Henry Ford and Thomas Edison built their reputations inside the machinery itself. They understood how the work was actually done because they were doing it, modifying it, breaking it, alongside the people whose hands were on it. For a long stretch of industrial history, that proximity was simply what business leadership looked like — not a founder's eccentricity, but the ordinary condition of running anything that depended on a new technology.
Then the centre of gravity moved. Management became a financial discipline first and an operational one second. Technology was cordoned off into its own department, staffed by specialists, reported on rather than touched. The people steering the business stopped needing to understand the thing they were steering, because steering had been redefined as something you could do from the boardroom with the right reports in front of you. This wasn't a failure of discipline. It was a deliberate trade, and for a long time it worked well enough that nobody questioned it.
AI has reopened the question, because AI doesn't behave the way the technologies of that earlier trade behaved. It doesn't sit quietly inside an IT department waiting to be specified. It is genuinely strange to use, surprising in what it gets right and wrong, and the only way to develop any real intuition for what it can and cannot do is to put your hands on it. None of that intuition is available secondhand, however good the report is.
Watch what boards do anyway. A board decides it needs "an AI strategy." It commissions a review, runs a vendor process, stands up a steering committee, signs off a platform. Every one of these steps is procedurally sound. None of them require a single person making the decision to have actually used a large language model, sat with its failure modes, or felt the specific discomfort of not knowing whether to trust its output. The decision gets made entirely in the register of information about the technology — capability decks, comparison matrices, analyst reports — and never once touches experience of it.
This is usually described, when it goes wrong, as a governance failure: the wrong vendor, insufficient oversight, a steering committee that didn't ask hard enough questions. That description is too generous, because it implies the failure happened during governance, and that a better-run process could have caught it. It didn't happen during governance. It happened earlier, at the moment the decision was handed to a process built to evaluate information rather than generate judgement. A steering committee can interrogate a vendor's claims. It cannot manufacture the intuition that comes from use, no matter how rigorously it's chaired. The steering wheel wasn't lost through bad governance. It was never connected to anything, because nobody driving had their hands on it to begin with.
It would be a relief if the next move corrected this. It doesn't. Watch what happens after the platform is live: the board, sensing it has lost some control it can't quite name, reaches for structure — an AI governance framework, a centre of excellence, a council that reviews use cases before they're approved. These are, on paper, the tools of adaptability — methods built, in other contexts, to let organisations sense and respond rather than plan and impose. And they get used, almost without exception, the way Scrum gets used inside organisations that adopted its rituals without changing how they relate to the work underneath them: as instruments of control wearing the language of flexibility. A council that reviews use cases before approval is not steering by immersion. It is the original mistake, repeated, now with a more reassuring name.
None of this requires heroics to avoid. It requires something much smaller, and much harder for a board to admit it needs: someone who has used this thing, on the business problem, recently, and felt exactly where it broke. Not a sponsor. Not a champion. Someone with the unglamorous, almost clerical experience of having sat with the technology long enough to have an opinion about it that didn't come from a deck. Organisations that have this rarely think of it as a strategy. It looks, from the outside, like someone junior being allowed to waste an afternoon. It is, in practice, the only mechanism by which the room develops any capacity to steer at all.
So the question worth sitting with isn't whether your organisation has the right AI governance framework. It's whether anyone who decided to build one has ever used the thing it's meant to govern. And once you start asking that about AI, it's difficult to stop asking it about every other technology decision the same room has already made the same way.