The AI use-case scorecard

If your scorecard has never killed an idea, it is not a scorecard. It is paperwork with a grid on it.

I say that because I have built a few of these and watched most of them get used backwards. The scorecard gets filled in after the decision, by the person who wants the project, to justify what they already told their boss they were doing. Value, feasibility and risk is a perfectly good rubric. Scoring it honestly is the entire difficulty, and nobody writes about that part.

The problem it solves

You have somewhere between fifteen and forty AI ideas. They came from a hackathon, a vendor lunch, a consultancy deck and your own team’s group chat. Every one of them is plausible. That is the trouble - plausibility is free.

Enthusiasm is not prioritisation. Without something to rank against, you end up funding whichever idea had the most senior sponsor, which is how organisations end up with an AI chatbot nobody asked for and no summarisation of the sales calls everyone wanted.

The three axes, defined properly

Value. A number, not an adjective. Hours saved times loaded cost, revenue affected, or a decision that happens days earlier. “Efficiency” is not a value score. If you genuinely cannot put a number on it, write that down, because that is the finding rather than an inconvenience.

Feasibility. Nine times out of ten this is a data question and not a model question. Do you have the data. Is it joined. Does anyone own it. The model is almost never the blocker, and the join almost always is.

Risk. What happens when it is confidently wrong, and who is the first person outside the room to notice? That second half is the useful bit. An internal draft that is wrong gets corrected in ten seconds. A customer-facing answer that is wrong gets screenshotted.

Four ways it gets scored dishonestly

This is the section I would actually keep if you only kept one.

Value gets inflated, because the person scoring wants the project. Feasibility gets scored by somebody who has never opened the table. Risk gets fudged, because it is the only axis without a natural number attached, so it drifts towards medium. And the whole thing gets scored exactly once and never revisited, which means a pilot can prove the feasibility score wrong and change nothing.

The fixes are dull and they work. Score it with two people who disagree in the room. Make the data owner score feasibility, not the sponsor. Score risk by naming the first outsider who notices. Re-score after the pilot, because the pilot exists to produce information and you may as well use it.

Do not average the three

This is the bit most templates get wrong. They add the scores up, produce a single number, and rank by it.

Risk is not an averageable axis. It is a gate. High value plus high risk does not come out as medium and it does not mean “medium priority” - it means do it, but internally, with a human in front of it, until the risk drops. Averaging hides exactly the thing you built the axis to surface.

Keep the whole thing to one page. If your scorecard needs a workshop to fill in, the scorecard has become the project.

[illustrative] A B2B SaaS company, around 2024, scored 22 ideas. The highest scorer was AI-written outbound sales email, which also carried the worst risk profile in the set and would have been the first thing a customer saw. Third place was summarising sales calls into CRM fields. That shipped in a fortnight and was still in daily use a year later.

Where I would argue with myself

Scorecards have a real cost, and it is speed. For a cheap, reversible experiment the correct move is to just try the thing. Scoring a two-hundred-pound test across three axes is bureaucracy wearing a governance costume, and I have watched that habit kill more good ideas than any bad score ever did. My rough line is that if you can test it inside a week for less than a rounding error, skip the grid.

The deeper objection is that every score is a guess, and the grid makes guesses look like measurements. Three numbers in a table carry an authority the underlying thinking has not earned. Worth holding loosely.

What to do this week

Score your current list. Then look for the ideas that scored badly enough to stop.

If nothing on your list is dead by the end of the exercise, you did not score it, you ratified it. Go back and get the person who owns the data to score feasibility.

---

Editing notes

• No dates or vendor claims in this piece, so nothing to fact-check. That makes it low-risk to publish but also means it leans entirely on your judgement and experience rather than on evidence - it is the most opinion-dependent of the four.
• [illustrative] on the 22-ideas SaaS example. A real version of this would do a lot of work, particularly the detail about the boring idea outperforming the exciting one.
• The “do not average the three” section is the most original claim here and the one most likely to get pushback. I think it is right.
• Deliberately avoided: giving the framework a name or an acronym. That is the fastest way to make this piece look like the consultancy decks it is arguing with.