Two years have passed since generative AI stopped being a novelty and became a working tool. And yet, most small and medium companies remain stuck at the door, between fascination and paralysis: "we know we have to do something, but where do we start?" The answer is not buying the most expensive technology — it is following a sensible roadmap, one phase at a time.
Phase 1: identify pains, not technologies
The classic mistake is starting with "we want to use AI" and then looking for where to fit it. Reverse the order. Make a list of the tasks that consume repetitive time in your company: answering the same customer questions, summarizing documents, drafting similar proposals, classifying emails. Generative AI shines exactly there — in language work that is voluminous and low on creativity. The right opportunity is a real, frequent, measurable pain.
Phase 2: choose a first small, safe case
From the whole list, pick one case with three properties: clear value, low risk and measurable result. A typical good candidate is internal support — an assistant that answers questions about your own procedures. If it errs, no one outside is affected; if it hits, you save hours every week. Avoid starting with something customer-facing or tied to critical decisions: that comes later, with earned confidence.
Phase 3: give it your knowledge (without retraining)
A generic model does not know your company. Instead of expensive, slow retraining, use the RAG approach: you connect the model to your documents and it answers based on them, always current when you update them. It is the difference between an assistant that makes things up and one that cites your own documentation — and it is affordable for any SME.
Phase 4: measure before scaling
Before expanding, prove the value with numbers. A concrete example: a services SME measured that its team spent about 40 hours a week answering repeated customer questions. With an assistant fed by the knowledge base, those answers became automatic drafts reviewed by a human — the time dropped to ~15 hours. That is 25 hours a week recovered, over half a job, at an operating cost of tens of euros a month. This is the kind of math that opens budgets.
Phase 5: scale with governance and supervision
With a proven case, extend to others. But grow with rules: define what the AI can and cannot do, always keep a human validating sensitive decisions, and log what the system produces. Autonomy without supervision is risk; autonomy with guardrails is productivity. It is this discipline that separates companies that reap value from those that collect failed experiments.
The mistake that stalls everything: waiting for the perfect solution
Many SMEs delay waiting for the definitive tool or the right moment. But AI is learned by using it. An imperfect pilot that teaches the organization is worth more than an ambitious plan that never starts. Start small, measure, adjust, repeat.
In practice
You do not need a hundred-page AI strategy. You need to pick a real pain this week, set up a supervised pilot next month, and measure the result. The rest builds from there, with confidence and data. Which repetitive task, if automated with supervision, would free the most time in your team as soon as next quarter?
