The Test for What Can Become an AI Skill
18
Overview
Most AI Skill pitches fail before they start, and nobody checks for it. Here's the one question that disqualifies half your list before anyone spends a dollar scoping it.
Year
2026
Industry
Financial Services / COOs & Operations
Challenge
Ask ten people how to prioritize AI Skills and you'll get ten different lists. Ask them how to tell whether something can become an AI Skill in the first place, and most go quiet. That's the real gap. Not sequencing. Screening. Last week's tier map showed a pattern: back office work clusters at the top, front office work clusters at the bottom. True. But the tier map is a symptom, not the cause. The test underneath it has nothing to do with which office someone sits in. It's this: can you write the finish line as a rule, or does finishing the task require someone to make a call? Write the rule and you have a candidate. Can't write it, and you don't have an AI Skill. You have a job. A workflow diagram doesn't change that, and neither does a good demo. A candidate that isn't one Take credit exception review. It looks automatable. There's an intake form, a checklist, a decision at the end. Watch what actually happens on a borderline request, though: the analyst doesn't run the checklist mechanically. They weigh the client relationship. The size of the exposure. What happened last time, and whether the desk head will sign off if this one goes wrong. None of that lives on paper. It lives in the analyst's head, and it moves case by case. Wrap that in an AI Skill and you haven't automated her judgment. You've hidden it. The system will spit out an answer every time, fast and confident, and nobody will know it's guessing at the part that actually mattered. That's not a model problem. It's what happens when you skip the test. The disqualifier table So before anyone scopes a project, run the one-question test against the list. Here's what disqualifies a candidate, what tips you off that it's failing, and what has to change before it's worth a second look.
Impact
None of these mean the activity is stuck forever. They mean it isn't ready yet, and most of the fixes are cheaper than people assume. Writing down one definition of “done” costs a meeting, not a project. Turning “ask the desk head” into documented criteria costs an afternoon with them. What doesn't work is skipping the test and hoping the model figures out the rule on its own. It won't. It was never going to. The rule has to exist before you hand the work over. Because that's what an AI Skill actually is: not intelligence. A codified rule, running at scale, with nothing underneath it holding it up. Run this test on your own list before anyone scopes a project. Half of what looked automatable will fail it, and that finding is worth paying for. Not “here's the AI you could use.” Here's what's actually ready, what isn't, and what has to change first. Fix the operating model. AI works after that, not before.