The Next Ten Years of AI Will Be Decided by Boring Things

The public conversation about artificial intelligence is organised around capability: what the next model will be able to do, and when. The things that will actually determine how much of this technology gets used are considerably less exciting.

Queues, permits, cover and 1990s interfaces

• Grid interconnection queues that determine when a data centre can operate.
• Municipal permitting for power, water and buildings.
• Insurance products: whether a model’s output can be covered, and at what price.
• Procurement rules in public sector and regulated industries.
• Liability frameworks that determine who pays when a system fails.
• The mundane cost of integrating with systems built in the 1990s.

Each one is a rate limiter

Each item is a rate limiter. A better model released next year is worth little to an organisation that cannot get approval to deploy it, cannot insure it, cannot connect it to its data, and cannot explain to a regulator what it does.

The organisations that will capture the most value are the ones that have solved these problems, and those solutions are administrative rather than technical.

Read utility filings, not research announcements

Anyone trying to predict the trajectory of this technology should be reading utility filings, insurance underwriting guidelines, and procurement notices rather than research announcements. Those documents move slowly, they are public, and they constrain the deployment curve far more tightly than any laboratory’s roadmap.

It is a less thrilling way to follow the field. It is also more likely to be right.

Insurers have already excluded AI claims

Nothing gets deployed at scale in an industry without cover. Insurers are currently trying to price policies for systems whose failure modes are poorly understood and whose liability allocation is unsettled, and in several lines they have simply excluded AI-related claims.

That exclusion is more consequential than any regulation currently in force. A hospital, a bank or a logistics operator cannot adopt a system that voids its cover, regardless of how good the system is.

The model is the small part of the project

Most large organisations run critical processes on systems built decades ago with interfaces that were never designed to be automated. Connecting a modern model to those systems is unglamorous, expensive and the actual work of deployment.

Companies that have done it describe the model as a small part of the project. That is a less quotable finding than a capability announcement and a considerably more accurate description of where the effort goes.

There is a second-order effect worth watching. Organisations that understand this tend to invest in the unglamorous capabilities early, which is why the gap between sophisticated and unsophisticated adopters keeps widening. The technical capability is available to almost everyone at roughly the same price; the ability to deploy it is not.