What the ROI of an AI automation is (and why almost everyone gets it wrong)
The ROI of automating with AI is a subtraction, not a slogan: what you save minus what it costs, divided by what it costs. The problem isn't the formula — it's first-day-of-spreadsheet stuff — it's that most people only write down one side. They count the hours it frees and forget the model gets paid per token, that someone has to maintain the system, and that the setup wasn't free. An ROI that only looks at the savings isn't optimistic: it's miscalculated.
This guide is about the method: how to put a number on both sides without inventing figures. It doesn't bring a magic savings percentage — every process is its own — it brings the calculation skeleton so you plug in your own numbers and know, before you sign anything, whether that automation pays for itself or is an expensive toy. If you don't yet know which process to automate, that decision comes first and is solved by which processes to automate with AI; here we already assume you have a candidate and want to know if it's worth it.
The cost side: the three blocks almost everyone forgets
The cost of an AI automation isn't one figure, it's three blocks — and the one that hurts most is the one that doesn't show up in the initial quote. Before you calculate anything, add all three:
- Setup (one-off). Design, integration with your systems, testing and go-live. It's the number they give you in the proposal and the only one almost everyone looks at. It amortizes over the system's life, so in the annual calculation it only weighs a fraction.
- Model and infrastructure (recurring, variable). LLM tokens, hosting, the APIs you consume. It scales with volume: the more cases the system processes, the more it pays. Predictable per unit, but only if you measured the real volume, not the one you wish you had.
- Maintenance (recurring, the forgotten one). An AI system doesn't sit still: input formats change, prompts degrade, edge cases appear that need reviewing. Someone — yours or the partner's — spends hours every month supervising, iterating and fixing. This block is what separates the spreadsheet ROI from the real one, and what makes many pilots die in production even when the gross savings looked good.
The savings side: how to put a number on the hours you free up
The savings look like the easy part — "it takes forty hours a month off us" — and it's where the most inflation happens. For the number to hold, run it through three filters before it goes into the formula:
- Real hours, not photo hours. Measure how long the process takes by hand today, with a stopwatch and on real cases, not the optimistic estimate of whoever does it. It's almost always less than people say — and sometimes much more.
- Loaded cost per hour. Don't use gross salary: use the loaded company cost per hour (salary + contributions + overhead). That's the number that hour actually costs you, usually 30-40% higher than gross.
- Hours freed, not hours eliminated. The system doesn't do the process 100%: it leaves a share of cases that still need a person (review, exceptions, escalations). If you automate 80% of the volume, the savings are on that 80%, not the total. The remaining 20% still costs, and some of those cases cost more than before because they arrive pre-filtered and hard.
Annual savings come from multiplying: real hours freed per month × loaded cost per hour × 12. Without the three filters, that figure is a wish formatted as a euro.
The formula: payback and 12-month ROI
With both sides measured, the math is direct. Two numbers are enough to decide:
| Metric | How to calculate it | What it is for |
|---|---|---|
| Payback (months) | Setup ÷ (monthly savings − monthly recurring cost) | How many months to recover the initial investment |
| 12-month ROI | (Annual savings − total annual cost) ÷ total annual cost | The net first-year return, as a percentage |
| Year 2+ ROI | (Annual savings − annual recurring) ÷ annual recurring | Whether the system holds once setup is amortized |
A practical guide to reading the numbers: a payback under 6-9 months on a stable process is a good sign; over 18 months, either the process is huge or something doesn't add up. And if year-two ROI isn't comfortably positive, don't automate: you're paying to have a system that doesn't pay off, with the excuse of the first launch's savings.
What to measure to know if it actually works
The spreadsheet ROI is a hypothesis; it works or it doesn't once the system has been in production for months. To know, measure four things from day one — if you don't measure them, your ROI is a belief:
- Automation rate. What share of cases the system resolves without touching a person. It's the engine of the savings: if it drops, ROI drops with it.
- Escalation and error rate. How many cases it sends to a human and how many it gets wrong. A system that automates a lot but errs generates rework that eats the savings.
- Hours actually freed. Not the ones estimated at signing: the ones the team no longer spends. Ask the team, not the proposal.
- Real recurring cost. What the model and maintenance actually bill each month. It usually drifts from the budget — upward as volume grows.
Calculation errors that inflate the ROI on the sheet and sink it in production
- Forgetting maintenance. The most common and most expensive error. An ROI without the maintenance line isn't optimistic, it's incomplete.
- Counting 100% of the process as automated. No serious system reaches 100%. Count only the share it truly resolves on its own.
- Using gross salary instead of loaded cost. It undervalues the savings on one side, but usually comes paired with overestimating the hours on the other. Honest numbers in both directions.
- Putting in savings that never materialize. "We freed 40 hours" is only savings if those hours get reinvested into something that creates value or are genuinely cut. If the team just works more relaxed, that's wellbeing, not ROI — legitimate, but don't put it in the financial formula.
- Calculating a year for a system you'll rebuild in three months. If the process is about to change, the ROI evaporates with the change. Stabilize first, automate later.
The formula is easy; the hard part is plugging in numbers that don't fool you. If you'd rather not build the sheet alone, operations automation does exactly this calculation with you: we measure the real hours, put in the full cost of the three blocks and tell you whether that process pays for itself — and if it doesn't, we tell you that too. We don't sell automation for the sake of it; we sell the kind that's worth it and leave it running.