Return on investment is the most quoted and least measured argument in AI projects: it gets promised on a slide and never checked. Yet an honest calculation fits on one page. It needs a measurement taken before AI, every cost, and a safety margin. Here is the four-step method, followed by a full example.

Step 1: measure the current time

On the target process, before AI arrives:

  • Volume: how many tasks per week (reminders, quotes, meeting notes, replies).
  • Time per task: in minutes, timed over a representative week, not estimated from memory.
  • Fully loaded hourly cost: salary, social charges and workplace costs expressed per hour, not just gross salary.

This day-0 measurement is your point of comparison. Without it, the calculation rests on a base that can be moved after the fact. The 30-day pilot builds it in from the scoping stage.

Step 2: count the net gain

The gain is not "the task disappears". It is "the task goes from so many minutes to so many minutes", measured during the pilot:

  • Time afterwards, including review and corrections. If AI saves ten minutes but four are spent correcting, the gain is six.
  • Actual use. Count the tasks actually handled with AI, not the potential. If eight reminders out of fifteen go through the tool, count eight.

Step 3: add up every cost

Over the same period:

  1. Licences for active users, at the vendor's public price.
  2. Set-up: the AI Workshop, CHF 1,900 excluding VAT for 4 hours.
  3. Follow-up: AI Adoption Support, for example the Essential plan at CHF 1,800 a month for 10 hours, with a three-month minimum.
  4. Running costs if there is automation: AI model usage, hosting, maintenance.

Step 4: calculate and decide

Net benefit = net hours saved × fully loaded hourly cost − full cost, over the same period.

ROI (%) = net benefit ÷ full cost × 100.

Redo the calculation with a cautious assumption: fewer active users, more corrections. If the result stays positive, the use holds up. If it is only positive under the favourable assumption, it is a bet, and should be treated as one. If it is negative in both cases, you do not sign.

A full example over three months

Calculation assumptions, to be replaced by your own measurements: an SME of thirty people, eight of whom use AI every week for quotes, letters and meeting notes. During the pilot, these eight people each saved an average of an hour and a half net per week, review included, making 12 hours a week for the team. Fully loaded hourly cost: CHF 80.

Costs over three months

  • AI Workshop: CHF 1,900
  • AI Adoption Support, Essential plan, three months: CHF 5,400
  • Licences: 8 users × CHF 30 a month (assumption) × 3 months = CHF 720
  • Total: CHF 8,020

Gain over three months (13 weeks)

  • 12 hours × 13 weeks = 156 hours
  • 156 hours × CHF 80 = CHF 12,480

Result

  • Net benefit: CHF 12,480 − CHF 8,020 = CHF 4,460
  • ROI: 4,460 ÷ 8,020 × 100 = about 56%
  • Break-even point: CHF 8,020 ÷ CHF 80 = about 100 hours over the quarter, or roughly one hour per week per active user.

If, under the cautious assumption, only five people save one hour a week, the gain drops to 65 hours and the quarter runs at a loss: that is exactly the information you need before you sign. After the three months, if the team carries on by itself, licences are the only fixed cost left, and the same gain becomes clearly profitable.

The three mistakes that inflate an ROI

  1. Counting freed-up time as money. Hours freed up are capacity. They become a saving if they avoid an expense or absorb a real workload.
  2. Forgetting the review. A gain measured without the correction time is a projection, not a measurement.
  3. Calculating on potential. Not everyone in the team will use the tool on every task. Calculate on observed use.

What ROI does not tell you

ROI tells you whether a use pays off, not whether it is desirable. A use with a modest return but no risk may be worth keeping because it improves quality; a highly profitable use that exposes protected data may be worth turning down. The decision combines the calculation with the gain / effort / risk grid from our guide on finding the three right use cases. For a breakdown of budget items, see what AI costs a Swiss SME.