Fix the Foundation Before the Agents
By Paul McLeod, Founder, Bespoke Analytics
In June, Insurtech Insights USA closed its New York event with a message from the main stage of an insurtech conference.
The organisers' own summary called on the industry to fix its data foundation as it integrates AI. One line from the coverage stuck with me: AI is only as powerful as the data foundation it sits on.
I've been saying a version of that to reinsurance clients in Bermuda and Cayman for a long time, so it's good to hear it at that volume. But reading through how the programme was organised, there's a gap. It happens to sit in the part of the business where a reinsurer's data foundation actually gets tested, and I think that part Bermuda reinsurers will have to finish themselves.
Three things the conference got right
There are three points from the coverage I'd underline.
The first is the order of operations. Fix the data, then deploy the agents. One point from the stage put it plainly: the problem is how the systems were built, not the AI. Anyone who has connected to an AS400 from the stone tablet days knows exactly what that means.
The second came from a session called "Data Foundations to Decision Power: Building a Single Source of Truth", with speakers from Swiss Re and AXA XL. Their advice was to start AI with moderate and lower-risk use cases, build a delivery track record, and only then move to financial and compliance applications. That's sensible sequencing, and I'll come back to it.
The third was the human in the loop. Panel after panel said AI can be overconfident in its output, and that in insurance the downside is not capped at the premium. A person reviews the answer, and the answer shows its working.
What the agenda left out
The published tracks included underwriting, claims, distribution, life and health, and specialty lines. The track overview did not list dedicated tracks for finance, actuarial or regulatory reporting, although investment was included.
That matters for a reinsurer. Underwriting is where AI gets the headlines. Greenlight Re, for example, talked about pointing AI at submission ingestion so underwriters can quote or decline faster. Good.
But a single source of truth isn't tested at intake. It's tested at quarter-end, when actuarial, underwriting, claims and finance all have to aggregate up into one set of numbers that the board signs off and the BSCR and EBS work depends on. If the foundation is going to break, that's where it breaks.
One measure, three answers
Take net written premium. This is a hypothetical, but you'll recognise it.
The underwriting system records it at bind. Finance carries it after cancellations, endorsements and FX. Actuarial looks at it by underwriting year. All three numbers are right in their own system. None of them is wrong until someone asks one question across all three.
Now put an agent on top. A CFO asks for net written premium for the quarter. She meant the Bermuda entity only. The agent included the sidecar as well, used the bind-date figure, and gave the answer in a clean sentence with no hint of doubt.
As an executive, I can get quite excited about an agent that answers in seconds instead of weeks. But I could also get nervous. How do I know that number is the one we'll file? How do I defend it when an auditor asks where it came from?
That's the reconciliation problem this industry has always had. Agents don't create it. They make it faster and more confident.
What a defensible foundation looks like in Fabric
Microsoft Fabric is where most of our reinsurance conversations land at the moment, and it suits this problem well. But a feature only matters if it changes what the CFO can say hand on heart. So here is each one, translated.
One copy of the data in OneLake. Actuarial and finance stop arguing about whose extract is right, because there is only one.
A certified semantic model. Net written premium is defined once, signed off by a named owner, and every report and every agent uses that definition.
Lineage from source to filing. When the auditor asks where a figure came from, you can walk them back from the board pack to the bordereau line and the bound contract system.
Access set by role, in the data. The AI answers with what the person asking is allowed to see, and nothing underneath it.
Capacity sized for quarter-end. Fabric capacity bought for an average Tuesday will struggle in the last week of the quarter. Size it for the close.
The false starts I see are usually the same few. Capacity bought before anyone knows the workloads. Every legacy report lifted across unchanged. Workspaces created per project until nobody can say which one holds the real numbers. And no named owner for the definitions, which means one semantic model quietly becomes three.
I'll speak as a partner here. On the projects we've run, TimeXtender has done an awful lot of the heavy lifting. The build is metadata-driven, so the documentation and lineage come out of the build rather than being written up after the fact. Your team gets to focus on the business problem instead of the plumbing.
Sequencing for a reinsurer
So how does the Swiss Re and AXA XL advice apply in Bermuda?
Start the AI with lower-risk work. Summarising submissions for underwriting. Letting the finance team ask questions of certified measures instead of waiting on an analyst. Both are useful, both have a person reviewing the output, and neither goes near a filing.
But build them on the same governed layer your reporting uses. The filing isn't your first AI use case. It's the first test your foundation has to pass. Get the definitions, the lineage and the access right for quarter-end, and every AI use case after that inherits them. Build the AI on a separate, quicker stack, and you've created a second version of the truth that someone will have to reconcile later.
I find this an exciting time. What used to take a year and an army of specialists to stand up can now be in place in weeks. That speed is exactly why the order matters.
Five questions boards are being asked to evidence
Boards in Bermuda and Cayman are being asked to show safe value from AI without creating an audit problem. These are the questions I hear most often. They aren't a regulatory checklist, and you should confirm with your own counsel or regulator what you specifically need to evidence.
Which definition of the number did the AI use, and who signed that definition off?
Can you reproduce last quarter's figure from source today?
Can you trace a figure in the board pack back to the system it came from?
Did the AI answer only from data the person asking is allowed to see?
When the AI gets it wrong, how would you know, and who reviews it?
If you can answer all five, you're in good shape. If you can't answer the first two, the agents can wait. We wrote about the first question in more detail in Show Your Working: AI Governance Starts With One Number.
Where to start
The industry has agreed on the order: foundation first, agents second. For a reinsurer, the foundation is the quarter-end, and that part isn't on anyone's conference agenda.
That's what our two-week Fabric Readiness and Roadmap Sprint is for.
We look at your source systems, the measures that matter at the close, your Fabric capacity and workspace design, and where AI can safely start.
You leave with a sequenced roadmap your team, your board and your auditors can follow.
