How to calculate AI ROI: a simple model that survives scrutiny

Most AI investments show no measurable return, and it is almost never because AI cannot help. MIT put the failure rate at 95%, usually because nobody defined the return before starting. A defensible AI ROI is simple: the time or cost saved on a task, times how often it happens, minus what the tool and its adoption cost, measured against a baseline you took first.

If your team is spending on AI and someone has asked whether it is paying off, this post is for you. We will show why most AI returns are never measured, give you a one-line model you can defend to a finance lead, walk a worked example with real numbers you can swap for your own, and separate the returns you can bank from the ones you cannot.

The evidence that this matters is stark. In MIT NANDA's The GenAI Divide: State of AI in Business 2025, a July 2025 study drawing on 300-plus AI initiatives, 52 interviews and 153 senior leaders, "95% of organizations are getting zero return", while "just 5% of integrated AI pilots are extracting millions in value". The gap is not the technology. It is that most firms never define, or measure, what "return" was supposed to mean.

Why most AI ROI is never measured

The common pattern is to buy a tool, feel faster, and never check. Everyone agrees AI is helping, no one can say by how much, and when budgets tighten the spend cannot defend itself because there is nothing to point at.

The reason is almost always a missing baseline. You cannot show time saved if you never recorded how long the task took before. By the time anyone asks for the return, the "before" is gone, and any number you produce is a guess dressed up as a result.

The fix is not a complicated model. It is deciding, before you start, what you will measure and writing down where you are starting from. Do that and the maths is easy. Skip it and no formula will save you.

The baseline-first ROI model

Here is the whole model in one line. We call it the baseline-first ROI model because the baseline is the part everyone skips and the part that makes the number defensible.

AI ROI = (time or cost saved per task × how often the task happens) − (tool cost + adoption cost), measured against a baseline you took before you started.

Three things make this survive scrutiny in front of a finance lead. It is per-task, so you are measuring a real change in real work, not a vague "productivity" feeling. It nets off the full cost, including the messy human cost of getting people to adopt the tool, not just the licence. And it is anchored to a baseline you recorded first, so the "before" number is real rather than remembered.

One honest caution on the "time saved" figure. Savings are real but they vary, and they are usually largest for less experienced staff. In the NBER study Generative AI at Work, a 2023 analysis of 5,179 customer support staff, access to an AI assistant raised issues resolved per hour by 14% on average, and by around 34% for the newest and least experienced workers, while the most experienced saw little change. Measure your own saving on your own task; do not borrow someone else's percentage.

Set the baseline before you start

The baseline is one afternoon of work and it is the difference between a number you can defend and one you cannot. Pick the task, then record three things about how it works today.

1. Time per task. How long does one instance take now? Time it three or four times rather than guessing; people routinely misjudge this by half.
2. How often it happens. Per day, per week or per month, whichever is natural. Frequency is what turns a small per-task saving into a number that matters.
3. Quality now. A rough measure of how good the current output is, so an AI version that is faster but worse does not read as a win. Error rate, rework rate, or a simple "good enough" tally all work.

Write those three numbers down and date them. That single note is what every later ROI claim rests on.

A worked example

Here is the model run end to end on a fictional but realistic firm. Every figure is illustrative; swap in your own.

Meet Harbourline, an eight-person professional services firm. Their chosen task: drafting first-draft client proposals. Three people do this work and adopt the AI; the numbers below are the whole task across the team.

Read what the table is doing. The benefit is time saved on a frequent task, valued at what that time actually costs the firm. The costs include the awkward ones: not just the three licences, but the two hours a month someone spends keeping the prompts and reviews sharp, and the one-off cost of setting it up and training the team. Even after all of that, the task pays back in about three months and clears roughly £4,260 in year one. That is a case a finance lead can sign off, because every line is a number you could check.

Change one input and watch it move. If only 10 drafts happen a month, the yearly saving drops to £1,800 and the one-off setup swallows most of year one. Frequency is the lever. A big saving on something you do twice a month is a rounding error; a small saving on something you do fifty times a week is a real return. That is the same test behind choosing the right AI use case in the first place.

Hard returns and soft returns

Not every benefit belongs in the business case. The discipline that keeps your ROI honest is separating returns you can bank from returns you can only feel.

The rule is blunt: build the business case on hard returns only, and treat soft returns as the bonus that tips a close decision. If the numbers only work once you count "happier staff" or "we feel faster", the return is not there yet, and it is better to know that now than after a year of spend. Feeling faster is not the same as being more valuable, a distinction we dig into in AI productivity versus business value.

This is also why AI spend so often cannot defend itself later. The gains were real but soft, nobody wrote down the baseline, and the hard number was never there to bank. Getting the measurement right up front is the same discipline that separates the pilots that scale from the ones that quietly never make it past the trial, and it is what lets you answer plainly when someone asks whether your AI is paying for itself.

What to do on Monday

Pick one task your team does at least weekly and that eats real time. Time it three or four times this week and write down how long it takes, how often it happens, and how good the output is today. That single dated note is your baseline, and it is all you need to run the model once AI is in place.

Do that, and the return stops being a matter of opinion. If you would like help choosing the task most likely to pay back, and setting the baseline so the number holds up, book a call and we will work through it with you.

Common questions

What is a simple formula for AI ROI?

Time or cost saved per task, multiplied by how often the task happens, minus the tool cost and the cost of getting people to adopt it, measured against a baseline you recorded before you started. The baseline is the part that makes the figure defensible.

Why do most AI projects show no return?

Usually because no one defined or measured the return up front. MIT NANDA's 2025 research found 95% of organisations getting zero measurable return, largely a failure to redesign the work and to record a baseline, not a failure of the technology itself.

How do you set an AI baseline?

Before you adopt the tool, time the chosen task three or four times, note how often it happens, and record a rough measure of current quality. Date the note. Every later claim about time or money saved is measured against those numbers.

Should you count soft benefits like less stress or happier staff?

Name them, but do not build the business case on them. Base the case on hard returns you can measure, such as hours saved on a repeated task, and treat softer gains like morale or faster client replies as a bonus that tips a marginal decision.