Your experts are already building with AI. Give them somewhere safe to do it
In most insurers and financial services firms, AI is already being used for real analytical work. It is just not happening where anyone can see it.
An actuary builds a reserving check with an AI assistant on their own laptop. A pricing analyst uses one to rewrite a model in Python. A claims lead has a script that summarises loss files. The results travel by spreadsheet and slide deck. Nobody else can rerun them, and nobody can say which version of which tool produced which number.
This is not a failure of discipline. It is what happens when capable people find a tool that saves them hours and there is no approved place to use it. Deloitte's 2026 UK GenAI workforce survey found that 19% of GenAI users in financial services still pay for their own tool, even in a sector that provides more in-house tools than most.
The question for leaders is not whether to allow this work. It is already happening. The question is where it should live.
Why the usual responses don't work
There are two instinctive reactions, and both tend to make things worse.
The first is to lock it down. Block the tools, issue a policy, remind people of the rules. The work slows down for a while, then moves somewhere less visible. The experts who were getting the most value are the most frustrated, and they are usually the people you can least afford to lose.
The second is to turn it into a traditional IT project. The business writes requirements, a platform team builds something over several quarters, and the experts wait. By the time it arrives, the tools have moved on and the original need has changed. Anyone who has lived through a long data warehouse programme will recognise the pattern.
What both responses miss is that the experts are not the problem. They understand the models, the data and the judgement calls better than anyone. What they lack is a shared, governed environment that lets them keep working the way they already do, with the controls a regulated firm needs.
What regulators are now saying
For UK firms there is a further reason to act. In October 2026 the FCA published AI in financial services: good practice and areas for improvement. It does not create new requirements, but it is a clear signal of what supervisors will look for.
Several of the areas for improvement read like a description of desktop AI. The FCA describes firms that deployed AI tools before clear governance arrangements were in place, and calls out "retrospective governance". It flags overreliance on hard-to-replace AI providers and dependence on a single platform without an exit option. On the good-practice side it points to auditable records of prompts, outputs and model versions, and to piloting lower-risk use cases before full deployment.
None of that is achievable when the work sits on individual laptops.
What a safe place to build actually looks like
The answer is not a large platform programme. It is a deliberately small, governed sandbox that the business can work in from the first weeks. In our experience it needs five things.
- One shared code base. Models, checks and prompts live in version control rather than in personal folders, so work can be reviewed, rerun and handed over.
- Governed access to the right data. Experts work against the same curated data the rest of the firm relies on, not extracts they have saved locally. That is usually where a modest amount of data engineering pays for itself.
- An audit trail by default. Every model change, every AI-assisted step and every result is recorded, so a number in a board pack can be traced back to how it was produced.
- No single point of failure. The work should carry on if one AI provider changes its terms, retires a model or goes down. That means designing for a fallback, not assuming one vendor forever.
- A named owner for every result. AI can draft and calculate, but an expert signs off. The environment should make that ownership explicit rather than implied.
Cloud platforms such as Azure, Databricks and Microsoft Fabric already provide most of the building blocks for this. The hard part is rarely the technology. It is shaping it around how a reserving, pricing or claims team actually works.
Start with one model, not a platform
The fastest route we have seen is to pick one non-production model in one line of business and move it into the sandbox. Something that matters enough to be worth doing well, but where a mistake does not reach a customer or a regulatory return.
Engineers sit with the experts for a few weeks, rather than collecting requirements from a distance. They rebuild the work in the shared environment, wire it to governed data, add the controls and make it repeatable. The experts keep ownership of the logic throughout.
At the end you have something working, a clear view of what it took, and evidence for the next decision: which models move next, what needs to harden before anything gets near production, and where the controls need strengthening. That is a far better basis for investment than a requirements document.
It also changes the conversation with risk and compliance. Instead of being asked to approve something abstract, they can look at a real model, a real audit trail and a real fallback, and challenge them.
What to do this quarter
If you lead an actuarial, pricing, finance or claims function, three questions are worth asking now. Where is AI already being used in your team's analytical work, and on whose machine? Could you reproduce last month's AI-assisted result if the person who built it left tomorrow? And if your preferred AI provider were unavailable for a week, what would stop?
If the honest answers are uncomfortable, that is not a reason to clamp down. It is a reason to give your experts somewhere better to work.
How we approach it
This is the kind of work our forward deployed engineers do. They embed with business experts, build in governed cloud environments on Azure, Databricks or Microsoft Fabric, and put controls in from the first day rather than after the fact. Our data engineering pods then make sure the data underneath is something people can trust. The aim is always the same: capability your team owns and keeps using after we step back.