The CFO as AI capital allocator

The capital allocation question, properly posed, is not whether to invest in AI. It is which AI investments, in which sequence, with what milestones, against what alternatives. That is the same question the CFO asks of every other capital request. The vocabulary is new. The discipline is not.

What follows is what changes — and what does not — when the capital being allocated funds an AI capability rather than a conventional one.

What capital allocation actually means for AI

The capital allocated to an AI deployment is not, in most cases, a single large commitment. It is a series of commitments — pilot, scaled pilot, production, refresh — each contingent on the previous one performing. The CFO who treats the AI investment as a single large capex line will fund the wrong thing, or will fund the right thing badly.

The right unit of analysis is the AI programme, not the AI tool. A programme has phases, gates, decision points, and a defined trajectory from problem statement to production capability. Each gate is a decision the CFO is entitled to make, with a defined set of inputs.

Milestone gating: the only investment structure that survives

An AI investment approved as a single large commitment — build the full thing, then see if it works — is a risk management failure. The technology moves too quickly for the assumptions in the initial business case to hold to completion. The use case clarifies in contact with reality. The cost surprises emerge in production.

The defensible structure is milestone-gated.

• Discovery. A short, low-cost phase that defines the problem, validates that AI is the right approach, and identifies the candidate solutions. The deliverable is a problem statement and an investment recommendation, not a working system.
• Proof of concept. A working system at small scale, with defined success criteria, run on a representative sample of the problem. The deliverable is performance evidence against the criteria.
• Pilot. The system run at scale on a defined subset of the actual workload, with the governance, the integration, and the cost emerging at production scale. The deliverable is a scaling decision.
• Production. Full deployment, with the operating cost, the monitoring, and the refresh discipline in place. The deliverable is a steady-state capability.

Each gate is a decision. The CFO's role is to design the gates, enforce them, and not allow the programme to skip ahead because the team running it is enthusiastic about the next phase.

Build, buy, partner: the analysis that must be done

The choice between building an AI capability in-house, buying it from a vendor, or partnering with a specialised provider is a strategic decision with long-term consequences for capability, cost, and differentiation.

The analysis runs along several dimensions.

• Strategic differentiation. Is the capability a genuine source of competitive advantage that should not be shared with vendors who serve competitors? If yes, the build case is stronger. If no, the buy case is stronger.
• Capability gap. Does the entity have the engineering talent to build to production quality, or would building require a hiring effort that takes longer than the capability is needed? Build cases that depend on a hiring miracle do not work.
• Vendor maturity. Is the vendor solution mature enough to deploy reliably? In moving markets, the vendor pitch deck and the vendor production reality differ. The diligence has to be genuine.
• Total cost. Build, buy, and partner have different cost curves. Build is high upfront and slow to ramp. Buy is faster but pays the vendor margin in perpetuity. Partner is somewhere in between, with the additional consideration of partner alignment.
• Exit cost. The cost of changing the decision later. Build locks in less; buy with deep integration locks in more.

The CFO must insist that this analysis is done rigorously for each material AI investment, not decided by whichever approach the technology team finds most comfortable. The instinct of an engineering team is to build; the instinct of a commercial team is to buy; neither is, by itself, a strategy.

The counterfactual test

For every AI investment, the correct comparator is not "what does this cost?" It is "what would we spend to achieve the same outcome without AI, and what is the gap if we cannot achieve it at all?"

Two cases emerge.

In the first case, the outcome is achievable conventionally — through additional headcount, traditional automation, or a vendor solution that does not depend on AI. The AI investment is then a productivity case. The decision turns on whether the AI route is cheaper, faster, or higher-quality. The discount rates and the comparison are conventional.

In the second case, the outcome is not achievable conventionally — because the data volume is too large, the speed requirement too demanding, or the pattern too complex for human analysis. The AI investment is then an enabler of new capability. The comparison is between having the capability and not having it. The decision turns on strategic value, not on productivity arithmetic.

The capital allocation framework must distinguish between these cases. A productivity case priced as a capability case overpays. A capability case priced as a productivity case underfunds.

Technology literacy without technology expertise

The CFO does not need to understand transformer architectures. She does need to understand:

• The difference between a model running on proprietary data within the entity's security perimeter and one that sends data to a third-party API. The cost, control, and compliance profiles differ.
• The difference between a fine-tuned model and a general-purpose one — when each is appropriate, and what each costs to maintain.
• Why a pilot that worked at small scale may not work at large scale — the assumptions on data quality, latency, and concurrency that hold at twenty users may break at twenty thousand.
• What "context window" means in cost terms, why retrieval-augmented generation is the workhorse pattern, and what evaluation looks like for an AI system.

This is informed executive judgement, not deep technical expertise. The CFO who has none of this literacy cedes authority to the technology team in a category where authority is increasingly consequential. The CFO who attempts to have all of it confuses her role with the CTO's. The middle is the right place.

The discipline of saying no, late

The hardest capital allocation discipline in AI is the willingness to stop a programme that has consumed significant investment and is not delivering. The sunk cost fallacy operates particularly strongly when the programme is technically interesting, when the team is enthusiastic, and when the alternative is uncomfortable.

The CFO's authority — and obligation — is to insist that the gate decisions are real. The programme that does not meet the proof-of-concept criteria does not advance to pilot. The pilot that does not scale does not advance to production. The production system that does not deliver against the success metrics in three quarters is reviewed for continuation. These decisions are not popular. They are the function the CFO performs that no one else in the organisation will.

This piece sits inside the CFO in AI framework. See also the full-stack cost of AI and gross-margin maths in the age of AI. Lorna writes from practice at IMPT. The verified page records what is and isn't published here.