Perspectives

Most institutions responded to the last three years of AI progress the same way: they accelerated. Task forces, pilots, platform commitments, a transformation narrative for the annual report. The instinct is understandable. Falling behind is legible and embarrassing; being pointed slightly wrong is invisible for quarters at a time.

But watch what actually happened. A large share of those institutions have now rewritten their AI strategy at least once, some twice. The technology did not reverse course — the capabilities of early 2023 still exist and still work. What changed was the institution's understanding of what the technology was for, in its own specific context. The first strategy was written before that understanding existed. It was speed without a datum.

Navigation has a precise word for what was missing. A datum is the reference point from which every measurement is taken; without one, position reports are noise and velocity is a liability. The organizations that have handled this period well are conspicuous for how little they resemble the fastest movers. They fixed a reference point first — a concrete view of which decisions the technology would be allowed to touch, what the institution must remain able to explain, and which dependencies it could not afford to accept. Then they moved quickly, once, in roughly one direction.

There is a reason this discipline is rare. Orientation work is invisible in the quarter it happens and only pays in the quarters after. Acceleration, by contrast, photographs well. Boards should be suspicious of that asymmetry: in stable conditions, momentum forgives small errors of direction. In a discontinuity, it multiplies them. The correction, as always, is cheapest before the speed.

Ask an executive team who is allowed to approve a credit line, sign a contract, or ground a fleet, and you will get precise answers. Institutions are, among other things, machines for allocating decisions. The org chart, the delegation matrix, the approval workflow — all of it encodes a long-negotiated settlement about who decides what, with whose money, at whose risk.

AI does not merely automate tasks inside that settlement. It quietly renegotiates it. When a model drafts the analysis, ranks the options, and pre-fills the recommendation, the human "decision" that remains is often a ratification made in ninety seconds by someone who trusts the pipeline. Formally, nothing has changed; the same name is on the approval. Functionally, the decision has migrated — into training data, prompt design, and vendor model updates that no one in the approval chain reviews. We have taken to calling this the ratification problem. Most institutions cannot yet see how much of their approval architecture it has quietly absorbed.

This is why treating AI as an adoption program understates the problem. Adoption asks "where can this make us faster?" Redesign asks harder questions: which decisions may a system recommend but never execute; where must a human be able to reconstruct the reasoning, not just audit the outcome; what happens to accountability when the person who signs no longer performs the judgment being signed for. These questions have owners in every institution — they are just scattered across legal, risk, technology, and operations, and no standing forum exists where they meet.

The institutions that get this right tend to do something unfashionable: they slow down at exactly one point. Before a system touches a consequential decision class, they write down the decision rights as if drafting a delegation of authority — because that is precisely what it is. Everything else can move fast. The settlement about who decides is the one thing an institution should never discover it renegotiated by accident.

When software fails, the cost is usually a bad output and a correction. When an autonomous system operating machinery fails, the cost is a bent asset, a stopped line, or a person. That single difference explains most of what leaders find confusing about the physical-AI market: why the demonstrations are dazzling while deployment is slow, why capital requirements refuse to behave like software economics, and why the vendor's reference customer is always in a different industry than yours.

The market is best understood not as one shift but as a set of venues, each with its own physics of risk. A warehouse is a controlled, mapped, semi-private environment where errors are mostly financial — which is why autonomy arrived there first. A port adds weather, unions, and national infrastructure designations. A surgical suite adds a liability regime measured in lives and a regulator that moves at the speed of evidence. A mine adds remoteness that makes autonomy attractive and consequences that make insurers nervous. The same perception stack, the same marketing language, radically different adoption curves.

For boards and investors, this venue-specificity is the entire game. The right questions are rarely about the robot. They are about the environment: how structured is it, who else is in it, what does an error cost, who bears that cost under current law, and what happens to the operating model when machines require supervisors instead of operators. Two opportunities that look identical in a deck can sit a decade apart once those questions are answered.

The discipline, as ever, is resisting the abstraction. "Robotics" is not a strategy, any more than "AI" was. The institutions that will do well in physical autonomy are already doing the unglamorous work: mapping their own venues, pricing their own error costs, and deciding — before the vendor visit — what evidence would actually justify letting software act on the world they are responsible for.