurance, caution around AI often rests on a single argument: the regulatory framework is supposedly too unclear to get started. It is a comfortable position, but it does not hold up. The principles that have governed our industry for decades, controls proportionate to risk, non-discrimination, documented decisions, corporate accountability, apply perfectly well to AI. What is missing today is not another piece of legislation. It is the will to translate these principles into practice and get moving.
The model that works: automate 70, arbitrate 30
Automated decision-making is nothing new in insurance: premiums have been calculated without human intervention and business rules applied automatically for years. So the real question is not whether a system can decide on its own, but where to set the dial. The model that works is 70-30: heavy automation on standard cases, where the productivity gains are massive, and a deliberate handover to humans on complex cases, where their judgment adds real value.
To move fast, start with the simplest tasks and the most repetitive administrative operations. That is where gains come quickly, where teams learn, and where trust is built.
A real difficulty, faced without drama
Generative AI models lack the determinism of fixed-rule systems: the same operation can produce slightly different results from one run to the next, and some will need rework. For information systems historically built on reproducibility, that is a significant shift in habits, and it would be dishonest to downplay it.
Yet this shift is more familiar than it looks. Our organizations have always managed a form of variability: that of human teams, who also make mistakes, differently depending on the day and the case. We learned to contain it through control testing, supervision, and error correction. That is exactly the approach to transpose to AI, rather than demanding from it a perfection we never demanded from humans.
The skill that will make the difference
The key skill of the coming years will not be knowing how to fine-tune a model. It will be knowing how to identify which workflows can be fully entrusted to AI and which require human judgment. That is business expertise: it takes deep knowledge of your processes, your risks, and your customers. The organizations that develop this discernment will build a considerable lead over those that automate at random, or automate nothing at all.
Let’s stop hiding behind regulation
The obstacles are real and numerous: data quality, change management, governance of edge cases. And precisely because they are real, they need to be tackled now, rather than waiting for a hypothetical perfect framework that will never arrive.
The legal framework exists, the principles are known, and the first use cases can be identified today. Regulation is not what is stopping insurers from deploying AI. It is sometimes just what they find convenient to believe.
This is exactly the ground on which Korint built its platform: AI agents embedded in business workflows rather than bolted on beside them, able to fully automate standard operations while handing control back to a claims handler the moment a case falls outside the frame, all running on event-based data that traces every action and makes it possible to document, at any time, who did what and on what basis. In other words, the conditions to apply to AI the safeguards insurance already knows, and to move forward without waiting.