Walk into any large institution today and ask to see its AI strategy. You will almost certainly be handed a list. It will be long. It will be colour-coded. It will be sliced by business unit, by function, by maturity. It will be presented with pride. And it will, in most cases, be a symptom of the problem rather than a solution to it.
The use-case backlog has become the dominant artifact of enterprise AI. It is also one of the clearest signs that an institution has not yet decided what it actually believes about AI, what it intends to do with it, or what it is willing to change in order to use it well.
A list is not a strategy
A strategy expresses a point of view about where the institution will create disproportionate value, what capabilities it will build to do so, and what it will deliberately decline. A backlog expresses none of these. It aggregates demand. It does not order it, fund it against an enterprise thesis, or test whether the underlying problems are worth solving with AI at all.
The honest reading of a 120-item AI backlog is usually this: the institution has not chosen. It has delegated choice to whichever team had the energy to draft a slide. That is not strategy. It is the appearance of motion in the absence of direction.
“Pilot proliferation is the appearance of motion in the absence of direction.”
The hidden cost of pilot proliferation
Every pilot in flight consumes scarce institutional inputs: senior attention, control-function review, data-engineering capacity, vendor governance, change-management bandwidth. Those inputs are not elastic. When they are spread across forty initiatives, none receive enough of them to cross the threshold from interesting to operational.
The result is a portfolio that looks busy and produces little. Worse, it conditions the organisation to believe AI does not work — when in fact the institution simply never gave any single use the depth of investment it required.
What replaces the backlog
A more disciplined alternative does not require fewer ideas. It requires a different relationship to them. Three shifts matter most.
- An explicit enterprise AI thesis — a written, leadership-owned view of where AI will and will not create material value for this institution given its franchise, its data, its regulatory profile, and its operating model.
- A small number of concentrated bets aligned to that thesis, each resourced to a level that makes success operationally plausible rather than merely demonstrable.
- A standing mechanism for evaluating new ideas against the thesis — including the discipline to decline ideas that are interesting but off-thesis, rather than absorbing them into a growing backlog.
Strategy as a function of constraint
Institutions tend to underestimate how much of their AI difficulty is self-imposed. The backlog is appealing because it avoids the harder question of what the institution is for, and where AI fits within that. Answering that question is uncomfortable, because it requires saying no to senior stakeholders whose ideas happen to be off-thesis. But that discomfort is the work.
The institutions that will produce real returns from AI over the next three years are not the ones with the longest backlogs. They are the ones whose leadership has done the harder work of choosing — and who have built the operating discipline to defend those choices against the constant pressure to add one more pilot to the list.
A simpler test
There is a simple diagnostic any executive team can run. Ask each member to describe, in two sentences, the institution's AI thesis. If the answers diverge, the backlog is not the problem. It is the evidence.