The ROI of automating a process with AI comes from a simple formula: what it gives back minus what it costs, divided by what it costs. The hard part is the two numbers inside it. The benefit comes from the hours you free up, the enquiries you stop losing, the errors you no longer redo and the days sooner you get paid. The cost includes the set-up, your time, the learning curve, supervision and maintenance. There is no average ROI for SMEs: your figure comes from counting your process for a week before you automate anything.
This article is about judgement: what to count, which costs get forgotten and how to spot an invented number. If you just want to run the numbers with your own figures, the calculator is in the guide work out the return on your AI project.
A scene shows why it pays to do the sum yourself. The salesperson opens the laptop before sitting down and shows you the slide. A big percentage, with a decimal so it looks measured, and underneath the line: this is what you will get back in the first year. They haven't yet asked what you do for a living.
That number isn't badly calculated. It isn't calculated at all. It can't be: to know what automating something gives back, you need to know what you do today, how many times a day you do it and what it costs you. None of that is on their slide, because they never asked you.
So this article won't give you your ROI. It will give you something more useful: a way to calculate it yourself, with your own data, and the judgement to tell when someone is making it up.
Why there is no "average ROI of automation for SMEs"
The word SME stretches too far. It covers a company of 180 employees with a finance department and a self-employed person who invoices from the van. An average between those two describes neither of them.
And in Spain the split is nowhere near the middle. According to the National Statistics Institute's Central Business Register (INE), on 1 January 2025, 54.4% of Spanish businesses had no employees at all, and adding those with one or two, 81.6% had two or fewer. In more than half the businesses in this country there is nobody to delegate to.
An "average ROI for SMEs" mixes all that into one number and hands it back as if it were yours. It speaks for a company that doesn't exist. Yours has a specific volume, a specific process and a specific bottleneck, and those three are what count.
The formula, the only thing you don't need to invent
The calculation itself is primary-school arithmetic, and it has never changed:
ROI = ((Benefit obtained − Cost of the investment) / Cost of the investment) × 100
You read it like this: take what you got back, subtract what it cost you, and compare the result with what it cost you. If something gives back exactly what you paid, the ROI is zero: you are where you started. If it gives back double what you paid, the ROI is 100%.
The formula has no opinion about your business. It can't. It's a division. All the real work is in the two numbers you put inside, and that is where nobody helps you, because it's the part that takes effort. That is why the market sells you the finished result: it's easier to show you a percentage than to sit down and count your hours.
The benefit side: what to measure and how to measure it properly
Gut feeling won't do here. Your memory inflates what annoys you and hides what you have already got used to. You have to count.
1. The work that is done by hand today
Take one process. Just one, the one that repeats most. For a week, note down three things each time it happens: how many times it happens, how long each one lasts and who does it. No rounding up and no rounding down.
At the end of the week you have a number of hours. Multiply it by what that hour costs you: if you have employees, the real cost of that person's hour; if it's you, what you charge for an hour of your work. That number is yours and nobody will know it better than you.
And a warning that almost never appears in these calculations: freeing up two hours is only worth something if those two hours go into something else you can name. If the gap just fills up with more of the same, you have bought convenience, which is fine, but don't count it as a return.
2. What you lose by not answering in time
This can be counted too, even if it doesn't seem so. Open the call log on the business phone and look at last week: how many missed calls there are, how many you returned, and of the ones you didn't return, how many were customers who never called again. Do the same with the messages that arrived outside opening hours.
You don't need a study to know what that is worth. You need your own count and what an average customer brings you. If you have never done that count, do it before asking anyone for a quote: it is the figure that changes the decision most, and it's free.
3. The errors that have to be redone
Count the times last month that something had to be done twice. A corrective invoice, a wrongly typed figure that turned up three weeks later, two appointments at the same time, an order that went wrong. Each one has a cost to redo, and some have a worse cost that you won't see in any spreadsheet: the customer's face.
4. Getting paid sooner
This is the one most often forgotten and often the biggest. Getting paid sooner isn't a cosmetic improvement: it's money that is already yours reaching your account earlier.
And here there is a figure with a name and a source. According to the CEPYME Late Payment Observatory, in its second-half 2025 edition, the average payment period in Spain across 2025 was 80.5 days. The legal maximum set by Law 3/2004 on combating late payment is 60 days. In this country people pay late as a rule, and an invoice that goes out on the 1st instead of the 20th starts counting that period nineteen days earlier.
Notice what I have just done, because it is exactly what this article asks you to demand: that 80.5 comes from CEPYME, it's the 2025 average, and the link is right next to it so you can check. And even so, it isn't your number. Your number is in your accounts: how long you take to get paid. The CEPYME figure tells you the problem exists and is widespread, but it doesn't belong in your formula.
How much it costs to automate a process: what almost nobody adds up
We are not going to give you a price list, because there isn't one: automating a process costs more or less depending on how many processes are involved, how many systems need connecting, how much volume you handle and how many exceptions the process has that nobody has written down. A flow that reads one type of document and drops it in a folder is nothing like one that touches your CRM, your invoicing and your accountant's software.
And besides the supplier's quote, there are costs that don't appear in it. If your cost calculation only has the monthly fee, your ROI is inflated and you already know it. This is what is missing:
- The set-up, counting your time. Not only what the supplier invoices. The meetings, giving access, explaining how your process really works (which is never how it's written down), checking what they deliver. Those are your hours and they are worth the same as any others.
- The learning curve. The first weeks go slower, not faster. You and your team are learning while you work. That dip always exists and has to go in the sum.
- Supervision. Someone has to watch what the system does, especially at the start. A system nobody checks isn't cheaper: it is cheaper until the day it makes a mistake that nobody sees.
- Maintenance. You change your card terminal, your supplier changes an integration, a new regulation comes in, you set up a service you didn't offer before. Each of those things touches the system. What isn't maintained degrades by itself.
None of these four appears on the salesperson's slide. All four get paid for.
Why the round figures going around are worth nothing
Let's be straight, and we can do it with first-hand knowledge.
Today we audited this entire blog. This article, the one you are reading, published six different ROI percentages in its previous version, exact annual savings figures, exact payback periods and a price table. None of those figures came from anywhere. It also cited a study by a well-known consultancy that, when we went to look for it, does not exist.
And there is a second part, which is the one that really matters. When we asked an AI search engine about us, it answered that we offered a "guaranteed" return. Nobody wrote that word. The model read the percentages on these pages, saw they were repeated, and promoted them from estimate to promise. That is how an invented figure turns into a commitment someone can hold you to.
From this comes the rule, and it is the only one you need: a figure without a source isn't data, it's decoration. It makes no difference whether it was written by an agency, an AI search engine or ourselves two months ago.
How to spot whoever is selling them to you
The most reliable sign isn't in the number. It's in the questions they asked before giving it to you. Anyone who is really going to calculate your return needs to know, at the very least:
- How many times a day the thing you want to automate happens.
- Who does it today and what they stop doing while they do it.
- Where the data comes from and which system it has to end up in.
- What happens when the system gets it wrong, because it will get it wrong sometimes.
- What you will look at in three months to know whether it worked.
If they give you a percentage without asking any of this, the percentage comes from a catalogue. And these four warning signs apply to any supplier, ourselves included:
- A number before a question.
- The word "guaranteed" applied to a business result. Nobody serious guarantees a percentage that depends on how you work.
- A study quoted with no link, no year and no name behind it. Ask for the link. If it exists, it arrives in ten seconds.
- The same percentage for every sector. A hair salon and an accountancy firm have nothing in common operationally. If the promised return is the same, neither has been measured.
What can be stated with a source
Take away the invented figures and what is left is true and checkable. It is less spectacular and it stands up.
This is no longer a minority thing, but it isn't the majority either. According to the INE survey on the use of ICT in companies, published in October 2025, 13.4% of Spanish companies with fewer than 10 employees used artificial intelligence technologies in the first quarter of 2025: 5.9 points more than the year before. That is exactly the size of most businesses in this country. It isn't even one in seven, and that is the interesting part: most of the people competing with you still don't use it.
And there is a real date in the official gazette (BOE), which is nobody's forecast. Royal Decree 238/2026 of 25 March, published on 31 March 2026, develops the compulsory e-invoicing system between businesses and professionals, with phased adaptation deadlines depending on company size. That doesn't tell you how much you will save. It tells you that the paperwork side of your business will have to be organised and digital anyway, and that whoever does it in good time will arrive without surprises. If that's your case, here we explain how it works to automate invoices with AI.
How to do the calculation, step by step
- Choose one process. One. The one that repeats most, not the one that bothers you most. If you can't decide between several, choose your first AI project with the impact and effort method.
- Count it for a week. Frequency, duration and who does it. Written down, not from memory.
- Put your number on the benefit side. Hours × cost of the hour, plus whatever you can count for unanswered enquiries, redone work and days to get paid.
- Add up the full cost. Fee, set-up with your time included, learning curve, supervision and maintenance.
- Apply the formula. It's a division. It will take you a minute. If you would rather have a calculator do it, with the months it takes to recover the investment and the three-year return, use the one in the guide to the return on an AI project.
- Decide the threshold before looking at the result. Which figure would make you say yes. If you decide after seeing the result, you will talk yourself into it.
- Measure again after three months, with the same sum you did at the start.
And the point everybody skips: if you don't measure before, you won't be able to measure after. The "before" only exists once, and once you have automated there is no way to get it back. That one-week count is the only thing that separates knowing it worked from believing it worked.
Frequently asked questions
How much exactly does a Spanish SME save by automating with AI?
There is no single figure, and anyone who gives you one without knowing your business is making it up. The saving from automating with AI depends on how many times a day the thing you automate repeats, how long it takes each time, who does it and what is done with the time that is freed up. A business that gets six enquiries a day and another that gets sixty have nothing in common, even if both are "an SME". To find yours, count for a week the process that repeats most: times, minutes each time and the hourly cost of whoever does it. That sum is your maximum possible saving; the real one usually comes in lower, because there is always something left to check.
What is the average ROI of AI automation in Spain?
We don't publish one, and be wary of anyone who does. An average calculated over companies with 1 to 200 employees in every sector doesn't describe any actual company. In Spain, moreover, 81.6% of companies have two employees or fewer, according to the INE, so an average mixes businesses that have nothing in common. Your return depends on your volume and your process. The formula is the same for everyone: benefit minus cost, divided by cost, times one hundred. What changes are the numbers you put inside, and only you have those.
How long does it take to recover the investment?
It depends on the same variables as the ROI: how often the process repeats, what each freed-up hour is worth and what it really costs to set up and maintain the automation. A catalogue payback period is an invention dressed up as data. What you can calculate is your own period: divide what the project costs by the net saving you expect each month, with maintenance already deducted. What you can settle from day one is what you are going to measure to check it, and that conversation happens before signing anything. If you want to do it with a calculator, you have one in the guide to the return on an AI project.
How do I know if my business has processes that can be automated?
The sign is repetition, not size. If you or your team do the same thing several times a day, if you can describe that process in predictable steps and if the answer to "what if the data looks odd?" is a rule rather than a hunch, there is something to automate. If every case is different and depends on your judgement, probably not. The most common candidates in an SME are answering repeated enquiries, booking and reminding appointments, reading invoices and delivery notes and moving data from one program to another. What you shouldn't automate are the decisions that depend on your eye, such as a complicated quote or a complaint.
Is there a minimum turnover for it to pay off?
No. The deciding factor isn't how much you invoice, it's how much of what you do repeats. A small business with a lot of repetition has more room than a large one where every job is different. That is why the calculation is done on the process, not on the profit and loss account. What does exist is a volume below which it doesn't pay: if a task takes you ten minutes a week, automating it costs more than it gives back. With few repetitions, doing it by hand can still be the sensible option, and saying so is also part of a good calculation.
What guarantees are there that the ROI is real?
None in the form of a percentage. We don't guarantee a return, and nobody serious does: the result depends on things on your side that we don't control, starting with what you do with the time that is freed up. If someone guarantees a specific ROI without having measured your process, that guarantee is worth nothing. What we do is measure with you before starting, put in writing what will be looked at afterwards and look at it again after three months with the same sum. That isn't a guarantee. It is the opposite of an invented number, and it is the only way to know whether it worked.
If you want your number
You can do the above on your own. It is written for that and you don't need to call us to apply it.
If you would rather we do it together, we sit down, count your processes, add up the full costs including your time, and a number comes out. It may be yes and it may be no: there are processes that aren't worth automating and saying so is also part of the job.
What we won't do is give you a percentage over the phone before asking what you do. Tell us which process is eating your week and we will calculate it with your data. You can also see what we work on or how we approach AI for the self-employed.
Xora IntelligenceAI consultancy and custom software in Los Barrios (Cádiz). We write about what we build for businesses in the province. If you have a question about your case, write to us.




