What's a good ROI or payback period for an AI project in an SME?

There's no universal threshold, but a useful rule for small companies is to favour projects that pay back within about 6 to 12 months, because anything longer ties up scarce cash and exposes you to the technology or your own processes changing before you've recouped the cost. Payback period is often more decision-useful than a percentage ROI for an SME, since it speaks directly to cash flow. Treat a sub-12-month payback as promising, and be cautious about multi-year payback stories that depend on optimistic assumptions.

Should I use an employee's salary or something else to value time saved?

Use the fully loaded (fully burdened) cost, not the bare salary. The loaded cost adds employer taxes, benefits, software, equipment, and overhead, and is commonly 1.25 to 1.4 times base pay, sometimes more. Someone on a €60,000 salary typically costs €75,000–€84,000 a year to employ. Using the loaded figure is the honest number: it makes genuine savings look as good as they really are and stops you overstating marginal ones.

Why can't I just multiply hours saved by an hourly rate?

Because that ignores two things that decide whether the project actually pays. First, scattered saved minutes usually get absorbed into the working day rather than turning into cash, unless you can convert them — by not backfilling a role, cutting outsourced spend, or taking on measurably more work. Second, it ignores the costs on the other side: software, integration, the recurring AI usage bill, and ongoing maintenance. Hours saved is the starting input, not the finished answer.

How much should I budget for ongoing AI maintenance?

A practical planning figure is roughly 15–20% of the original build effort every year, covering keeping prompts and knowledge bases current, monitoring output quality, fixing breakages, and absorbing vendor price changes. This is recurring forever, not a one-off, which is exactly why leaving it out is the most common reason a project that looked profitable becomes a slow loss. Always model a full year of running costs against a full year of benefit.

What costs do people most often forget in an AI ROI calculation?

Three stand out. Integration — connecting the tool to your real, messy CRM, spreadsheets, and email and cleaning the data so a machine can read it, which is frequently the single largest hidden cost. Ongoing maintenance — the recurring annual upkeep described above. And internal time — the hours your own team spends building, training, and reviewing, which feel free but are not. The software sticker price is usually the smallest line on the list.

How do I estimate the AI usage bill before I build anything?

Model it per transaction, not as a flat fee, because it scales with how often the automation runs. For a typical small-business automation handling a couple of hundred tasks a day, the language-model usage commonly lands somewhere around $10–$60 a month depending on the model, and can be under $25 on a budget model. Estimate your expected volume, pick a model tier, and multiply — then remember a successful, popular automation will cost more as usage grows. Caching repeated context can reduce this bill substantially.

Does keeping a human reviewing the output ruin the ROI?

No, but it changes the maths and you must account for it. If a person still reviews and approves the output — which is the right pattern for anything customer-facing, financial, or contractual — then the AI is saving the drafting and fetching time, not the reviewing time. Calculate the saving on the part you actually removed, not the whole task. Over-claiming here is how a headline '90% time saving' turns out to be 40% in practice; modelled honestly, human-in-the-loop automations still pay off, just less dramatically.

What if my ROI calculation says no — is the AI project a waste?

Not at all; a clear 'no' on paper is the calculation doing its job and saving you money. Often the issue isn't AI itself but that the target process is bloated and should be simplified before any tool touches it, or that the saved time doesn't convert to cash for this particular use case. Either fix the process and recalculate, or pick a different, higher-impact process — the back-office and operational tasks tend to return more than flashy sales-and-marketing tools. Finding the 'no' before you spend is far cheaper than discovering it afterwards.

AI ROI for SMEs: How to Calculate the Return

How do you calculate the ROI of an AI project before you spend the budget?

You calculate AI ROI before spending by building a simple before-and-after model on one specific process: estimate the annual value the project creates (mostly hours of work saved, valued at the fully loaded cost of the people doing it, not their bare salary), then subtract the total annual cost of getting and keeping it running (build, software, integration, and — the part everyone forgets — ongoing maintenance). The formula is ROI = (annual benefit − annual cost) ÷ annual cost, and the more honest sibling metric is the payback period: total upfront cost ÷ monthly net benefit. Do this on paper, for a single high-impact process, before you sign anything. If the numbers only work when you ignore maintenance and assume every saved hour turns straight into cash, the project isn’t ready — and you’ve just spared yourself the most expensive lesson there is.

That is the whole method in one paragraph. The rest of this guide shows you exactly which costs belong in the model, how to put a defensible number on “time saved” without fooling yourself, the three traps that make almost every back-of-the-envelope AI calculation lie, and a worked example you can copy.

This is written for the owner or operations lead of a 10-to-250-person company who has a quote on the table, or a tool they’re tempted by, and wants a sober way to decide. No finance background required — if you can use a spreadsheet, you can run this.

Why does this matter so much for a small company?

Because you have no slack, and the odds are worse than the sales deck admits. MIT’s NANDA initiative studied 300 public AI deployments in 2025 and found that roughly 95% of generative-AI pilots delivered no measurable impact on profit and loss. The technology mostly worked; the return didn’t show up, largely because of how projects were chosen and integrated, not because the models were bad.

A large enterprise can absorb a failed six-figure pilot as a rounding error. A 30-person company cannot. The calculation in this article is your defence against joining that 95% — not by being pessimistic, but by being specific. You are not trying to prove AI is good or bad in the abstract. You are trying to answer one narrow question: does this particular automation, on this particular process, return more than it costs me to run for a year?

One more finding from that same research is worth holding onto: the biggest returns showed up in unglamorous back-office automation — cutting outsourced work, reducing manual operations — not in the flashy sales-and-marketing tools that absorbed most of the budget. The boring processes are usually where your real ROI lives. Your calculation should start there.

Which costs do you actually have to count?

The number-one reason AI ROI calculations are too optimistic is that they count the sticker price and stop. The sticker price is often the smallest line on the list. A defensible model includes five cost categories.

Cost category	What it includes	Easy to forget?
Build / setup	Time or fees to design and build the automation, write prompts, connect systems. Internal hours count too — they aren’t free.	Internal time is almost always undercounted.
Software & licences	Subscriptions, the AI API usage itself, the automation platform, any connectors.	The per-use AI bill, which scales with volume.
Integration	Connecting to your CRM, email, accounting, database; cleaning the data so the tool can read it.	Frequently the single largest hidden cost.
Training & change	Teaching staff to use it, time lost to the learning curve, writing the new procedure.	The productivity dip in week one.
Ongoing maintenance	Keeping prompts and knowledge bases current, monitoring output quality, fixing breakages, vendor price rises.	Almost universally ignored — and it’s recurring forever.

Two of these deserve special attention because they sink more SME projects than anything else.

How big is the AI usage bill, really?

If your automation calls a language model, you usually pay per use — priced in tokens (roughly, pieces of words in and out). For a typical small-business automation, this lands somewhere modest: a business agent handling a couple of hundred tasks a day commonly runs from about $10 to $60 a month depending on the model you pick, and can be under $25/month on a budget model. That’s small — but it is variable, and it scales with success. An automation that proves popular and runs ten times more often costs roughly ten times more. Model it as a per-transaction cost, not a flat fee, so a win doesn’t quietly become a budget problem. (Caching repeated context can cut this bill sharply, which is worth asking your builder about.)

Why is integration the cost that ambushes everyone?

Because the demo runs on clean, fake data and your business doesn’t. Connecting an AI tool to the messy reality of your CRM, your inconsistent spreadsheets, and your half-documented email folders is real work — sometimes more work than building the automation itself. If the information a process needs lives only in someone’s head or scattered across formats a machine can’t read, you’re paying to fix that before you see a cent of return. Budget for it explicitly. A useful rule: if you can’t already point to where the data lives in a structured, machine-readable form, double your integration estimate.

Why must maintenance be in the model from day one?

Because an AI system is not a project you finish; it’s a system you operate. Models change, your processes change, and a knowledge base that was accurate in January is confidently wrong by June if nobody updates it. Plan for roughly 15–20% of the original build effort, every year, just to keep the thing honest. Leaving maintenance out of the calculation is the single most common way a project that looked profitable on paper turns into a slow loss — you booked the savings once and the costs forever.

How do you put a defensible number on “time saved”?

Time saved is where most of the benefit comes from and where most of the self-deception happens. Three steps keep you honest.

First, value the hour correctly. Do not use the bare salary. Use the fully loaded cost of the person whose time you’re freeing — salary plus employer taxes, benefits, software, equipment, and overhead. As a rule of thumb that’s commonly 1.25 to 1.4 times base pay, sometimes higher. Someone on a €60,000 salary realistically costs you €75,000–€84,000 a year to employ. Using the loaded figure makes a genuine saving look better and stops you overstating thin ones; either way it’s the honest number.

Second, capture a real baseline before you change anything. You cannot prove a saving against a number you never measured. For the target process, write down today’s reality: hours per week it consumes, how long it takes end to end, how often it produces an error or a complaint, and what it costs. Guessing the “before” is how projects claim victories that evaporate under scrutiny. (Doing this baseline work is also the heart of an honest AI readiness audit — find where the value actually is before you build.)

Third, and most important: decide whether the saved time becomes money. This is the trap that wrecks more calculations than all the others combined. Saving each of ten people thirty minutes a day does not automatically save you the cost of 1.5 employees. Scattered minutes usually get absorbed into the working day — a longer coffee, a slower email — unless you can actually do something with them. Saved time converts to ROI in only two ways:

Hard savings — you genuinely reduce cost: you don’t backfill a departing hire, you cut outsourced or agency spend, you stop paying overtime, you avoid a planned hire as you grow. These are real euros and belong fully in the model.
Soft savings — the same people now handle more volume, respond faster, or do higher-value work. Real, but only countable if it converts into something measurable: more revenue, capacity you’d otherwise have paid for, a churn improvement you can point to.

Be ruthless here. Count hard savings at full value. Count soft savings only to the extent you can name the number they unlock — and when in doubt, discount them or leave them out. A model built on “everyone will be more productive” is not a model; it’s a wish.

What are the typical pitfalls that make the math lie?

Almost every over-optimistic AI business case fails in one of these three ways. Sanity-check yours against all three before you trust it.

Pitfall 1 — Counting only the hours

The classic error is hours saved × hourly rate = ROI, full stop. It ignores three things at once: that scattered minutes rarely convert to cash, that you used the bare salary instead of the loaded cost, and that there are costs on the other side of the ledger. Counting hours is the start of the calculation, not the end. The question is never “how many hours?” — it’s “how many hours, valued correctly, that turn into money I can actually point to, minus everything it costs me to run.”

Pitfall 2 — Ignoring maintenance and the running bill

Treating AI as a one-off purchase. Almost every cost in the model except the initial build is recurring: the usage bill, the subscriptions, the 15–20% annual upkeep, the monitoring time. A project showing a healthy return in year one can be underwater by year two once the recurring costs lap the one-time savings you booked. Always model at least a full year of running costs against a full year of benefit — never the build cost against forever-savings.

Pitfall 3 — Forgetting the human-in-the-loop cost

If your automation keeps a person reviewing and approving output — and for anything customer-facing, financial, or contractual, it absolutely should — then the AI is not saving 100% of that task’s time. It’s saving the drafting-and-fetching time and leaving the reviewing time. An automation that drafts a customer reply in seconds still needs a human to read and approve it. Calculate the saving on the part you actually removed, not the whole task. Over-claiming here is how a “90% time saving” turns out to be 40% in practice.

A fourth, quieter pitfall: building the wrong thing efficiently. If you automate a bloated, badly-designed process, you’ve spent money to do something pointless faster. Fix the process first — eliminate and simplify before you automate — or your beautifully calculated ROI is measuring the wrong activity.

A simple model you can copy

Here is the whole calculation on one process. The numbers below are illustrative placeholders — replace every one with your own measured figures.

The process: inbound customer enquiries that staff currently read, sort, and draft replies to by hand.

Step 1 — Benefit (annual).

Baseline: handling enquiries takes 20 hours/week across the team.
The automation drafts and triages, removing the reading-and-drafting work but keeping human review. Realistic time removed: 60% → 12 hours/week saved.
Loaded cost of that staff time: €35/hour (salary × ~1.3).
Does it convert to cash? You’re growing and this lets you avoid one planned hire — so yes, it’s a hard saving.
Annual benefit: 12 hrs × 52 weeks × €35 = €21,840.

Step 2 — Cost (annual).

Build/setup (mostly internal time): €4,000 one-off.
Software + AI usage: €40/month → €480/year.
Integration with the inbox and CRM: €2,000 one-off.
Training and the week-one dip: €800 one-off.
Maintenance: ~18% of build = €720/year, recurring.
Year-one total: €4,000 + €2,000 + €800 + €480 + €720 = €8,000.
Steady-state annual (after year one): €480 + €720 = €1,200.

Step 3 — ROI and payback.

Year-one ROI: (€21,840 − €8,000) ÷ €8,000 = ≈173%.
Payback period: €6,800 net one-off cost ÷ (€21,840 − €1,200 = €20,640/year ≈ €1,720/month net) ≈ 4 months.
Steady-state ROI: (€21,840 − €1,200) ÷ €1,200 = strongly positive — but only because the saved hours convert to an avoided hire. Remove that assumption and the benefit collapses to soft savings you’d have to justify separately.

The model’s verdict isn’t “AI is worth it.” It’s “this automation, on this process, with that hire genuinely avoided, pays back in about four months.” Change the one assumption that the time converts to cash, and the same spreadsheet tells you to wait. That sensitivity is the point: a good model shows you exactly which assumption your decision rests on.

What’s the shortest honest version?

Calculating AI ROI before you spend is not hard maths — it’s disciplined accounting against reality. Pick one process. Count all five cost buckets, especially integration and the recurring maintenance everyone forgets. Value saved time at the loaded cost of the people, and count it only to the extent it becomes money you can name. Then run ROI and payback on a full year. If it only works when you ignore maintenance, use bare salaries, and assume scattered minutes turn into cash, it doesn’t work — and finding that out on paper costs nothing. The companies that beat the 95% failure rate aren’t the ones with the best models. They’re the ones who did this arithmetic honestly before they signed.

This article is general guidance, not financial or accounting advice. Use your own measured figures and consult a qualified advisor for decisions that matter to your business.

Sources: Fortune — MIT report: 95% of generative AI pilots are failing, HiBob — fully burdened labor rate guide.

Educational material, not legal advice. As of 2026 — interpretation of the EU AI Act may change.