The AI Investment Calculator: The Math Every AI Budget Needs

Photo via Unsplash

The exchange goes the same way in almost every board meeting where an AI budget is on the agenda. The CFO asks what the return is. The vendor answers with a word like transformation. The budget passes anyway, because AI is on the strategic priority list, and the firm now has a programme with no target.

MIT's State of AI in Business 2025 found that about 95 percent of enterprise AI pilots produced no measurable profit. Gartner puts at least 30 percent of generative AI projects as abandoned after proof of concept. Those figures get argued over, definitions of failure differ between studies, and I would not lean the whole case on either of them. What is not in dispute is the direction: the ordinary outcome of an AI pilot is that nothing measurable comes out of it. Something that happens to most projects is not a risk to manage in a footnote. It is the base rate, and a budget that does not account for it is priced wrong.

What the calculator does differently

A normal business case starts with a quote and works forwards, guessing at the benefit. That is the wrong direction, because the quote is the one number the vendor controls and the benefit is the one you have to defend.

This calculator inverts it. You state the annual value you are trying to produce, split three ways into cost savings, new revenue and retention value, because those three behave differently and mixing them hides the weakest one. Then you give it your context: whether you are buying off-the-shelf, configuring and integrating, or building custom; your sector; how many people are affected; the most sensitive data the system will touch; how ready your organization actually is for the change; and the horizon you are planning against.

What comes back is the first-year investment that value target implies, in three buckets.

Direct AI cost. Licences, models, data work, integration. This is the part a vendor quote covers, and in most budgets it is the whole budget.

Change management. Training, process redesign, communication, and the productivity dip while people adjust to working differently. This is the bucket that gets left out, and it is not a rounding error. BCG's 10-20-70 finding is that roughly 10 percent of the effort in a successful AI deployment goes to algorithms, 20 percent to technology and data, and 70 percent to people and process. If technology is most of your line items, you have not sized the thing that decides the outcome.

AI risk provision. Governance, security, oversight and rework, scaled to your sector and how sensitive the data is. IBM's Cost of a Data Breach research puts around 670,000 US dollars in additional cost where shadow AI is involved in a breach. A provision is cheaper than that conversation.

Alongside the buckets you get the yearly run cost, time to value, payback period, three-year net and return, with the cash flow modelled month by month so value starts at go-live rather than on day one of the project.

What it is for, and what it is not for

It is not for producing the exact return figure. Nobody has that, including the vendor whose slide says four months. It is for producing an honest range with the assumptions visible, so the board conversation moves from marketing language to numbers people can disagree with specifically.

If your calculator says 18 months and the vendor's slide says four, that gap is the most productive discussion available to you that quarter, because one of you is wrong about something concrete and it can be settled. If your calculator says 40 months, you probably need a different vendor or a smaller purchase. Either outcome beats approving the spend and finding out in year two.

Where I would not trust it

The ratios inside this calculator come from research on large organizations. A twelve-person firm in Kingston is not the population BCG derived 10-20-70 from, and I do not know how well it holds at that scale. My honest read is that the change management share is directionally right and probably too low for a small firm rather than too high, because a small firm has no slack to absorb a productivity dip, but I cannot prove that and the calculator will give you a confident number regardless. It also cannot tell you whether your value target is realistic, and in my experience that is the assumption that most often turns out to be wrong. The calculator will happily size an investment against a savings figure that was never achievable. Garbage in, professionally formatted garbage out.

So use the output as an argument rather than an answer. The number it produces is not the point. The list of assumptions it forces into the open is the point, and that list is useful even if you throw the total away.

Who should run it

Whoever is answerable for the outcome. In firms with an AI Officer, that person, against every AI purchase over a threshold you set. In firms without one, the CFO's team. In firms with neither, the sponsor of the initiative should run it and send the output to the CFO before the budget conversation rather than during it, which changes the meeting from a pitch into a review.

It takes under an hour, and most of that hour goes on arguing about the value target, which is where it should go.

If the payback number surprises you in either direction, the surprise is worth more than the number. That is the assumption you did not know you were carrying.

"The vendor prices the software. The overrun comes from everything the software makes people do differently. One of those is on the quote and the other is not." - Adrian Dunkley, AI Boss