AI automation ROI: how to calculate it before you buy
By Jacob Slot · Published
ROI on an AI automation is the annual gain divided by the total cost of building and running it. Before you have bought anything, both numbers are estimates, and you can make that estimate precise enough to decide on, if you collect four numbers first: how often the task runs, how long it takes, what an hour really costs you, and what the build and running costs are.
Three things decide an investment in AI automation: ROI, payback time and the risk that your estimate is wrong. You can work out the first two in a morning. The third is the one most vendors skip.
We hear the same sentence from new customers every month: "We think there are a lot of hours in it, but we can't put a number on it." As long as that is true, the decision is a gut feeling, and a gut feeling is easy for a vendor with a polished slide to talk down.
What AI automation ROI means before you have built anything
Most articles on the subject calculate backwards: they take a finished project and show what it ended up delivering. That is easy, and it does not help you, because you are holding a quote and have no measurements.
The formula is the same, but the inputs are different. ROI is the annual gain divided by the total cost in the first year. The gain has three parts: the time that disappears from an employee's calendar, the errors that no longer happen, and the capacity you gain without hiring.
People forget the third part, and it is often the biggest. When an order can be processed overnight instead of tomorrow morning, you are not moving hours, you are moving lead time. Ask what one extra day of delay costs you in complaints or lost orders.
Be honest about the cost too. It is the build, plus licences and usage, plus the time your own people put into the project, plus maintenance when one of the systems changes. If you only calculate on the front page of the quote, you get a number that disappoints in month four.
Then there is the difference between ROI and payback time. ROI is a percentage for one year. Payback time is the number of months before the gain has covered the expense. For a company with two to 50 employees, the second number matters most, because cash flow decides what you can do this year.
The four numbers you need before you can calculate
The chain is short. You are missing four measurements, and three of them you can get in a week.
1. How often, and how long each time
Count for one week. Not an estimate from a meeting, but a count by the person who does the work: how many times was the task done, and how long did three of those runs take on average.
A week is enough to catch the rhythm and short enough that someone will actually do it. If the task is monthly, take the last time it was done and ask whether that run was typical.
This is where most projects go wrong, long before they are built. Without a before-number, the later discussion about ROI becomes a matter of opinion.
2. Your own hourly cost, not an average
Take the actual salary of the person who does the task, add holiday pay, pension and ATP, and divide by the number of hours they are really available in a year. That number is yours, and it is the only one that holds up in an internal discussion.
Industry averages are convenient and dangerous. They make the calculation impossible to attack, because nobody recognises themselves in it, and so impossible to act on as well.
If several people do the task, weight by who does the most of it. If it is you or your operations manager keying it in, use that hourly rate. It is typically the most expensive in the building, and that is exactly why the task is worth removing.
3. What the errors cost
The time at the keyboard is the cheap part. A wrong account number, an order with the wrong item number or an address that sends the parcel back only costs you once it is discovered, and it is discovered late.
Put a price on one error: what does the clean-up cost in time, in freight, in credit notes and in the conversation with the customer that follows? Multiply by how many times it happened last quarter.
You do not need to be exact. A conservative number you can defend often moves the calculation more than the hours do.
4. The price of the build and of running it
Ask the vendor for three numbers: the one-time price, the ongoing price in DKK per month and how much time is expected from your own people. If the third is missing, the quote is not finished.
With us, a scoped build starts from $1,500 as a one-time payment, and an ongoing agreement with further development starts from $750/month. We have laid out the full ranges in our guide to what AI automation costs, so you can see where a task like yours usually lands.
Worked example: three processes side by side
Here are three tasks we meet in almost every Danish company, calculated with the same method. The numbers are an example, not a price list, and the assumptions are conservative:
- Working weeks per year: 46
- Hourly cost in DKK, including holiday pay and pension: 350
- Share of the task that can actually run on its own: 80%
| Process | Times per week | Minutes per run | Hours per week | Saved per year in DKK | Build plus first year of running costs in DKK | Payback in months |
|---|---|---|---|---|---|---|
| Invoices from email to bookkeeping | 40 | 9 | 6.0 | 96,600 | 38,000 | 5 |
| Weekly report from three systems | 1 | 150 | 2.5 | 40,250 | 32,000 | 10 |
| Checking new customer agreements | 2 | 20 | 0.7 | 11,270 | 30,000 | 32 |
Three tasks that feel equally annoying day to day, and that end up in three very different places. The invoice task pays for itself in just under half a year, because the frequency is high. The report is on the fence. The checking task is a no, and it is a no even though it is the one the employees complain about most.
That is the whole point of calculating up front: frequency beats irritation. A task that runs 40 times a week and takes nine minutes is gold. A task that runs twice a week and takes 20 minutes is not, however stupid it feels.
If you want to try it with your own numbers instead of our example, you can use our ROI calculator and get the same four fields set up.
When ROI is good enough, and when you should say no
A percentage is hard to relate to. Payback time is not. Here are the thresholds we use ourselves when we advise on a task.
| Payback | What it means | Decision |
|---|---|---|
| Under six months | The gain is big enough to absorb an estimate that turns out too optimistic | Build now |
| Six to twelve months | Only holds if the process is stable and not about to change | Build, if the systems stay as they are |
| Twelve to eighteen months | The hours alone do not carry the project | Only if the errors or the capacity count too |
| Over two years | The uncertainty in the estimate is bigger than the gain | Don't, and spend the money elsewhere |
The last row is the most important, and it is the reason we say no to tasks. A project with a two-year payback is not a bad project, it is a project where you cannot tell whether it was a good one until you have forgotten what you paid.
Be specific about what could overturn the estimate, too. If one of the systems is due to be replaced within a year, the calculation falls apart, whatever the table says. Then you wait, or you write the automation into the requirements for the new system.
The trap: saved time is not saved money
This is the mistake that most often turns a good-looking calculation into a disappointment. A salaried employee costs the same before and after she gets six hours a week back. The salary does not fall by itself.
The hours only become money in one of three ways: she spends them on work that creates revenue, you avoid the next hire, or you avoid overtime during peaks.
So the calculation needs an answer before you buy: what will the freed hours be used for, and who decides? Write it down with the numbers. It is the difference between a gain and a good feeling.
One comparison is worth having. A full-time hire costs 25,000+ DKK a month, and keeps costing that. An automation is built once, runs around the clock and does not hand in its notice. At Markeity they save 20,000 DKK a month on systems we built once.
How grants change the calculation
Almost none of the Danish articles on the subject include grants, which is odd, because it is the single factor that moves payback time the most.
If you apply to the SMV:Digital digitalisation pool, the state covers 50% of your spend on external advisory, and the grant is capped at 150,000 DKK (source). The pools open in rounds and close when they run out, so the date matters as much as the application.
Calculate two versions: one with the grant and one without. If the advisory part is halved, a task moves from the fence zone around one year to typically under six months. If you do not get the grant, the project still has to carry itself. We have collected the schemes, the amounts and the application rounds in our guide to grants for your AI project, and the programme's own page is at SMV:Digital.
From estimate to certainty
An estimate never becomes a measurement. But you can do two things that remove most of the risk before you sign.
The first is to measure for a week, as described above. It costs nothing and gives you the before-number that the whole later discussion rests on. The use of artificial intelligence in Danish companies with more than ten employees rose from 15% in 2023 to 28% in 2024 according to Statistics Denmark, but very few of them have a before-number. That is why so many projects end in disagreement about what they delivered.
The second is to get the result written into the agreement instead of onto a slide. With us, measurable ROI within 90 days is a contract condition: if we cannot show it, we keep working at no extra charge. Why we dare to do that, and how we scope tasks so that it holds, is described under our 90-day ROI guarantee, and the timeline from mapping to running is in our process.
Key numbers
15 hours
freed up per week for a typical customer
90 days
to measurable ROI, written into the contract
$1,500
starting price for a scoped build, one-time payment
The order is worth holding on to: measure, calculate, decide. The reverse order turns AI into an expense instead of an investment, however good the solution is. The method behind the build itself is in our guide to workflow automation.
FAQ about AI automation ROI
What is a realistic ROI on AI automation?
For a scoped, high-frequency task, payback time is typically between three and eight months, which corresponds to an ROI of several hundred percent over the first year. Low-frequency tasks land far worse, and there is no average ROI that can replace your own calculation.
Be sceptical of a vendor who quotes an ROI figure before asking about your frequency and your hourly cost. Then the number comes from another company's process.
How long before an automation pays for itself?
It depends almost entirely on frequency. A task that runs daily has typically paid for itself within half a year. A weekly task takes about twice as long, and a monthly task rarely pays for itself on time alone.
The build itself is not the long part. Expect about three weeks from mapping to a running flow.
Which processes give the fastest ROI?
The ones where the same information is moved many times a week between two systems that do not talk to each other: orders to bookkeeping, receipts to approval, leads to CRM, hours to payroll. They are fast because no decisions have to be made along the way.
Conversely, processes with many exceptions and few repetitions give the worst ROI, even when they are the most irritating. Start where the frequency is, not where the frustration is.
How do I calculate ROI when the automation prevents errors instead of saving time?
Put a price on one error and multiply by how often it happens. The price is made up of the clean-up time, direct costs such as freight or a credit note, and possibly the lost customer. Use a conservative number you can defend internally.
That kind of gain is harder to measure afterwards, because errors that do not happen are invisible. So agree in advance how you will count: typically the number of corrections in a period before and after.
Can I get a grant for the advisory and the calculation?
Yes, in most cases. The SMV:Digital advisory grant covers 50% of the cost of external advisory in digitalisation and AI projects, and the mapping can be part of that process. The pools open in rounds, so timing matters.
With us, the mapping itself is free as part of an AI Opportunity Audit, regardless of any grant. You keep the mapping and the calculation, even if you choose not to build anything.
If you need to calculate ROI on AI automation, start by counting for a week and finish with the payback time. That week decides whether you are buying on a number or on a feeling.