The CFO's AI Dilemma: Funding Innovation Without Delivering a Blank Cheque

Traditional capital investment frameworks are built around a simple premise: you specify in advance what the investment will produce, you apply a discount rate, and you approve investments where the NPV is positive. This works for capital expenditure with predictable returns, for acquisitions where the synergies can be estimated, and for technology investments where the cost savings are definable in advance.

AI investment has a different structure. The returns are partially predictable (productivity improvement from well-understood tools like Copilot) and partially emergent (the value that comes from discovering the highest-impact use cases through experimentation and learning). Requiring full specification of returns before investment eliminates the exploratory investments that often produce the highest eventual value. But accepting a general commitment to innovation as the financial justification enables unlimited expenditure with unlimited deniability when returns do not materialise.

The most effective AI investment framework for CFOs adapts the stage-gate model from R&D investment management. Rather than requiring full justification upfront, it allocates investment in stages with defined gates at which evidence of progress is required before the next stage is funded.

Stage one is exploration: a small, time-limited budget for a specific use case, with the objective of generating evidence about whether the use case delivers the hypothesised value. The financial justification at this stage is the hypothesis and the cost of testing it, not a fully specified ROI projection.

Stage two is validation: a larger budget to deploy the use case with a representative cohort, measure business outcomes against the pre-defined metrics, and produce a ROI case based on actual evidence rather than projections. The financial justification at this stage is the evidence from stage one.

Stage three is scale: full investment based on the validated evidence from stage two. At this stage, the ROI case is evidence-based and can support a traditional investment framework.

This approach funds exploration without requiring premature certainty, but also does not fund scale without evidence. The gate between each stage requires real evidence, not optimistic projections.

Individual AI investments should be managed as a portfolio with different risk/return profiles and different investment frameworks.

Fix-and-operate investments are AI deployments that address known, quantifiable inefficiencies with well-understood tools. The ROI case for deploying Copilot to a specific team with a defined productivity measurement framework is a fix-and-operate investment. It should meet a standard hurdle rate with a defined payback period.

Grow investments are AI deployments that create new capability or open new market opportunities where the return is less certain but the strategic value is high. These merit a higher risk tolerance and a longer evaluation horizon, but not unlimited tolerance.

Explore investments are early-stage experiments where the primary return is learning. These should be subject to a defined budget and timeline, with a specific learning objective, and a clear decision point about whether to proceed to grow investment or discontinue.

A portfolio that has all three categories, with appropriate allocation and governance for each, gives the CFO both the rigour that expenditure governance requires and the flexibility that innovation requires.

Beyond individual investment ROI, CFOs should be tracking several portfolio-level AI metrics that give them a picture of whether AI investment is working at the programme level.

AI investment as a percentage of total technology spend: is AI investment proportionate to its strategic priority? Most FTSE 250 CFOs who have not yet defined this ratio are surprised to find that AI spend is either much larger or much smaller than they expected.

AI return on investment by category: are fix-and-operate investments meeting their hurdle rates? Are grow investments showing trajectory consistent with their strategic case? Are explore investments generating learning that informs subsequent investment decisions?

AI value at risk: what is the exposure from AI investments that are underperforming relative to expectations? This metric forces an honest assessment of which AI programmes are not working and creates the discipline to redirect or discontinue them rather than continuing to fund underperformance.