AI ROI

The formula is the same as for classic ROI: (benefit − cost) / cost. Cost covers rollout, models, data access, support, and review and rework; that last item is missing from many calculations, even though every AI result has to be checked before it goes out. The benefit side usually shows more output per working hour, but it only counts as a return once costs go down or additional revenue comes in.

Quality and lower error risk are real effects, but they belong in a separate column because they can only be put into money terms through assumptions. Without review, rework and a fixed period, AI ROI remains an estimate. An investment decision also needs the payback period: the time it takes for the cumulative benefit to cover the cost.

The formula and a worked example.

The math is the same as for classic return on investment: AI ROI = (benefit − cost) / cost. The result is a ratio, usually given as a percentage. A positive value means the benefit exceeds the cost over the period in question. A negative value means the initiative cost more than it delivered in that period.

A typical example: an initiative costs a mid five-figure sum over three years and delivers one and a half times that in financially effective benefit over the same period, for example because less outside work has to be bought in. AI ROI is then 50 percent, because the surplus equals half the cost.

The payback period complements this figure. If, in the example, all costs fall at the start and the benefit is spread evenly over the three years, the investment pays for itself after two years. ROI says how much an initiative returns; the payback period says when the money is back. Both belong in a decision memo, along with the period and the assumptions behind them.

Getting to numbers that hold up.

The formula is simple; the inputs are the hard part. A reliable calculation therefore follows a fixed path. First, the use case is narrowly defined: one specific task, one team, a clear start and end to the workflow. Then the effort before the rollout is measured, including review and corrections. After the rollout, it is measured again with the same method so both values are comparable.

Next, all costs over the period are added up, including the ongoing ones. Finally, financially effective benefit and quality gains are listed side by side but separately, each with the assumptions behind it: how many employees actually use the tool? How much review does each result need? How will model prices develop?

When assumptions are uncertain, a range works better than a single figure, meaning one cautious and one optimistic calculation. That shows how strongly the result depends on individual assumptions. After a few months in operation, it is worth checking which assumptions held and which did not.

What belongs on the cost side.

With AI, the cost side has five items. Rollout covers setup, connecting to existing workflows and training. Many providers bill model costs by units of processed text, called tokens, separately for input and output. These costs therefore rise and fall with usage: the more text a system passes to the model, such as documents from a knowledge base, the more each call costs. Data access means connecting and maintaining the sources, and support means ongoing operations with monitoring, bug fixing and keeping up with new model versions.

The fifth item is review and rework: the time people spend checking and correcting AI results. Review and rework are missing from many calculations, even though every AI result has to be checked before it goes out. This item in particular can eat up part of the time saved.

The cost side becomes reliable when calls are recorded individually with their cost, for example through a central layer between the applications and the models. Without that kind of booking, model costs disappear into a lump-sum invoice and the cost side remains an estimate.

When benefit becomes return.

The benefit side usually shows more output per working hour: a team produces more reports, proposals or texts in the same time. It only counts as a return once costs actually go down or additional revenue comes in. Freed-up time is not yet a payout. If the hour gained goes into additional orders or into work that used to be bought in from outside, it turns into money. If it disappears unnoticed into the workday, nothing shows up in the calculation.

There are also effects that are real but hard to put a dollar figure on, above all better quality and lower error risk. They belong in a separate column. Adding them to cost savings and revenue in a single total makes the calculation easy to challenge, because every dollar in it rests on an assumption.

For leadership, this creates a planning task. Before the rollout, it should be settled what the time gained will be used for: more orders, less outside work or tasks that have been piling up. Only then can the benefit be proven later.

Common calculation errors.

The most common error is measuring a single work step. If drafting a report gets faster but review takes longer, less is left than the faster draft suggests. So the whole workflow is measured, from the request to the approved result.

A second error is the missing baseline. Without a comparison value from before the rollout, every number is just a claim. A third is extrapolating pilot results. Pilots are often run by especially interested employees on hand-picked cases; in daily work the mix looks different. Counting licenses instead of usage is just as misleading: paid seats that hardly get used sit fully on the cost side and deliver no benefit.

Finally, costs are often assigned to the wrong place. The foundation, meaning users, permissions, access and training, serves many initiatives at once. Charging all of it to the first use case makes that one look bad. Leaving it out entirely makes all the following ones look too good. The clean approach is a disclosed split or a separate line item for the foundation, carried across several initiatives.

Where the term stops working

Freed-up time is not yet a payout; if the hour gained goes unused, nothing shows up in the calculation. AI ROI assesses individual, clearly defined initiatives. It does not measure the overall value of an AI transformation, because the foundation, rules and training cannot be attributed to any single initiative.