A WORKED EXAMPLE
The capability you let lapse
A worked example – What an adopt-first AI policy cannot see – the one cost that shows on no dashboard, and arrives only on the day the model is wrong and no one is left who can tell.
Most AI policy starts from the wrong question. It asks how much to adopt, how fast, and how to drive usage across the organisation – and in asking it that way, it has already made adoption the goal. A means has been seated where an end belongs before any particular use has been examined. From there the policy will work, efficiently, to increase a thing whose worth it has not established.
The wrong question has a number waiting for it, the way every means-made-end does. Make adoption the aim and the figure that lands on the dashboard is usage: how much of the work now passes through a model. It reads well when the model truly helps, and just as well when people route work through it to make the number rise – usage measures motion, not benefit, and a workforce told that usage is the target will produce usage. The line glows while the question of whether the work is any better goes unasked.
The right question is not how much but, task by task, for what. A model is a capability, not a direction; the policy that holds does not bless it or refuse it, but asks of each use whether handing this task to a model builds up what the organisation is for, or feeds on it. That much is the discrimination any change demands. What is particular to a model, and what an adopt-first policy is structurally unable to see, is a second cost – and it is the one that appears nowhere on the dashboard.
When a task is handed to a model, the faculty that used to do it begins to lapse. Not dramatically; quietly, the way an unused capability always wastes. For a while nothing seems lost, because the model is right often enough that the lapse does not show. The cost arrives later and all at once, on the day the model is wrong – confidently, plausibly wrong, as models are – and the people who would once have caught it can no longer tell. The judgement that would have known the output was off has gone soft from disuse, and the organisation discovers, too late, that in delegating the doing it also delegated the ability to know when the doing was bad.
This is the failure no usage metric can register, because it is an absence – a capability no longer there – and absences do not show up where the dashboard is looking. It has one faint early sign, the same one every act-level failure has: the unease of the person who can see the model producing fluent, well-formed, wrong work while the numbers stay green. That unease is worth more than the dashboard, and an adopt-first policy, busy driving usage upward, is exactly the regime most likely to wave it away.
So a policy that holds does three things, none of which is a position on the technology. It refuses to make adoption the goal, and asks of each use what it is for. It keeps people capable of catching the model wrong – deliberately, by keeping the underlying faculty in use wherever a confident error would be costly, treating the preserved capability as a price worth paying rather than an inefficiency to be removed. And it pairs its measures, so that usage never travels alone but always beside some sign of whether the work the model now does is work that serves.
None of this turns on which model, or this year’s capabilities, or how good the tools become. If anything the better and more fluent the model, the sharper the second cost, because the more trustworthy the output looks, the faster the faculty that would check it lapses. So when the next tool arrives and the question in the room is how quickly the organisation can adopt it, the useful question is the older one: for what, and what must we stay able to do ourselves. The capability you let lapse is invisible until the day you need it – and on that day, the dashboard that told you adoption was going well will have nothing to say.
This is one of the moves the programme makes visible. The full account of what it is and how it works is here.