AI ROI for SMEs: how to calculate it properly

AI ROI for SMEs: how to calculate it properly

AI ROI for an SME comes down to net gains over full project costs: (annual gains minus annual costs) divided by annual costs, times 100. The formula is not the hard part. The hard part is measuring your starting point, pricing the time you actually free up, and counting every cost, including the ones that never show up on an invoice.

Most articles on this topic hand you the formula, then a list of impressive numbers pulled from vendor studies. That is not what SME owners are missing. What they are missing is a method that holds up in front of an accountant or a board. That is what this article covers, as a companion to our complete guide to AI in business for SMEs.

The formula is simple, everything depends on what you put in it

AI ROI fits on one line: ROI = (gains - costs) / costs x 100. A project costing 12,000 euros over twelve months and producing 21,000 euros in gains shows a 75 % ROI. Nothing mysterious.

The trouble starts when you fill in the two terms. Gains are almost always estimated from a feeling ("we save a good two hours a week"), and costs are almost always understated because only the software subscription gets counted. The result: ROI figures announced at 300 % that collapse after three questions.

Two guardrails before you calculate anything:

• Set a clear reference period, typically a rolling twelve months. ROI without a horizon means nothing.
• Express gains in euros, not hours. An hour of accounting work and an hour of sales work are not worth the same.

Without a measurement beforehand, you have an impression, not an ROI

You cannot calculate a gain if you never timed the task before AI arrived. This is the most common mistake, and the most expensive one in terms of credibility.

The approach that works in an SME has three steps and takes two weeks:

1. Pick one specific, recurring task: writing up meeting notes, first-level replies to customer emails, preparing quotes, extracting data from supplier invoices.
2. For two weeks, ask the people involved to log the actual time spent each day on that task only. A shared spreadsheet is enough. Change nothing else during this period.
3. Roll out the tool, then measure again over two comparable weeks (same workload, not a holiday week).

This baseline has a second benefit: it tells you whether the task was worth automating at all. A task that feels time-consuming sometimes turns out to take forty minutes a week. In that case, the best ROI is not running the project.

Public figures confirm the discipline is missing. According to France's national statistics institute (Insee Première no. 2120, July 2026), 18 % of companies using AI report no specific purpose for that use, and the share rises to 28 % among firms using a single technology, most often generative AI. No stated purpose means no denominator, and therefore no measurable ROI.

Time saved only becomes money when it is reallocated

This is the point almost nobody makes, and it is the one that sinks most calculations. If four people each save thirty minutes a day, you have not recovered 0.4 of a full-time equivalent. You have four people finishing a task slightly earlier, and the freed-up time evaporates unless someone decides what to do with it.

Time saved turns into euros in three cases only:

• It replaces a hire or an outside supplier. Example: not renewing an outsourced copywriting contract at 800 euros a month. The gain is immediate and visible in the accounts.
• It is reallocated to revenue-generating work. Example: a salesperson going from six to nine meetings a week. The gain is then measured in margin, not hours.
• It absorbs growth without a new hire. Example: handling 30 % more orders with the same headcount. The gain is the cost of the role you did not create.

If none of these three applies, be honest in your calculation: the gain is comfort and quality of working life, not money. That is a perfectly good reason to continue, but it is not an ROI.

A prudent convention used by many finance teams: only value 50 to 70 % of the measured time saved, to account for that dilution. An ROI calculated with that discount will survive a conversation with your bank or your board.

Count every cost, including the ones with no invoice

The denominator is nearly always wrong because only licences go into it. Here are the items to include over twelve months.

Two items deserve particular attention.

First, the owner's time. In a company under twenty employees, the owner is usually the one testing, comparing and configuring. Twenty hours at a fully loaded cost of 80 euros comes to 1,600 euros that appear nowhere but weigh heavily on a 12,000 euro project.

Second, compliance. Since 2 February 2025, Article 4 of the European AI regulation (Regulation EU 2024/1689) requires organisations deploying AI systems to ensure a sufficient level of AI literacy among the people using them. This is not an optional budget line you can leave out of the calculation: it is an obligation and part of the project cost. We cover it in Article 4 of the AI Act and the AI training obligation and in our AI Act compliance guide for SMEs.

Three worked examples, from the cleanest to the most debatable

An example beats an abstract method. The three cases below illustrate the mechanics; they are not promises of a result.

Case 1, a gain that replaces an existing expense. An eight-person consultancy outsourced its meeting write-ups to a supplier for 650 euros a month. It switches to a transcription and summarisation tool, with internal review.

• Costs: 45 euros a month subscription for 4 users (2,160 euros a year), 12 hours of scoping at 80 euros (960 euros), half a day of training (600 euros). Total: 3,720 euros.
• Gains: 7,800 euros of supplier spend removed, minus 2 hours a week of internal review at 45 euros (4,680 euros). Net gain: 3,120 euros.
• ROI: (3,120 - 3,720) / 3,720 = -16 %.

The project loses money slightly in year one. It turns profitable in year two, once scoping and training costs disappear: costs drop to 2,160 euros for the same 3,120 euro net gain, an ROI of 44 %. That is exactly the nuance a twelve-month-only calculation destroys.

Case 2, a gain reallocated to sales. A thirty-five-person manufacturer automates quote preparation. Pricing was still manual, based on a 400-item catalogue.

• Costs over the year: 4,800 euros of tooling and integration, 3,200 euros of sales team training, 2,400 euros of internal time. Total: 10,400 euros.
• Gains: 6 hours a week freed across three salespeople, or 810 hours a year. With a 60 % prudence discount, 486 hours reallocated to prospecting. Those hours produced 9 extra meetings a month, of which 1.5 converted, at an average margin of 1,100 euros per deal. Gain: 19,800 euros.
• ROI: (19,800 - 10,400) / 10,400 = 90 %.

Case 3, a comfort gain. A twelve-person nonprofit uses generative AI to draft activity reports and social posts. The time saved is real, roughly 3 hours a week spread across four people, but no expense is removed and no additional revenue is generated. The financial ROI is zero. Continuing is still the right call, but it should be owned as an investment in working conditions, not dressed up as profitability.

The indicators to track alongside ROI

ROI is an annual snapshot. To steer, you need monthly indicators. Four are enough in an SME.

• Real usage rate: number of people using the tool at least once a week, over the number of licences paid for. Below 40 %, the ROI calculation stops meaning anything; you are paying for empty seats.
• Processing time per unit: minutes per quote, per write-up, per ticket. This is your baseline, still alive.
• Rework rate: share of AI outputs that need manual correction. Past 30 %, the net gain collapses.
• Time to break-even: how many months until cumulative gains cover cumulative costs. For an owner, this is often more meaningful than a percentage.

These indicators double as a measure of AI maturity. An organisation that can answer all four is ready for a second use case. One that tracks none of them should consolidate before expanding.

Five mistakes that make an ROI calculation useless

Most ROI figures presented to a board fall into one of these five traps.

1. Counting theoretical rather than measured time saved. An expert estimate is worth nothing to a finance director.
2. Forgetting rework costs. An AI output needing ten minutes of correction out of the fifteen minutes saved is not a fifteen-minute gain.
3. Adding up gains on overlapping tasks. The same quarter of an hour counted twice across two use cases inflates the result artificially.
4. Leaving training and compliance out of the denominator. These are project costs, not overhead. Excluding them skews the result and puts you in breach of the AI Act.
5. Measuring too early. The first four to six weeks are a learning period during which productivity often dips. An ROI measured in that window understates the project as much as one measured after six months of enthusiasm overstates it.

On that last point, public data offers a useful marker. Insee reports that 53 % of companies already using AI say they are held back by a lack of expertise, and 33 % by excessive costs. Missing internal skills, not tool pricing, is the main factor degrading observed ROI.

What official figures say about adoption in France

Two public sources let you position your company without relying on vendor studies.

Insee (Insee Première no. 2120, July 2026) finds that 18 % of French companies with 10 or more employees reported using at least one AI technology in 2025, against 10 % in 2024 and 6 % in 2023. The gap by size is stark: 15 % for companies with 10 to 49 employees, 31 % for 50 to 249, and 58 % above 250. Among users, 73 % cite cost optimisation (time savings, automation of repetitive tasks) as a motive, and 64 % improved process quality and reliability.

The France Num 2025 barometer, published by the French Directorate General for Enterprise on a sample of 11,021 companies including 7,878 micro-businesses, puts AI use at 26 % of small and medium-sized businesses, double the previous year, with wide sector gaps (41 % in the digital sector against 9 % in agriculture).

Two practical takeaways. First, if you are an SME under 50 employees, you are not behind: most of your peers have no organised use yet. Second, the dominant stated motive is time savings, which brings us straight back to the central question of this article: is that time reallocated, or simply absorbed?

Frequently asked questions

What ROI should an SME expect from an AI project? There is no reliable range that applies to every SME, and you should be wary of articles that give one. ROI depends on the hourly cost of the people involved, the volume of the automated task and, above all, on whether the time saved is reallocated. The same tool can produce 150 % ROI in one company and zero in another.

How long before measuring ROI? Expect a first serious measurement at three months, for break-even, then a full calculation at twelve months. Before three months, you are mostly measuring the learning curve.

Should training be counted in costs or separately? In project costs. Training is not overhead: without it, the time saving never materialises, and since February 2025 Article 4 of Regulation EU 2024/1689 requires a sufficient level of AI literacy among the people using these systems. Our article on funding AI training through your OPCO explains how to reduce this line.

How do I value an hour of work saved? Use the fully loaded employer cost, not gross salary: loaded salary divided by annual hours worked. For an SME, this often falls between 35 and 90 euros an hour depending on the role. Then apply a 30 to 50 % discount to account for time that is not reallocated.

Does a negative ROI in year one mean the project failed? No, provided you have isolated the one-off costs (scoping, integration, initial training). Recalculate ROI on a full year excluding those. If it is still negative, the use case was poorly chosen.

How do I prove ROI to my accountant or my bank? With three things: the before-and-after measurement, documented and dated; the exhaustive cost list with supporting evidence; and the concrete counterpart of the time saved (supplier spend removed, role not created, additional revenue). A sense of improvement will not do.

Do I need an audit before starting? Not always, but a short AI audit helps identify high-volume, low-value tasks, which are the ones where an ROI calculation is most likely to come out positive. It is also a chance to map your existing AI systems, which the AI Act will require of you anyway.

How GrowthPerf helps SMEs on this

We step in at the two moments where ROI is decided.

Before the project, we help choose the right use case and build the baseline: which tasks to measure, for how long, with which indicators. It is short work, a few hours, but it is what makes the calculation defensible later.

During rollout, we train the teams. This is where the gap is decided between a tool used by two people out of ten and a tool that is genuinely adopted. Our training is Qualiopi-certified, therefore fundable through your French OPCO, and it covers the AI literacy obligation set out in Article 4 of Regulation EU 2024/1689. See our AI training for business and our operational AI training.

To place this calculation inside a broader approach, from strategy to rollout, our complete guide to AI in business for SMEs walks through the steps in order.

If you want to know within thirty minutes which use cases are most likely to produce a positive ROI in your company, book a free audit. You leave with two or three quantifiable leads and the method to measure them, whether you then work with us or not.