The constraint is no longer what is technically possible. It is deciding what deserves funding, and most shortlisting processes reward enthusiasm over evidence.
Filter candidates on four questions. Does a measurable baseline already exist, or can one be captured cheaply. Does a named person own the outcome the use case would improve. What is the cost of the system being wrong, and who absorbs it. Would anyone notice if it stopped working. Use cases that pass all four are usually unglamorous administrative workflows, which is consistent with where surveyed adoption has concentrated.
Two years ago the question was what AI could plausibly do in a health system. Deloitte's September 2025 survey of 100 US healthcare technology executives suggests the constraint has moved: 40% said technical talent limitations were no longer a major challenge, 35% reported improved leadership buy-in, and 32% reported reduced data quality concerns.
When the barriers ease, the scarce resource becomes judgment about what to fund.
If you cannot state what the process costs today, you will not be able to demonstrate improvement later. Use cases where the current state is already instrumented start with a large advantage.
Not the project. The outcome. If no named person's performance improves when this works, nobody will drive adoption past the first obstacle.
A scheduling error is recoverable. A missed clinical finding is not. This filter should determine how much oversight the use case needs, not whether to proceed.
A useful proxy for whether the use case addresses real work. Tools that could quietly fail for a month without complaint were solving a problem nobody had.
Applied honestly, these filters tend to eliminate the most exciting proposals and keep the dull ones. That matches where adoption has actually concentrated: repetitive, rules-heavy, administratively painful workflows such as prior authorisation, documentation, scheduling and revenue cycle, rather than autonomous clinical decision-making.
Shortlisting processes usually fail for a political reason rather than an analytical one. The proposals arrive attached to sponsors, and the ranking exercise becomes a negotiation between departments in which the best-argued case wins rather than the best case.
The fix we recommend is to score against the filters before anyone knows which use case is whose. Strip the sponsor names, score the workflows, then reveal. Teams resist this and it consistently changes the ranking, which tells you what the original process was actually measuring.
One more thing worth saying plainly: a good shortlist has things on it you have decided not to do, with reasons recorded. If everything proposed survives to a pilot, you did not run a shortlist, you ran an intake process.