How Do You Measure the Return on an AI Investment?

In short: the return on an AI investment cannot be calculated without a baseline measured before the project starts. Measurement happens across four headings: time recovered, error rate reduced, capacity gained, and direct revenue impact. According to McKinsey's 2025 survey, 88% of organizations use AI in at least one function, yet only 39% can show any impact on earnings. The gap usually comes not from the technology but from the absence of measurement discipline.

Why most organizations cannot show a return

Saying "our work got easier" after an AI project is easy. Turning that into a number is impossible in most organizations, because the state that came before was never measured.

McKinsey's data shows the gap clearly: adoption has reached 88%, but the share able to demonstrate an impact on earnings stays at 39%. In the same survey, only 21% of organizations using generative AI have redesigned a single workflow around it. When the workflow does not change, no measurable difference appears either.

That is why the return calculation is set up at the start of a project, not at the end.

Baseline: what to measure before the project

The baseline is a numerical snapshot of the process as it stands today. It is collected over at least two weeks, using real transaction volume. The data to gather is:

• How many minutes one transaction takes, as an average and as a worst case
• Monthly transaction count
• How many people touch the process and how many hours each one spends
• Error or rework rate, meaning the percentage of work that has to be redone
• Waiting time, meaning the time work sits on a desk before moving to the next step

Any return calculation made without these five items is an estimate.

The four metric families

1. Time recovered. The difference in minutes per transaction multiplied by monthly volume. Before multiplying recovered hours by a labour rate, one question has to be answered: where did that hour go? If the freed hour was not redirected to new work, the saving on the sheet is not real.

2. Error rate reduced. Rework, late delivery and correction cost all belong here. This is usually a larger line than time saved, and a less visible one.

3. Capacity gained. The increase in transactions the same team can handle. In growing organizations this is where the return shows most concretely: volume absorbed without hiring.

4. Direct revenue impact. Outcomes that touch sales, such as shorter quote preparation, more inquiries answered, or fewer lost opportunities. It is the hardest heading to measure and the one that carries the most weight in the boardroom.

The 30-60-90 day schedule

Measuring once gives a misleading result, because the learning curve pulls the numbers down in the first weeks. This schedule is more reliable:

Time	What is measured	What to expect
Day 30	Adoption rate, open defect records	Time savings are not visible yet; the goal is that the system is genuinely used
Day 60	Duration, error rate, transaction count	First measurable difference; compared against the baseline
Day 90	All four metric families	A stable result; payback period is calculated from this data

A result taken before day 90 should not be used as the basis for stopping or scaling a project.

What belongs on the cost side

The denominator of the return calculation is often written short. The full list is:

• Setup cost, meaning consulting and development
• Model and infrastructure usage cost, monthly
• Licence and integration expenses
• Hours the internal team spends on the project, at a labour rate
• Maintenance and improvement after go-live

The last two items are usually skipped, which makes the payback period look shorter than it is. This is the most common reason projects miss expectations in their second year.

Where measurement goes wrong

Counting a freed hour as a saving. If nobody left and the freed hour was not redirected to new work, no cash equivalent appears. In that case the gain belongs under capacity, not under savings.

Multiplying a pilot result across the whole organization. Pilots usually run with the most willing team and the cleanest data. Assuming the same ratio in other departments builds the budget on the wrong number.

Presenting one big figure. A sentence like "two million in annual savings" loses trust precisely because it cannot be questioned. Showing the four metrics separately, each with its source, is more convincing.

Frequently asked questions

How long before a return becomes visible? In process automation, the first measurable difference usually appears between 60 and 90 days. Payback on the investment ranges from 6 to 18 months depending on the volume of the process.

What happens if we start without a baseline? Even if the project works, you cannot prove it did. That makes the next budget harder to get. A baseline can be reconstructed roughly after the fact, but its reliability is low.

Do small teams need measurement too? Yes, but it can be kept simpler. One process, one duration metric and a monthly transaction count are enough for most small teams.

Who should do the measuring? The business unit that owns the process measures; the consultant sets up the method. When only the outside party measures, nobody inside owns the result.

Sources

• McKinsey & Company, The State of AI: Agents, innovation, and transformation (2025) — 88% of organizations use AI in at least one function, only 39% can show an impact on earnings; 21% of generative AI users have redesigned a workflow.
• Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (June 2025) — unclear business value ranks among the leading reasons for cancellation.

Closing

The return on an AI investment is not something you look for after the project ends; it is a measurement discipline set up before it starts. If the baseline was taken, the calculation is simple. If it was not, no method makes up for it.

See our enterprise AI consulting service or look at the research data behind our approach.