AI Implementation ROI Calculator

Why AI ROI Calculations Miss the Point

Most AI return-on-investment sums stop at the subscription price. A tool costs a small monthly fee, it saves an afternoon of freelance work, so the arithmetic looks trivially favourable. Real deployment carries four cost layers, and only two of them appear on an invoice: selection and setup, the recurring licence, the time spent on training and workflow redesign, and the cost of work that quietly breaks when the output is subtly wrong and nobody catches it. This calculator models the first two against the value of hours recovered. The third and fourth need judgement that no spreadsheet supplies.

The Time-Saving Multiplier

Everything here rests on one input, and it is the input nobody can pin down: hours saved per week. The costs are known, invoiced numbers. The benefit is an estimate, and the gap between plausible estimates is enormous.

Published evidence gives some shape to that estimate, and the striking thing is how far apart task-level and job-level findings sit. In a controlled experiment with an AI coding assistant, developers asked to build an HTTP server finished 55.8% faster than the control group, a single scoped task with a clear right answer. In a field study of 5,179 customer support agents, access to a generative AI assistant raised issues resolved per hour by roughly 14% on average, with about 34% for the newest and least experienced staff and close to nothing for the most experienced. Both numbers are real. They measure different things: one a task, the other a job that contains that task alongside everything else a working day involves.

Realistic Hours-Saved Benchmarks

That spread is why a single run of this calculator says very little on its own. Entering a figure derived from a demo of a narrow task, then applying it to whole roles, is how these projections go wrong. A figure measured on the job, across a mix of tasks and a mix of experience levels, behaves very differently from one measured on a benchmark exercise, and the research above suggests the difference can be a factor of three or four rather than a rounding error.

Two further patterns come out of the same studies. Gains concentrated among less experienced staff mean a team of specialists may see far less than a mixed team of the same size. And a task that is genuinely well suited, with a clear right answer and low review burden, behaves nothing like one requiring context or judgement, so a single blended hours figure across a whole department averages away the thing being measured.

The Hourly Rate Gotcha

The hourly rate carries a similar trap, in the opposite direction. Payroll cost is not what an hour of staff time costs an employer. Fully-loaded cost adds employer contributions, equipment, workspace, management overhead and non-billable time, and the multiple over base pay varies by country, sector and how the employer accounts for overhead. Entering take-home pay understates the benefit; entering a billable client rate overstates it, because that rate contains margin the employer never spends. The same distinction applies to outsourced work, where the landed cost includes the oversight time somebody spends managing the supplier.

Worked Example

Take a marketing team of four adopting an AI writing tool. Setup, integration and training come to 5,000 one-off. The licence runs 75 per user per month across four users, which is 3,600 a year. The team estimates 25 hours saved per week in total, fully-loaded cost is 85 an hour, and the horizon is three years.

Annual hours saved come to 1,200, worth 102,000 at 85 an hour. Take off the 3,600 licence and the annual benefit is 98,400. Over three years that is 295,200, less the 5,000 setup, giving a net benefit of 290,200 and an ROI of 5,804%.

That ROI figure deserves suspicion rather than celebration, and not because the hours estimate is wild. Twenty-five hours across four people is roughly 16% of a 37.5-hour team week, which sits between the two research findings above rather than beyond them. The percentage is enormous because labour is expensive and software is cheap: 1,200 hours of staff time is worth twenty-eight times what the tool costs to run for a year.

Which means the ROI percentage is nearly useless as a decision input here. Work out where this deployment breaks even and the reason becomes obvious: over three years, at these costs and this hourly rate, the arithmetic turns positive at about 1.3 hours saved per week. Halve the hours estimate to 12.5 and the three-year net is still 137,200. Cut it to 5 hours and it is 45,400, an ROI of 908%. On numbers like these the question is not whether the percentage clears a threshold. It is whether the hours are real at all, and whether the freed time goes anywhere useful.

When ROI Goes Negative

The result turns negative in a handful of recognisable situations. An enterprise-scale implementation cost sits against the savings of a two-person team. The hours estimate came from a vendor demonstration rather than a measurement. Pricing scales with usage, so the ongoing cost climbs with the very activity that generates the benefit. Output needs enough human review that the net time saved is a fraction of the gross. Or the workflow redesign and retraining absorbed months that nobody counted as a cost.

There is a quieter failure mode that the arithmetic cannot detect at all. A calculation can come out positive while the organisation has no capacity to realise it. Twenty-five hours a week freed across a team is only worth 98,400 a year if those hours move to work of comparable value. Where the time refills with low-value activity, or with more meetings, the tool works exactly as advertised and the money never appears. The model counts hours recovered; it cannot see what happens to them next.

Broader labour-market analysis makes the same point at a different scale: the International Labour Organization's global assessment of generative AI concludes that its main effect is likely to augment jobs rather than automate them, which is why the benefit lands as reclaimed hours inside a role rather than as headcount removed from it. Whether reclaimed hours turn into value is an organisational question, not a financial one.