Every AI steering committee we sit in on has the same problem, and it is not a shortage of ideas. It is an abundance of them. Sales wants a chatbot, operations wants a forecasting tool, finance wants anomaly detection, and someone senior wants something with the word generative in it because a competitor announced one. The list grows faster than anyone can evaluate it, and evaluation, when it happens at all, tends to reward whichever idea has the loudest sponsor rather than the one that would actually move the business.
This is not a technology problem. Model quality has stopped being the binding constraint for most practical use cases; usable models are now a commodity input, not a differentiator. The binding constraint is organisational: the ability to triage a pipeline of use cases with the same rigour applied to capital spending, and the willingness to kill ideas that fail the test. Few organisations have built that muscle, and it shows in how many AI pilots quietly die in a proof of concept purgatory that never gets reported upward.
Novelty is not a selection criterion
The single biggest mistake in use case selection is letting novelty stand in for value. A use case that sounds impressive in a steering committee slide, summarising legal documents, generating marketing copy, answering customer questions with a large language model, is not automatically worth doing. The question that matters is duller: how much value does solving this problem create, how often does the problem recur, and how expensive is it today.
Value density, the amount of benefit per unit of effort, is a far better filter than technical sophistication. A modest automation that removes two hours of manual reconciliation from every finance close is often worth more than an ambitious model that impresses in a demo but touches a process nobody spends much time on. We have watched organisations pour a year of engineering effort into a flagship AI use case that, on close inspection, addressed a task consuming perhaps three person-days a month. The effort would have paid for itself faster as a spreadsheet macro.
The binding constraint on AI value is rarely the model anymore. It is the discipline to say no to the wrong use case before it consumes a year of engineering time.
Data readiness decides more than ambition does
The second filter, and the one most steering committees skip, is whether the data underneath the use case is actually fit for the job. Not whether data exists somewhere, but whether it is labelled, current, consistent across systems and legally usable for the purpose intended. A use case can be strategically brilliant and still be twelve months away from viable simply because the underlying data has never been cleaned, structured or governed with this application in mind.
This is why we ask teams to score use cases on data readiness before they score them on business value. A high-value idea sitting on unusable data is not cancelled, it is resequenced; someone needs to own the unglamorous work of preparing the data before a model gets anywhere near it. Skipping that step is how organisations end up with a model that performs beautifully in a sandbox and fails quietly the moment it meets production data.
- 路Score every candidate use case on value density, not on technical interest
- 路Separate data that exists from data that is usable for the specific purpose
- 路Assign an accountable owner to data preparation before any model work starts
- 路Require a kill decision date for every pilot, not just a start date
The politics of saying no
Even with a sound framework, the hardest part of triage is social rather than analytical. Killing a use case usually means telling a sponsor, often a senior one, that their idea is not going forward. Most governance processes are built to avoid this conversation, which is why so many low-value pilots survive: nobody wants to be the person who says no to the executive whose idea it was.
The organisations that do this well have removed the decision from any individual and placed it in a standing framework applied consistently, so that a no is a function of the criteria rather than a personal judgement. That depersonalises the decision and makes it survivable for the relationships involved. It also means the same rules apply when the CEO's pet idea comes up for review, which is the real test of whether a governance process has any teeth.
Building the muscle, not just the model
None of this argues against ambition in AI. It argues for spending ambition on the ideas most likely to earn it back. A portfolio approach, a small number of well-chosen use cases moved quickly through to production, consistently outperforms a scattergun approach where twenty pilots start and two ever ship.
The organisations getting genuine value from AI right now are not the ones with the most advanced models. They are the ones that built a disciplined process for choosing what to build in the first place, and had the nerve to use it.
If your AI programme has more pilots than it has production deployments, the problem is very unlikely to be your data scientists. It is more likely to be a selection process that never learned how to say no.