AI in insurance will be transparent, or it won't be

Let's say it upfront: we believe in AI for insurance, and not in small doses. Document analysis, preparation of policy administration tasks, portfolio steering: the gains are real, and it would be absurd to pass them up. But to deploy AI at scale in such a heavily regulated industry, you have to be able to answer for it. Insurers, and the SaaS platforms that embed AI, like to showcase figures on the accuracy and reliability of their models. That matters, but it is not enough. Another element is just as essential: transparency. The question is not only what AI can do, but whether you can clearly demonstrate how it operates and where it fits into your processes.

What regulators actually look at

No regulator asks for a model that is 99% accurate. What the EU AI Act and supervisors like the ACPR, France's insurance regulator, require comes down to a few questions: which AI systems do you use? Can you trace how a decision was made? Did the AI act autonomously, and if so, was it justified to let it act without supervision? Who is accountable? In practice, this means documenting your systems, logging what they do, and being able to explain any automated decision that affects a customer.

Performance is a commercial argument; AI compliance rests on the ability to document, trace and take responsibility. And insurance cannot afford to be approximate on this ground: pricing, underwriting, claims, the duty to advise. Every decision engages the customer's trust, and sometimes their financial protection.

Transparency happens around the model, not inside it

An AI agent cannot spell out its internal reasoning. It can justify its conclusions and cite the elements it relied on, but the billions of parameters in a model do not read like a decision tree. That is not a flaw to hide; it is a characteristic to design around.

What can be made transparent, and what regulators actually ask for, is the framework the agent operates in: the workflow is documented, every step is identified, the risk of each intervention is assessed, and every action is logged and linked to the file, the policy, the claim. You do not read the model's mind. You know exactly where it acts, on what, within which limits and under which controls. The AI is treated like a human operator, trusted to a greater or lesser degree depending on its track record and the level of risk involved.

That is what makes it possible to push automation far without losing control. Transparency is not what holds AI back. It is what earns AI the right to run at scale.

It all rests on the quality of context

There is one prerequisite, though: technology foundations that are actually AI-ready. A reliable AI workflow requires two things.

1. Very high-quality core data, with events captured at the right level of granularity, a complete history, and the ability to reconstruct the exact state of a file or a portfolio at any point in time.
2. Documented, digitalized business processes. If most decisions and operations happen outside the system, verbally, over email or across scattered tools, the AI will not have enough context to operate effectively.

This context is what allows an agent to produce accurate, verifiable analyses.

And this is where legacy weighs heavily: decades of aging systems, migrations and transferred portfolios have left data scattered, incomplete and rarely historized. Most decisions rely on implicit processes and manual rules. AI layered on top of these foundations fixes nothing; it amplifies. Questionable input data plus vague processes produce questionable recommendations, delivered with extra confidence.

Trust will be the real adoption driver

The future of AI in insurance will not be decided on performance alone. It will be decided on the ability to explain how these new AI-augmented processes work. The players who can document, trace and contextualize their AI use will be able to industrialize it at scale without eroding the trust of their customers, their distributors and their regulators.

That conviction shapes our work at Korint: event-based, granular, historized data as the foundation; AI workflows that are documented and risk-assessed; actions traced end to end. Insurance does not need a restrained AI. It needs an AI it can prove, explain and control, so it can use it to the fullest.