How To Prioritise AI Investments Without Guessing

Most AI investment decisions get made by whoever argues loudest. A scoring model built on strategic impact, delivery risk, and capability fit stops that.

Your AI backlog is not the problem. Every business unit has ideas. The problem is that the ideas compete for budget without a shared decision rule, so the one with the most enthusiastic sponsor wins. That is not a prioritisation system. It is a lobbying contest.

The pattern that keeps repeating

Enterprise AI programmes show a consistent failure mode: organisations fund a pilot, generate learning, then fund a different pilot from a different advocate rather than scaling what worked. The CSIRO's guide on evaluating and prioritising AI projects identifies the mechanism directly. Decisions driven by enthusiasm or internal politics, not scored criteria, determine which initiative gets funded next. The result is a portfolio of perpetual proof-of-concept work with no production-scale outcomes and no reusable capability infrastructure underneath.

This is worth naming precisely because it does not feel like a governance failure from the inside. Each pilot looks reasonable. Each sponsor has a plausible business case. The problem only becomes visible at the portfolio level, when you notice that the organisation has run twelve pilots in three years and scaled none of them.

What a scoring model actually forces

The CSIRO framework scores AI initiatives across strategic alignment, technical feasibility, ethical risk, and net present value expressed as scenario ranges rather than point estimates. Hansen and Svejvig's review of seven decades of portfolio management research shows why scenario ranges matter more than single-number ROI forecasts: AI value estimates carry enough uncertainty that a point estimate is mostly a confidence signal, not a financial projection.

A scoring model does not eliminate that uncertainty. It applies the same uncertainty consistently across every candidate initiative, which makes comparisons between projects more honest even when the absolute scores are imprecise. That is the actual value. Not precision. Consistency.

The false precision problem cuts both ways

A reasonable CFO objects here. Scored multi-criteria models aggregate uncertain inputs into a number that looks rigorous but is not. If strategic impact scores and capability fit ratings are judgment calls, multiplying them together and ranking the results produces false confidence, not better decisions. Fund the use cases with clear 12-month financial returns instead. At least those predictions are falsifiable within a budget cycle.

This argument is correct about the uncertainty. It fails on the evidence. The research on enterprise AI programmes shows that ROI projections for AI use cases are routinely optimistic and routinely wrong. The gap between stated AI ambition and realised value is a documented pattern across large organisations, not an edge case. A CFO who funds the three initiatives with the strongest ROI cases is not funding the three with the most accurate financial projections. They are funding the three with the most persuasive sponsors, which is the exact failure mode the scoring model exists to prevent. Fast ROI filters are just as vulnerable to political manipulation of input assumptions as any other method, and they generate no comparison record that accumulates over time.

The third axis most models skip

Strategic impact and delivery risk get included in most prioritisation approaches. Capability fit gets dropped. This is a mistake with a structural consequence.

Strict financial filters systematically defund the foundational data, platform, and responsible AI work that makes high-return use cases viable. You cannot scale a customer-facing AI application on a data infrastructure that was never built to support it. Google's 70-20-10 innovation allocation model exists partly to protect this kind of investment from being killed by near-term ROI demands. The OECD Due Diligence Guidance for Responsible AI and the EU AI Act add a second reason: regulatory risk class is now a required input into AI investment sequencing, not an afterthought. An initiative that scores well on financial return and strategic alignment but sits in a high-risk regulatory category under the EU AI Act carries compliance exposure that a ROI filter will not catch.

Capability fit scoring forces this question before budget is committed. Does the organisation have the data, the platforms, and the talent to deliver this initiative at production scale, or is the plan to build all three simultaneously while also delivering business value? The answer changes the risk score significantly.

What to do with the scoring output

Score every candidate initiative on the same criteria in the same budget cycle. Weight the criteria through explicit stakeholder discussion rather than defaulting to financial return. Use scenario ranges for value estimates, not point forecasts. Build a comparison record across cycles so the organisation can test whether its scoring weights predicted actual outcomes over time.

The investment committee that does this will still make wrong calls. The difference is that the wrong calls become visible and correctable rather than invisible and repeated.