To calculate ROI of AI automation you need two formulas: annual saving minus annual cost divided by the build cost, and build cost divided by the saving per month. Three of the four inputs can be looked up. The fourth, the time the task takes today, you have to measure yourself.
This is for an owner holding a proposal and wondering whether the number at the bottom holds up. There is no percentage here and no average payback period. There are two formulas, the lines that belong in them, and the three numbers you count for two weeks first.
How do you calculate the ROI of an AI automation?
With two formulas that answer different questions. The first gives the return across a year. The second gives the month you break even.
Annual net saving = (S x 12) - R
First-year ROI = ((S x 12) - R) / B
Payback in months = B / (S - (R / 12))Where:
- S = saving per month, in euros
- R = recurring cost per year, in euros
- B = one-off build cost, in euros
S is not a percentage and not a feeling. S is how often the task occurs per month, times the time it takes now, times your fully loaded hourly cost. That last figure sits in your own payroll and nowhere else. Every benchmark you find online describes a different company than yours.
Note the second formula. The denominator is the saving per month minus the recurring cost per month, not the saving on its own. Leave R out and you are dividing by money that never nets out, which produces a payback period that is structurally too short. The larger the platform, model and maintenance lines are relative to the saving, the further those two answers drift apart.
Which numbers do you need before you can calculate?
Three, and you probably do not have them yet. That is not a failing: almost nobody logs how long recurring work takes. So count it for two weeks before you calculate anything.
- Frequency. How often the task occurs, per week or per month. Count it, do not estimate it.
- Duration per run. How long it really takes, including the looking up, the waiting and the double checking. Not the time it would take if everything were correct first time.
- Error rate. How often it goes wrong and what the rework costs. This number is usually zero in a proposal and rarely zero in practice.
- Fully loaded hourly cost. Salary plus employer costs for whoever does the work, divided by hours worked. From your own records.
- Who does it. The same task done by a junior and by an account manager produces a different S.
This list is why this article differs from the rest of the results. Every page that offers an ROI formula fills it in with a percentage that comes from nowhere. The formula is the easy part. The inputs are the work, and without them the output is an estimate of an estimate.
Two weeks of tallying is enough. One row in a spreadsheet each time the task comes round, with the date and the minutes. That is not research, but it is yours.
What belongs on the cost side of the calculation?
Four lines, exactly one of which has a public price list. The full breakdown sits in how much AI automation costs. Here it only matters which variable each line lands in.
| Line | Variable | Where the amount comes from |
|---|---|---|
| One-off build cost | B | agreement with whoever builds it, for your process |
| Platform cost per month | R | depends on the choice between cloud and self-hosted [S-006] |
| Model cost per use | R | published price list per million tokens [S-040] |
| Maintenance after handover | R | hours from your own team, at your own hourly cost |
The second row has no amount. That is a choice, not an omission: the pricing pages of the automation platforms could not be recorded as a source, and an amount without a source does not belong here. What the choice itself means is documented: cloud needs no installation, self-hosted means you manage your own deployment [S-006].
Model cost, the one line with a published price list
This line you can calculate to the cent. The price per million tokens is published per model: Haiku 4.5 costs 1 dollar per million input tokens and 5 dollars per million output tokens, Sonnet 5 costs 2 and 10 dollars, Opus 5 costs 5 and 25 dollars [S-040]. One token is roughly four characters or 0.75 English words [S-045].
The worked example the vendor publishes: 10,000 customer conversations averaging about 3,700 tokens come to roughly 37 dollars in total on Haiku 4.5 [S-043]. That is the whole line, not the price per conversation. Two discounts push it down further. Work that does not need an immediate answer can run through the Batch API at a 50 percent discount on both input and output tokens [S-041], and prompt caching costs 1.25 times the base price to write with a five-minute cache while a cache hit costs 0.1 times the base price, so it pays for itself after a single hit [S-042]. One line pushes the other way: the web search tool costs 10 dollars per 1,000 searches on top of standard token costs [S-044].
Those amounts are in dollars because that is how they are published. Convert them when they enter your budget, and record the rate you used.
Maintenance, the line that never makes the spreadsheet
Maintenance is work, so maintenance sits in R. A running automation needs an error workflow that picks up the details of a failed workflow and acts on them [S-031], and version history so a change can be rolled back to an earlier version [S-033]. Both exist, and both have to be set up and reviewed by a person. Who that person is and what it costs per month is covered in who owns an automation after launch.
Running the numbers without invented percentages
This is what the calculation looks like when you fill it in honestly. The letters stay letters until you have your own numbers.
Step 1 S = runs per month x duration in hours x fully loaded hourly cost
Step 2 R = platform cost per year + model cost per year + maintenance hours x hourly cost
Step 3 Annual net saving = (S x 12) - R
Step 4 Payback in months = B / (S - (R / 12))One of those four steps you can finish today. In step 2 the model cost comes straight from the price list: at a volume in the order of 10,000 conversations a year on a light model you are looking at tens of dollars, not hundreds [S-040, S-043]. That is often the smallest line in the entire calculation and the one that gets the most airtime in meetings.
Where the answer actually turns is step 1. Double the duration per run and the payback period halves. Leave the error rate out and the whole business case shifts. Which is why an honest calculation starts with counting, not with arithmetic.
What a high ROI on paper does not tell you
A good answer is not a decision to build. The formula does not know whether the task is stable, and that is the question sitting above it. If the process arrives differently every month, if the exception changes more often than the rule, or if the judgement is the task rather than the execution, the formula will produce a saving you never collect.
That is not theory. Work that resists a fixed shape generates change requests every month after handover, and those change requests sit in R. Where the floor is, and the three conditions around it, is covered in when automation is not worth it. Read that first if your answer looks too good.
The reverse happens too. A task with a mediocre ROI that irritates three people every week and regularly goes wrong is sometimes the better first workflow automation: the system starts running, the team starts trusting it, and the next calculation uses real numbers instead of assumptions.
Frequently asked questions
What does the model itself cost per month? Less than most people assume. The published example of 10,000 conversations averaging about 3,700 tokens comes to roughly 37 dollars in total on a light model [S-043]. The full breakdown per line sits in how much AI automation costs, including the discounts that apply.
How often does a task have to repeat before automating pays off? That depends on S and B in the formula above, not on a rule of thumb. The lower the build cost and the higher the frequency, the shorter the payback period. The floor itself, and the three conditions a task has to meet, is covered in when automation is not worth it.
What are the hidden costs in this calculation? Maintenance and rework. An error workflow that responds when a workflow fails [S-031] and version history to roll a change back [S-033] are both somebody’s work. Count those hours in R, or your payback period comes out flattering and wrong.
Where does a WeAdapt engagement start? A Quickscan starts from 995 euros and runs 1 to 2 weeks [A-006]. It quantifies the opportunities using your numbers, so the frequency and duration you counted yourself. From discovery call to live is four weeks [A-007].
Want the calculation with your own frequency, duration and hourly cost? Request an audit or book a call. What an engagement looks like and what falls under consultancy is on the service page.