AI strategy for SMEs: where to start without burning the budget

A business owner with 60 employees called me in July. He had already spent €35,000 over six months — on AI tool subscriptions, a generic online training course, and an “AI transformation” consultant who had produced a 40-page strategy. The concrete result: three people on the team were using ChatGPT to write emails faster. Everyone else kept working exactly as before.

This isn’t an isolated case. It’s the most common pattern I see in Italian SMEs with 20 to 200 employees: awareness of the opportunity, competitive anxiety, rushed investment, disappointing results, cynicism. The problem isn’t AI. The problem is the order of operations.

The systematic mistake: buying a tool before defining a problem

SMEs that fail at AI adoption almost always follow the same script. Someone on the board reads a McKinsey article or hears something at a conference. A meeting is called. The decision is made to “do AI.” IT is asked to evaluate tools, or a consultant is brought in to propose a roadmap. Within a few weeks, licences are purchased, a training day is organised, and results are expected.

The problem is that nobody has answered a basic question: which specific process do we want to improve, by how much, and what metric tells us whether we’ve succeeded?

Without an answer to that question, any tool is equally wrong. With an answer, the choice of tool becomes almost obvious — and often costs far less than originally imagined.

Nobody has answered the basic question: which process do we want to improve, by how much, and what metric tells us whether we’ve succeeded?

Three questions to ask before any investment

Before buying any subscription or engaging any consultant, these three questions must have concrete answers — not generic ones, not aspirational ones.

Which process consumes the most time on your team? Not “where could AI help”, but where people’s time is being spent on repetitive, low-value tasks. The answer comes from talking to teams, not management. Operations managers know where the bottlenecks are. Sales reps know how many hours go into filling in reports nobody reads. Customer service knows which questions arrive every day, identical to the last.

Where does quality suffer because of human fatigue? Processes where error rates rise at end of day, where quality depends on who executes rather than on the process itself, where variability is high — these are the best candidates for AI support, because the benefit is immediate and measurable.

What data do you already have that you’re not using? Most SMEs have more data than they realise: customer emails, support tickets, contracts, post-sale feedback, production data. Often this data is inaccessible not for lack of AI tools but for lack of structure and cleanliness. An AI assistant fed poor data produces poor output. The question about data comes before the question about tools.

The four most common mistakes — and how to avoid them

Mistake 1: Starting with the tool, not the problem. The inevitable result is that the tool gets used for the use cases the tool itself promotes in its demos, not for the company’s real problems. ChatGPT Enterprise gets used to write emails. Copilot gets used to autocomplete sentences in Word documents. Nothing wrong with that — but it doesn’t justify the investment, and it doesn’t produce transformation.

Mistake 2: Treating it as an IT project instead of an operational one. AI adoption isn’t a software update — it’s a change in how people work. If the project is delegated to IT without involving operations, it produces technically correct and practically unused solutions. The team that runs the process must be a co-designer of the solution, not a passive recipient.

Mistake 3: Skipping the data infrastructure. Generative AI models can do a lot with limited data. But the most valuable use cases for an SME — answering customers with the right tone and information, analysing contracts with industry-specific terminology, categorising feedback against company criteria — require that the relevant data be accessible, structured, and up to date. Investing in AI before data is in order is building on sand.

Mistake 4: Expecting ROI in three months. For simple, well-defined processes, results can be visible within a few weeks. For complex processes requiring integration with existing systems, team training, and workflow adjustment, a realistic breakeven is between six months and a year. Companies that abandon the project at three months due to lack of results often do so precisely when they’re about to see the payoff from the work already done.

The minimum viable path for an SME of 20 to 200 people

This isn’t a universal framework — every company is different. But it’s the starting point I use most often, adapted to the specific context.

Step 1 — Audit of repetitive processes (2–3 weeks). Structured interviews with 8–10 people in different roles. The goal is to map high-volume, low-variability processes: how many weekly hours, what the cost of error is, how much depends on implicit context (“knowing how we do things here”) versus explicit rules. The deliverable is a list of 5–8 candidate processes, ranked by potential impact and ease of implementation.

Step 2 — Pilot on the most painful process (6–10 weeks). Take the process at the top of the list and build a minimum solution: not the final system, but one that lets you measure whether the approach works. A clear hypothesis (“if we automate ticket categorisation, average response time drops by 30%”), a defined success metric, a weekly feedback loop. If the pilot works, you have the internal evidence to justify scaling. If it doesn’t, you understand why before committing significant budget.

Step 3 — Build internal capability before scaling. The pilot must not create dependency on an external consultant. It must transfer competence to 2–3 internal people who understand how the system works, how to update it when the process changes, how to intervene when it produces wrong output. Without this step, every update requires going back outside, and total cost grows indefinitely.

What it actually costs — realistic numbers

Costs vary considerably depending on the use case, but here are the order-of-magnitude figures that reflect what I see in the projects I work on.

Consulting to set the strategy and run the first pilot: between €5,000 and €15,000 for an 8–12 week engagement. AI tool licences: between €200 and €2,000 per month depending on the tools and number of users. Internal team time, which is the hidden and most underestimated cost: allow for 4–8 hours per week for people involved in the pilot, for at least three months. At an average hourly cost of €30–50, that’s €4,000–10,000 in opportunity cost that rarely shows up in project budgets.

The breakeven relative to current costs depends on the process. For automatic categorisation of support tickets in a 10-person support team, the typical breakeven is between 4 and 6 months. For generating standard contract drafts in a law firm, even less. For processes requiring integration with ERP or legacy systems, the timeline extends and costs rise.

The hidden and most underestimated cost isn’t the software. It’s internal team time: 4–8 hours per week for three months that appear in no project budget.

A final word on patience

The pressure to “do AI” is real and growing. Competitors are experimenting. Clients are raising expectations. Vendors promise results in weeks. In this context, the most contrarian thing I can suggest is also the most effective: slow down at the beginning to go faster afterwards.

A company that spends six weeks thoroughly understanding the problem before choosing a tool will reach concrete results in four months. A company that starts immediately with the tool will spend a year justifying why it hasn’t yet produced measurable ROI.