How to implement AI in an SME: a step-by-step guide

Many SMEs are in the same place: they know AI could help them, but they do not know where to start. Some people on the team already use ChatGPT on their own, someone has seen a spectacular demo and leadership is wondering whether to invest, how much and in what.

This guide sums up the method we follow in our projects: start with a specific problem, validate it with a small pilot and scale only what proves its value. Here is the route map; each stage follows in detail.

1. Start with a problem, not a tool

The most common mistake is starting with the technology (“we want a chatbot”, “we need to use agents”). It works much better to start with business questions:

For example:

• classifying and answering incoming emails and requests;
• extracting data from invoices, delivery notes or forms;
• an assistant that answers internal questions from the company’s documentation;
• generating reports and summaries from data that already exists.

2. Get the existing use in order

Before launching anything new, it is worth knowing what is already happening. In most companies there are people using AI tools on their own, sometimes with customer data. Three quick actions:

1. Inventory: which tools are used, by whom and for what.
2. Usage policy: which tools are approved and which data is never entered.
3. Basic training: as well as reducing risks, it has been an AI Act obligation since February 2025 (we cover it in this guide to AI literacy).

3. Prioritise with an impact and effort matrix

With the list of possible cases, rate each one on two axes:

• Impact: hours saved, errors avoided, revenue or customer satisfaction.
• Effort: availability and quality of the data, integration with your systems, process changes and risks (personal data, decisions about people).

The first project should come from the top-left quadrant: it has to go well, because the team’s and leadership’s confidence in everything that follows depends on it.

4. Validate with a pilot of a few weeks

A pilot is not a demo: it is a scaled-down version that works with real data and real users. For it to support a decision, define beforehand:

• What is measured: time per task, percentage of cases resolved without intervention, errors, satisfaction.
• The baseline: how long it takes today. Without it, there is no way to show improvement.
• The success criterion: what result would justify moving to production.
• Who supervises: AI makes mistakes, so the pilot must include human review.

A good pilot usually lasts four to eight weeks. If the data does not back it up after that time, that is also a good outcome: you have avoided a bigger investment.

5. Buy, configure or build

Not everything has to be custom-built. In general:

• Buy or switch on what already exists (for example, Microsoft 365 Copilot or the AI features of your CRM or ERP) when the case is generic: drafting, summarising, searching email.
• Configure and integrate existing tools and models when the case depends on your data and processes: an assistant with your documentation or automating a specific workflow. That is what we do in AI integration in apps and systems.
• Build to measure when AI is part of your competitive advantage or of the product you sell. That is the territory of custom AI development.

When choosing, also weigh up vendor lock-in: where the data is processed, whether you can switch models and what happens if the price goes up.

6. Look after data and security from the start

• Use business editions of the tools, with a data processing agreement and without your data being used to train models.
• Give the AI only the permissions it needs: an internal assistant should not show an employee documents they cannot see.
• Check whether the case involves decisions about people (recruitment, credit, assessment): these may be high-risk uses under the AI Act.

7. Measure, adjust and scale

If the pilot works, moving to production includes:

• integrating the solution into the tools the team already uses (nobody wants yet another screen);
• documenting the new process and who supervises what;
• continuing to measure the impact for the first few months;
• choosing the next case, building on what you have learned.

Common mistakes

• Waiting for “perfect data”. Many cases work with the documentation you already have; you tidy it up along the way.
• Trying to automate 100%. If AI resolves 70% of cases and a person reviews the rest, that is already a great result.
• Not involving the people who do the work. The people who do the task are the ones who know where the odd cases are.
• Not measuring. Without a baseline, the project depends on opinions.
• Leaving training until the end. Without a team that trusts the tool, there is no adoption.

Where to start today

1. Take our AI maturity test to see which area is weakest.
2. List the five tasks that take up most of your time.
3. If you would like help prioritising them and launching the pilot, that is exactly what our AI consulting for SMEs does.