A powerful engine is only worth what reaches the road. Put it in a car on worn tyres and all you get is noise and wheelspin.
I think about that whenever I hear AI discussed, because nearly every conversation is about the engine. Which model is ahead this month. What the latest agents can do. How fast it is all moving. Those are fair things to be interested in. I’m just not convinced they are where the advantage lies.
The engine is no longer the differentiator
Start with an uncomfortable fact for anyone hoping to win on the model alone: The gap between the leading models is closing. Stanford’s 2026 AI Index reports that performance at the top is converging, with the leading models from four different companies now clustered within a narrow band of one another[1].
Models are not identical. They have different strengths, and the one in front changes regularly. But those differences are small, they don’t last, and above all they are for sale. Whatever model you choose, your closest competitor can be running the same one by this afternoon.
An advantage anyone can buy isn’t an advantage for long.
What nobody else can rent
So what’s left? If the engine is broadly the same for everyone, the difference has to come from what you give it to work with.
Here established organisations have something a newer competitor can’t acquire at any price: Their own record. Thirty years of contracts and claims. Correspondence with clients. The reasoning behind decisions that still shape how the business operates. The cases that went well, and the ones that didn’t.
None of that is in any model. A model knows a great deal about insurance and nothing about your policyholders. It can explain how a lease works, but it has never seen yours. That knowledge exists in one place only, and it belongs to you.
That ought to be good news. In my experience it is usually a more awkward discovery, because owning the knowledge and being able to use it turn out to be different things.
Where the power gets lost
After nearly thirty years in information management, I’d say historical information loses its grip in four places.
Format
A great deal of it has never been digital. It sits on paper in storage, or it was scanned years ago as a picture of a page rather than as text. A scan saved as an image has no text to search, so until it has been processed, the systems AI relies on to retrieve information won’t find it. A box on a shelf doesn’t exist at all.
Structure
Even when the text is readable, it is often one undifferentiated mass: A four-hundred-page file with nothing to show where one document ends and the next begins, or which pages matter.
Context
A document on its own rarely tells you whether it is the final version or a draft, whether it was later superseded, who wrote it or why. The people who were there carry that in their heads. AI has only what is written down, and it will treat a superseded policy with the same confidence as the current one.
Permission
Some information shouldn’t be used at all, or only for certain purposes and by certain people. If you can’t say what a set of records contains, you can’t say what AI should be allowed to do with it, and the safe answer becomes “nothing”.
These are not exotic problems. Gartner found that 63% of organisations either don’t have, or aren’t sure they have, the right data management practices for AI, and it predicts that through 2026 organisations will abandon 60% of the AI projects that aren’t supported by AI-ready data[2]. That is a prediction rather than a measurement, but notice where it puts the risk: Not in the model.
What grip looks like
The instinctive response is to digitise everything. I’d resist it. It is expensive, it is slow, and most of what you processed would never be asked for.
A better approach starts from the other end. Pick a question the business would really value an answer to. Which historical claims look like the ones now costing us the most? What did we commit to in agreements signed before the current team arrived? Then work out which historical information would answer it, and test that information against the four points above: Can it be read, is it organised, does it carry its context, and are we permitted to use it?
That turns an overwhelming estate into a defined piece of work with a visible return. It also shows you, usually quite quickly, how far your AI ambitions have run ahead of your data.
The question worth asking
I understand why this gets less attention. Choosing a model feels like progress. Sorting out thirty years of records feels like housekeeping.
But the engine is the part everyone can buy. The grip is the part you have to build, and it is built from material only you have.
So before the next discussion about which model to back, it might be worth asking a different question: How much of what your organisation knows can your AI actually reach?
You already have the power. How much of it is reaching the road?
References
- AI Index Report 2026: Technical Performance Stanford HAI[↩]
- Lack of AI-Ready Data Puts AI Projects at Risk Gartner[↩]