Almost every large company I talk to today has its own list of AI use cases. A customer service assistant here, a coding copilot there, a triage automation in some department nobody outside it has heard of. The list keeps growing. Until, in some board meeting, someone raises a hand with a hint of discomfort and asks whether all of this adds up to anything, or whether the company is just collecting projects.
A collection of use cases is a set of separate bets. Each team justifies its own project inside its own silo, and life goes on. An actual AI strategy is a decision about where the company is going to put capital and attention, which necessarily means deciding where it isn’t going to invest right now. A list can keep growing forever without ever becoming a strategy. In fact, when the list grows too fast, that’s usually the clearest sign the strategy hasn’t been decided yet.
Some people already have a name for this stage: pilot purgatory, or the PoC graveyard. What’s usually missing isn’t a better name for the problem. It’s the decision that would get a company out of it.
Here’s a quick test for telling which situation your company is actually in. Ask the executive team: if we had to cut 80% of the AI initiatives running today and keep only the ones that matter most, what would survive, and why? If the answer comes quickly, with a clear and defensible criterion, a strategy exists. If the answer takes a while, or turns into “it depends on the team,” or triggers quiet political pushback because no one wants their pet project to be the one that gets cut, what you have is a collection being managed like a portfolio, without anyone actually in charge of the allocation.
Herbert Simon, the economist and cognitive scientist who spent his career studying how decisions actually get made inside organizations, had a term for this behavior: satisficing, the tendency to settle for the first good-enough option when the cost of deciding properly feels too high. Approving one more isolated use case is politically cheap. Nobody has to say no to anybody, and the committee leaves the room with a sense of progress. Building an actual strategy costs the opposite. Someone has to say no to entire departments, and sit with the discomfort of justifying why their area didn’t make the priority list.
That’s why the list grows on its own, through sheer organizational inertia, while a real strategy has to be forced into existence against the current of internal politics. No committee meeting applauds the person who proposes cutting something. Every committee meeting applauds the person who shows up with one more pilot.
None of this is unique to AI. The same pattern showed up, almost scene for scene, in earlier waves of corporate technology. Companies spent the 2000s piling up dozens of half-implemented ERP systems for the same reason: it was easier to approve one more module in one more business unit than to decide, from the top, which architecture the whole company would standardize on and which legacy systems would get switched off. The lesson never came from the technology itself. It came from how committees decide under pressure from competing internal interests, and AI is just the newest chapter of that same story.
One practical way to tell a collection from a strategy is to look at where the conversation actually happens. Isolated use cases tend to get approved inside each department, one level below the board. A real strategy has almost always been argued out and decided at the level where trade-offs between departments can actually be made, because that’s the only place with the authority to say one department accelerates on AI now and another one waits.
It’s worth admitting that having a collection isn’t, by itself, a mistake. Early on, it makes sense to let different teams experiment, fail, and learn before any bigger allocation decision gets made. The problem isn’t starting that way. The problem is staying that way after the learning phase has already produced enough signal to decide, and nobody decides, because deciding hurts more than approving one more pilot.
In the assessments I’ve run, the most common pattern isn’t a company with no idea where AI creates value. It’s the opposite: too many ideas, competing with each other for executive attention, with no one holding a clear mandate to choose between them. What’s usually missing isn’t vision. It’s the decision that would turn several partial visions into one actual bet.
If the 80% test made you uncomfortable just reading it, that’s a fairly reliable sign your company is still running a collection, not a strategy. There’s no shame in that. It’s simply the next decision leadership hasn’t made yet, and it won’t show up on its own, disguised as one more use case on a slide.