How to Build the Business Case for AI Spend

A business case for AI spend is a written argument connecting a proposed AI investment to a specific number in your profit and loss, with an honest payback estimate and a defined point at which you stop. Most businesses approving AI budgets have never written one, and the results show.

The RAND Corporation found that more than 80% of AI projects fail to deliver their intended business value, roughly twice the failure rate of comparable IT projects without AI (RAND, 2024). MIT's Project NANDA, drawing on interviews and surveys with senior leaders and an analysis of around 300 enterprise deployments, found 95% of organisations saw no measurable profit and loss impact from generative AI, with about 5% capturing real value (MIT Project NANDA, The GenAI Divide, 2025). The authors describe that finding as a directional snapshot rather than a definitive market verdict, which is fair, and the direction is still worth taking seriously.

A business case improves your odds by forcing three questions into the open: which number moves, by how much, and how long until the investment pays for itself. This guide covers how to answer all three, with the arithmetic worked through on a real-shaped example.

Why Most AI Business Cases Fall Apart

Most AI business cases fail because they start from the technology and reason backwards towards a justification.

The evidence points the same way. RAND attributes 84% of AI project failures to leadership decisions rather than technical limitations, with unclear success metrics the most common single cause (RAND, 2024). The models mostly work. The decisions around them mostly do not.

The pattern is recognisable. Someone sees a demo, the capability is genuinely impressive, and the question becomes where in the business this could be applied. That order of operations produces a solution hunting for a problem, and the case has to work hard to sound convincing because the benefit was reverse-engineered after the decision had already been made emotionally.

Vendor numbers make it worse. A vendor claiming a 40% efficiency gain is describing a best case from a reference customer with different data, different volumes and a different team. Those numbers are real for that company. They tell you very little about yours, and quoting them in a board paper transfers someone else's outcome onto your balance sheet.

The third failure is scope. AI business cases often bundle several loosely related ideas into one approval so the total sounds strategic. When the project underdelivers, nobody can identify which part failed, because the case never separated them. Approving three small cases beats approving one large vague one.

The consequences are visible in the abandonment data. S&P Global Market Intelligence found 42% of companies abandoned most of their AI initiatives in 2025, up from 17% the year before (S&P Global Market Intelligence, 2025). A great deal of approved spend is quietly being written off.

Start From a Number You Already Track

The strongest business case for AI spend begins with a number that already appears in your reporting and has been annoying someone for at least a quarter.

That constraint does useful work. A metric you already track has history, an owner and a definition everyone agrees on, which means you can measure whether the investment moved it. A metric invented for the business case has none of those things, and it will be quietly redefined the moment the results disappoint.

Good candidates share a shape. They are operational, measured monthly or better, and expensive enough that a 10% improvement is worth someone's attention. Support tickets resolved per agent. Days sales outstanding. Time from lead to first response. Cost per unit of content produced. Manual reconciliation hours per close.

Weak candidates are the ones with no baseline. Customer experience, brand perception, employee satisfaction and innovation velocity all matter enormously, and none of them will settle an argument about whether $200,000 of AI spend paid for itself. Leave them out of the case and mention them as secondary benefits if they arrive.

The frustration test matters as much as the measurement test. A metric that has irritated its owner for a quarter comes with someone motivated to see it improve, which is what carries a project through the messy middle. A number nobody cares about produces a pilot nobody champions.

Once you have the number, write down its current value before anything starts. A business case with no baseline cannot be evaluated later, which is convenient for whoever proposed it and expensive for everyone else.

How to Estimate the Benefit When It Feels Soft

Soft benefits become estimable the moment you convert them into a unit of time or money that someone already owns.

The technique is decomposition. "Better customer support" resists measurement. "Forty hours a week of agent time spent on password resets and order status questions, at a fully loaded cost of $45 an hour" is a figure you can multiply, and it survives scrutiny because every input is checkable.

Build the estimate as a range and argue from the bottom of it. A case that only works at the optimistic end of its range is a bet dressed as an analysis. If the pessimistic case still clears your hurdle, you have something worth approving.

Subtract the costs everyone forgets. Licence fees are the visible part. The rest includes integration engineering, the time your team spends reviewing and correcting model outputs, the ongoing cost of monitoring quality, and the migration cost if the vendor reprices or disappears. AI carries heavier hidden costs than most software, because the output requires human judgement before anyone trusts it.

One structural finding is worth weighing here. MIT's research found that buying AI capability from specialised vendors reached successful deployment roughly twice as often as building internally, around 67% against 33% (MIT Project NANDA, 2025). Most businesses default to building. The evidence suggests the build option should carry a heavier discount for implementation risk than the buy option.

What Payback Period Is Reasonable for AI Spend

A reasonable payback period for AI spend is shorter than the technology's shelf life, which is currently the binding constraint.

This is what separates AI from most capital decisions. A warehouse upgrade earns for a decade. An AI implementation built on today's model landscape faces repricing, deprecation and replacement on a horizon measured in quarters. A three year payback assumes a stability this market has yet to demonstrate.

For most businesses, a target of six to twelve months on operational efficiency cases works well. It keeps the scope small enough to finish, it forces you to pick problems where the value is concrete, and it means that when the landscape shifts you have already recovered the investment.

Longer horizons can be justified where the investment builds something durable underneath the model layer. Gartner expects 60% of AI projects unsupported by AI-ready data to be abandoned through 2026, which is a useful way of reading the priority: clean, well structured, well governed data outlives whichever model you point at it (Gartner, 2025). Spend on data infrastructure earns over a longer period than spend on any particular tool.

The calculation gets contentious on projects promising revenue growth. Cost savings are easier to verify, because the baseline is known and the saving appears in a line someone already watches. Revenue claims depend on assumptions about customer behaviour that the AI itself does not control. Treat revenue-side cases with more scepticism and a shorter leash.

Write the Kill Criteria Before You Start

Kill criteria are the conditions under which you stop, agreed in advance and written into the business case.

They exist because of how AI pilots actually end. Very few get cancelled cleanly. They get extended, rescoped and quietly renamed until the original business case is unrecognisable and nobody can say whether it worked. Sunk cost does the rest, and each extension feels smaller than admitting the first decision was wrong.

Useful kill criteria are specific and dated. Using the worked example above: if deflection sits below 40% at week twelve, the payback no longer clears twelve months, so you stop and revert. That criterion falls directly out of the arithmetic, which is the advantage of having done the arithmetic.

Agreeing them early also improves the honesty of the original estimate. Proposing a benefit you will be measured against in twelve weeks produces more conservative forecasting than proposing one nobody will revisit.

A pilot killed on schedule is the process working. It cost a small defined amount and bought you certainty.

Who Should Own the AI Business Case

The AI business case belongs to finance, with technology supplying the feasibility assessment and the cost estimate.

The split matters because the two roles test different things. A fractional CTO tells you whether the thing can be built, what it costs to integrate, and where it will break. A fractional CFO tests whether it should be built at all: whether the number is real, whether the payback clears your cost of capital, and whether this beats the other three things competing for the same money.

RAND's finding that 84% of AI failures trace to leadership decisions rather than technology is the argument for putting genuine financial rigour on the approval side (RAND, 2024). The projects are rarely beaten by the models.

For businesses without a full-time CFO, this suits a fractional engagement well. The work runs intense for a few weeks, then becomes periodic. In Australia a fractional CFO runs $7,000 to $15,000 per month, against a full-time CFO at $215,000 to $235,000 in base salary (SEEK, 2026) before the 12% superannuation guarantee (ATO, from 1 July 2025) and other on-costs are added. The Australian cost page breaks the numbers down, and our guide to what a fractional CFO actually does covers the wider remit.

A business case for AI spend is easier to write with someone who has read a dozen of them. Fractionus accepts fewer than 3% of executive applicants and matches on the decision in front of you rather than a job title. Tell us what you are trying to approve and we will shortlist fractional CFOs within 2 to 5 days.

Frequently Asked Questions

How do you build a business case for AI spend?

Start from a number you already track and that already frustrates someone, record its baseline, estimate the benefit conservatively as a range, subtract every cost including integration and human review, set a payback target of six to twelve months, and write kill criteria before you begin. A business case for AI spend without a recorded baseline cannot be evaluated afterwards.

What percentage of AI projects actually fail?

The RAND Corporation found more than 80% of AI projects fail to deliver their intended business value, around twice the rate of comparable non-AI IT projects (RAND, 2024). MIT's Project NANDA found 95% of organisations saw no measurable profit and loss impact from generative AI, with roughly 5% capturing value (MIT Project NANDA, 2025). Definitions of failure differ between studies, and both point the same way.

What payback period should I expect from AI investment?

Six to twelve months suits most operational efficiency cases, because the technology landscape reprices and deprecates far faster than traditional capital investments. Longer horizons can be justified for data infrastructure, which outlives whichever model sits on top of it. A three year payback assumes a stability the AI market has yet to demonstrate.

How do I estimate the benefit of AI when it is hard to measure?

Decompose the soft benefit into time or money someone already owns. Calculate the agent hours currently spent on password resets and order status queries at their fully loaded hourly cost, which produces a figure you can multiply and defend. Build the estimate as a range and argue from the bottom of it rather than the vendor's number.

Should the CFO or the CTO own the AI business case?

Finance should own the business case, with technology supplying feasibility and cost inputs. A fractional CTO assesses whether the thing can be built and what integration costs. A fractional CFO assesses whether the number is real, whether payback clears the cost of capital, and whether this investment beats competing uses of the same money. RAND attributes 84% of AI failures to leadership decisions rather than technology.

What costs do businesses forget in AI business cases?

Licence fees are the visible part. The commonly missed costs are integration engineering, the ongoing human time spent reviewing and correcting model outputs, quality monitoring, and migration if a vendor reprices or shuts down. In a typical support automation case, licence and review time together can consume more than half of the conservative saving.

How do I stop an AI pilot that is not working?

Agree the kill criteria before the pilot starts and write them into the business case with a date attached, derived from the payback arithmetic. If the deflection rate your case depends on has not appeared by week twelve, you stop. Deciding in advance removes sunk cost from the conversation, which is what usually keeps failing pilots alive through repeated rescoping.

Can a fractional CFO help build an AI business case?

Yes, and the work suits the fractional model because it runs intense for a few weeks then becomes periodic. A fractional CFO who has evaluated several AI cases across different businesses spots the optimistic assumption faster than someone doing it for the first time. Australian retainers typically run $7,000 to $15,000 per month.