Agree what an AI investment should return before you fund it.

Why the return is hard to show

The return on AI spending is hard to show. Researchers working with MIT Project NANDA reported in 2025 that 95% of the organizations they studied were getting zero return. The figure counts organizations, and the report calls its findings preliminary and describes some of its figures as self-reported.

The report’s own method shows why measurement is the difficulty. It defined success as deployment beyond the pilot stage with measurable KPIs, and it measured the effect six months after the pilot. Its authors say that window may be too short for complex systems, that success measures differ across organizations, and that other operational changes at the same time complicate any reading of ROI.

The same report suggests that what is easy to measure draws the money. Executives asked to divide a hypothetical budget put most of it into sales and marketing, and the authors attribute this to outcomes there being easier to measure. One procurement leader they quote asks how to put a number on a tool that makes a team faster.

Agree the measure first

A leader can answer that question before the money is committed. Decide what the investment should change, which existing report will show the change, who answers for it and what the figure is today. Without today’s figure there is nothing to compare the result against.

Public audit and appraisal bodies ask for the same discipline. The UK National Audit Office says performance indicators should be tracked, including baseline measures set at the outset. It noted in 2024 that these were not yet in place for the government’s AI adoption strategy.

HM Treasury’s guidance for UK public spending, The Green Book, asks for an evaluation plan before a proposal goes for final approval. The cost of monitoring belongs in the case. Both are UK public-sector guidance, and a private organization can borrow the method without being bound by it.

Our view is that the expected return and its assumptions should be written down before any money is committed. National statistics will not supply the figure either. Statistics Canada reported that 19.2% of businesses used AI in the 12 months before its 2026 survey, and the release reports adoption without any measure of return.

Count the whole cost

The purchase price is one part of what an AI system costs. The Green Book says costs and benefits should be calculated over the lifetime of a proposal, and that maintenance costs must be reflected alongside capital and operating costs. It leaves out sunk costs, which have already been incurred and cannot be changed.

Running costs continue after launch. The UK government’s AI Playbook says AI systems need ongoing investment to maintain them, update them with new data and keep performance stable. That is guidance for public bodies, and it does not say how costs change with use.

The Green Book also asks for a deliberate correction for optimism. Appraisals raise estimated costs and timeframes and lower estimated benefits, because it finds a systematic tendency to be over-optimistic about both.

It describes this for public projects generally and did not test it on AI. We would apply the same correction to any AI case, and the case should say which adjustment was made.

Keep the benefits separate

A business case usually claims three different kinds of benefit. Savings are money no longer spent. Freed capacity is staff time that becomes available.

Reduced risk is a smaller chance of something going wrong, or a smaller loss if it does. These differ in kind, so we keep them as three figures and never add them into one.

The distinction matters because freed time is not the same as saved money. In the NANDA report, the best-performing organizations reported savings mainly from reduced outside spend, such as agency fees and outsourcing contracts. The tools sped up the work without changing team structures or budgets.

Our view is that freed capacity becomes a saving only when something else changes. The team is staffed differently, or it takes on work it could not before. The case should name which of these it expects and who decides.

Write a case that can say stop

A case that can only approve is a request for funding. In our view a sound one can also recommend stopping, and it states the condition that would change its answer; none of the sources above sets this as a rule. The Green Book offers two useful anchors.

It keeps business as usual as the benchmark that every shortlisted option is compared with, and says this does not mean doing nothing. Current arrangements have their own costs and risks. It also asks for switching values, the point at which a key variable would make an option stop being good value.

Put the recommendation in one sentence near the front, and show the arithmetic and its assumptions below it. A reader who disagrees can then see exactly which assumption to challenge.

What this means for you

Before the next AI investment goes to a decision, take four steps.

• Name the measure and the report where it will appear.
• Record today’s figure.
• List every cost over the system’s expected life.
• Show savings, freed capacity and reduced risk as separate lines.

None of this promises a return, and a case built this way may well recommend against going ahead. That is a useful result when it is reached before the money is spent.

Our work on return and cost describes how we help leaders build this kind of case. A briefing is the place to start.