Last year I let an AI agent loose on a national catalogue of public-sector semantic models and asked it how interoperable the country’s regions were. In a single session it had read the catalogue, compared the models, scored each domain and explained its reasoning. The scores needed expert review, and I said so at the time. What stayed with me was the speed, and the confidence, with which a piece of software had formed a view about which models belonged together.
That experiment points to something I now consider very likely, and I would like to put it plainly as a prediction.
The prediction
Within a few years, a large share of the modelling decisions in governed ecosystems will be drafted by AI agents. Agents will propose data models for new services, generate mappings between systems, choose which vocabulary to reuse for a new attribute and transform data on the fly when two parties need to exchange it. Humans will review, approve and correct, and the first draft of most of those decisions will come from software.
The pressure in that direction is already strong. The EU Data Act has applied since September 2025, and it widens the circumstances in which data has to be shared between organisations. Data spaces are moving from pilots into operation. Every new exchange needs its semantics settled, and there are far more exchanges than there are skilled modellers. Organisations will reach for agents because the alternative is a queue.
Two futures from the same technology
What happens next depends on a single condition: what the agent can see when it makes a choice.
In the first future, agents choose the way a hurried modeller does today, by plausibility. They pick the vocabulary that appears most often in their training data, or the one that fits the immediate task most neatly, or they generate a fresh model because that is quickest. Each choice is reasonable in isolation. Multiplied by thousands of agents working at machine speed, they produce divergence faster than any programme could ever map, and the alignment debt I have written about elsewhere grows at a pace no budget can follow.
In the second future, agents can see what the parties around them have actually converged on: which models their peers import and build on, which they map to, which they run in production for the scope at hand. An agent with that evidence will tend to choose what its peers already use, and every such choice leaves one more piece of evidence for the next agent. In that future, the speed of the machines works for convergence. Ecosystems could align faster than at any time in the history of our field.
The technology is the same in both futures. The difference lies in the information the ecosystem makes available to it.
Why language models alone will not settle it
It is tempting to think that capable language models will simply understand each system well enough to translate on demand, so that shared models matter less. I find that view optimistic. A model can translate between two descriptions it has never seen together, and often it will do so well. It cannot know which of several plausible interpretations a particular organisation has committed to, or which design its partners rely on in production, unless that information is published somewhere it can read. Meaning that has been agreed has to be recorded, and agents need to find the record.
Questions to prepare your ecosystem
- When an agent in your ecosystem chooses a model, what will it read? If the answer is “whatever it was trained on”, the first future is the default.
- Is the evidence of what your parties have converged on published in a form a machine can query? Catalogues that list assets are a start; evidence of who builds on them is what an agent needs.
- Have you named the parties and the scope an agent should align with? Without a circle and a purpose, an agent has no referent either.
- Who reviews the models and mappings agents produce, and against what?
- Are agent decisions recorded, so that each one adds to the evidence for the next?
Getting ready for the speed
I do not expect the transition to be gradual for long. The organisations and ecosystems that prepare will find that agents make convergence cheaper than it has ever been. Those that wait will discover their divergence growing at machine speed. Making convergence evidence readable, by people and by agents, at the moment a decision is made, is the problem we built EVIDENT around at METALEVELS.
If you would like to think through what this means for your own ecosystem, start a conversation with us.