How to prioritize AI projects

Why prioritisation is where AI spend is won or lost

AI budgets are finite; AI ideas are not. The question is never “is this a good idea?”, almost every pitch is. The real question is “is this a better bet than everything else competing for the same capital?” Teams that skip that comparison fund the loudest champion, not the highest-value opportunity.

The cost of getting this wrong is well documented. MIT’s NANDA initiative found 95% of enterprise generative-AI pilots produce no measurable P&L impact, and RAND puts the AI project failure rate above 80%, twice that of non-AI IT. Prioritisation is the cheapest place to avoid that waste, before the build budget is committed.

The five criteria to score every AI bet

Comparability is the whole point. Score each candidate on the same five dimensions, on a simple 1-5 scale, so a customer-service agent can be weighed against a document-processing workflow without comparing apples to oranges:

• Business value: the size of the P&L impact if the bet pays off, in revenue, cost, or risk reduction.
• Strength of evidence: how much real proof, not opinion, supports the expected value today.
• Adoption likelihood: whether the people who must use it actually will, day to day.
• Governance and compliance risk: the regulatory, security, and reputational downside of being wrong.
• Time-to-impact: how quickly the bet can produce measurable value.

Rank the portfolio, then fund from the top

Weight the five criteria to your context, most enterprises weight business value and strength of evidence highest, and combine the scores into one ranked list. That single view is the portfolio: every AI bet, ordered by how much value the evidence says it will create per dollar.

Funding then flows top-down until the budget is spent. Bets below the line aren’t rejected forever; they’re held until they earn more evidence or the portfolio is re-scored.

Use pretotyping to break the close calls

Scoring gets you a ranked list, but the bets clustered right around the funding line are where the real money is made or lost, and they usually differ on one thing: the strength of evidence. That’s the criterion you can cheaply improve.

Pretotyping, created at Google and taught at Stanford, is how you improve it fast. A fake-door test, a Wizard-of-Oz run, or a concierge trial produces real behavioural evidence, did people use it, did they pay, in days rather than quarters. Run one on each close call and the tie breaks itself on data, not debate.

Set fund, fix, or kill thresholds up front

Prioritisation only holds if the rules are agreed before the scores come in. Decide in advance the score and evidence a bet needs to be funded, sent back for more proof (fix), or stopped (kill). Pre-committed thresholds turn a hard political conversation into a routine decision the board can defend.

This is exactly the discipline an AI Bets Audit installs: in two weeks we map and rank your AI bets, pretotype the highest-stakes ones, and hand you a defensible fund / fix / kill call for each, drawing on the same method we’ve run across 4,000+ enterprise experiments.