Automation ROI: how do you calculate it? and give the board a payback date

01

What is the ROI formula for automation?

The formula is simple, and that is its weakness. Anyone can drop unverified numbers into it. ROI = (value of the saving − cost to build and run) / cost to build and run × 100%. Use the same period on both sides, usually one year.

The second number, the one that actually matters in a board meeting, is payback: the one-off cost divided by the monthly saving less the monthly running cost. The result is a number of months, and a number of months converts into a date.

• Value of the saving = hours of manual work × hourly cost + cost of the errors removed + value of the process running faster
• Cost = workshop + build + licences + AI model usage + team time for testing and training + support
• Payback = one-off cost / (monthly saving − monthly running cost)

02

What does the manual work really cost? A worked example

The most common case from our workshops: three people move data between systems, two hours a day each. Nobody counts it, because it is „just a minute” and „while I am in there anyway”. Counted, it looks like this.

Errors and delays are not in here. Add one wrong invoice a month and the number grows, but keep it on its own row so the board can see what was measured and what was estimated.

03

What goes into the cost of building and running it?

This is where vendor calculations part company with reality. The quote shows the build price and the first annual invoice carries four more lines. Below are all of them we know about, with our own rates where we publish them openly.

• Payback in the example: 12,000 / (6,600 − 500) = 1.97, so inside the second month
• First-year ROI: (79,200 − 12,000 − 6,000) / 18,000 = 340%
• The second year carries only the running cost, so ROI rises to (79,200 − 6,000) / 6,000 = 1,220%

The worked example assumes EUR 12,000 one-off, including a one-week workshop at EUR 3,500, and EUR 500 a month for licences and models. That is the upper end of the usual range for a single integration.

04

Which automations pay back fastest?

The order below is what we assume when we quote, not market research. It is how we estimate the return when we tell a client where to start. The common thread: a repeatable task, data already sitting in a system, and a result you can count.

Follow-up and routing are harder to count, because the return depends on conversion rather than on hours. We quote those paybacks more cautiously and measure them at 90 days, not at 30.

05

How do you count soft benefits so the board believes them?

Less stress, cleaner data, a faster answer for the customer. These are real effects, and they are the most common reason a calculation loses credibility. Not because they do not exist, but because they get dropped into the formula as money with no method behind them.

One rule: a soft benefit enters the ROI only when it has a measure and a source. Otherwise it stays in a separate „benefits not counted” box, which a board reads with more trust than an inflated total.

• Errors: invoice corrections per month × time per correction × hourly cost, plus discounts lost to late payment
• Speed: the time cut between enquiry and quote, converted into conversion only if you have last year's CRM data
• Turnover: the recruitment cost for a role people leave out of boredom, if you already have that cost calculated
• Risk: a task only one person can do carries the cost of their holiday and their resignation; write it in words, not in money

06

What does the payback look like in practice, here?

The only proof we can show without asking a client for permission to publish their figures is what runs inside our founders' own companies. Two AI agents have been in production there since March 2026. We say that plainly, because proof you cannot check is not proof.

At the performance marketing agency (around 30 people) the margin in May, June and July 2026 was 5.55 percentage points higher than a year earlier, moving from minus 0.59 to plus 4.96 percent, with revenue up 1.5 percent. At the clothing brand with an online shop, return on ad spend went from 3.53 to 4.19 on a budget 6.9 percent lower, February to August 2026 against the same months of 2025.

What these numbers do NOT prove: over the same period the agency's client mix changed and the shop's marketplace channel grew. That is a coincidence in time, not proof of cause, and your own calculation after go-live deserves the same honesty. The full figures with the measurement method are on the AI automation agency page.

• The same months in both years, so seasonality is taken out
• Months not closed in the books are excluded from the comparison
• Caveats written next to the number, not hidden in a footnote

07

How do you measure the return after go-live, not just promise it?

A calculation made before the build is a hypothesis. The return exists only once you can read it, at 30 and 90 days, off the same metrics you wrote down before the start. Here is how we run it.

1. 01

   Baseline

   During the workshop we measure the process for a week: hours, number of operations, number of errors, time from event to result. That is the „before” number.

2. 02

   Metrics agreed before the start

   We write down three to five metrics and how each one is read out of a system. A metric that needs a person to produce it does not make the list.

3. 03

   Reading at 30 days

   The first reading, usually incomplete: the team is only just stopping doing things by hand „just in case”. We compare it with the baseline and fix whatever is out of line.

4. 04

   Report at 90 days

   The second reading on the same metrics. This is the number you take to the board, together with everything that changed in the meantime outside the automation.