Short answer: yes, it's worth it, but probably not where you'd think first. Most companies want to start with the big transformation: "AI will run our customer service", "AI will write our proposals". That rarely works the first time. What works is one narrow, frequent, text-based task whose results you can measure. You build from there.
Where companies stand
According to Eurostat, 19.95% of EU enterprises used some form of AI technology in 2025, 6.5 percentage points more than a year earlier. Size makes a big difference: 17% of small companies, 30.36% of medium-sized ones and 55.03% of large ones.
Adoption also varies widely between countries; in some Central and Eastern European countries it is still around half the EU average. If you run a small company and feel behind, you probably are behind the large companies in your own country. That says more about resources than about whether AI is useful to you.
What holds smaller companies back
Not money. Among EU companies that considered AI but did not adopt it, Eurostat found the most common reason was a lack of relevant expertise (70.89%). Next came unclear legal consequences (52.52%) and data protection concerns (48.83%).
That is an important message. The monthly fee of a general AI assistant is negligible in an SME budget. What is expensive is a bad start: a poorly chosen task where, six months later, nobody can say whether it delivered anything.
What it actually delivers
Two studies give the best footing, and both measured real work, not lab exercises.
Customer support. In a study of 5,179 customer support agents, an AI assistant increased the number of issues resolved per hour by 14% on average. For the least experienced agents the gain reached 35%. In other words, AI does not mainly make the best people better; it brings newcomers up to speed faster.
Consultants. In an experiment by Harvard and BCG, 758 consultants worked with GPT-4. On tasks within the AI's capabilities they were more than 25% faster, produced work of more than 40% higher quality and completed 12% more tasks. The researchers called the boundary a "jagged frontier": it does not hold for every task. On tasks outside the frontier, the consultants using AI did worse, because they trusted confident but wrong answers.
Two practical lessons follow:
- An AI assistant is strongest in routine text work: summarising, drafting emails, translating, finding internal knowledge.
- It does not help with every task, and it is not always obvious where it hurts. Someone has to check the output, and you need a rule for what it can be trusted with.
Where to start: four good first tasks
A good first task meets four conditions: it is frequent (happens many times a week), text-based (where today's tools are strongest), checkable (a person can quickly see whether the result is right) and measurable (you know how long it takes now).
1. Drafting emails and proposals. Replies to customers, cover text for quotes, routine updates. The AI writes the draft, a colleague edits and sends it. Quality is not the main gain; time is.
2. Summaries. Action items from meeting notes, three sentences from a long email thread, the gist of a 40-page tender. This is where most companies get their first real value.
3. Translation. If the company exports or works with foreign partners, translation is often the fastest payback of all.
4. An internal knowledge assistant. An assistant that answers from your policies, product data and process descriptions. This is a bigger step because it needs tidy documents, but once it works it speeds up onboarding and internal questions.
Where not to start
- Where mistakes are expensive and nobody checks. Legal opinions, tax decisions, the final wording of a contract without human review.
- Where the input is a mess. If your documents live in five places and nobody knows which version is current, an assistant just gives wrong answers faster. Tidy up first.
- Where nobody owns it. If nobody is responsible for people using and improving the tool, it quietly dies within six months.
- Where the volume is tiny. A task that happens twice a month won't repay the effort of setting up prompts, rules and checks. Save it for later, once the frequent tasks are running.
What we've seen
On a previous project, our team worked with a home furnishings retail chain where three employees monitored competitors' prices and promotions manually. We built an AI-based system on a no-code platform that tracks the prices, product data and promotions of 20 web shops, matches the products, suggests prices and helps plan promotions and purchasing. It went live in four months.
What generalises from it: the project didn't work because the AI was clever. It worked because the task was narrow and frequent, and we knew exactly how long it took before. That pattern is worth copying, even if your first task is just an email draft.
Five common mistakes in the first months
Too many tasks at once. If AI starts in five areas in the first month, there's no time to see what works in any of them. One task, one team, one month.
No baseline. If you don't know how long something took before, you can't tell how much you gained. You're left with impressions, and impressions flip at the first failure.
Everyone with their own tool. If every colleague uses a different AI, you can't build shared rules, shared training or shared experience. Pick one tool for the company.
Blind trust. AI writes confidently even when it's wrong. In the first weeks everybody experiences this and becomes either too suspicious or too credulous. The goal is in between: every output is reviewed by someone who knows the subject.
No owner. You need one person who collects good and bad examples, updates the rules and answers questions. Not a full-time job, but it has to be somebody's job.
What to tell your team
The first obstacle to AI adoption is often human, not technical. Some colleagues are enthusiastic; others worry the machine will take their job. If you don't talk about it, the uncertainty lingers, and the tool is either not used or used in secret.
Say three things openly. First: what you want to use it for, and what not. Second: the AI makes a draft, the responsibility stays with the colleague who reviews and approves it. Third: what the team will do with the time saved. If there's no answer to that, people will reasonably assume the goal is cutting headcount.
The best ambassadors are colleagues who have tried it themselves and can show a real example from their own work. One such demo is worth more than any presentation about the future of AI. If you have someone like that, give them time to help others, and recognise the work.
Which tool?
For most SMEs the choice is simpler than it seems: pick the tool that fits your existing office environment. If you're on Microsoft 365, Copilot sees emails and documents according to each user's permissions. If you're on Google Workspace, Gemini is already included in the plan. If neither, or if writing, analysis and long documents are the main work, a standalone AI assistant is the right start: Anthropic's Claude or ChatGPT's business plan. We work primarily with Claude ourselves. Claude Team costs $20 per user per month billed annually or $25 monthly, from two users, and connects to Microsoft 365 as well.
Don't compromise on one thing: use company accounts, not staff's free personal ones. In business plans, vendors commit contractually not to train their models on your company data. With free personal accounts that is not guaranteed.
How to measure whether it paid off
Before you roll it out, measure the starting point. Simple questions will do:
- How long does a proposal draft, a weekly summary or a translation take now?
- How often does it happen in a week?
- How many errors does the approver find?
Count the costs honestly too: licences, the hours spent on setup and training, and the time reviewers spend checking the output. Review time is the one most often forgotten, and in the first weeks it can eat a good part of the gain. It usually drops as people learn where the tool can be trusted and where it can't.
A month later, ask the same questions. If the answers show no real improvement, don't buy the next tool; check whether you picked the right task. A typical AI project pays back over months rather than weeks, but with a well-chosen first task the difference shows within the first month.
What to do now
- Pick one task using the four conditions above: frequent, text-based, checkable, measurable.
- Measure the current state for a week.
- Choose a tool that fits your environment, with company accounts.
- Write a one-page rule: what it may be used for, which data must never be pasted in, who checks the output.
- Run a short training session on common failure modes. Under the EU AI Act, this also counts towards your AI literacy obligation.
- Measure again after a month, and only then expand.
If you'd like us to tell you, after a short assessment, which task is the best first step in your company, have a look at our AI solutions.
FAQ
How much does an AI assistant cost an SME?
In Google Workspace, Gemini is already part of the plan. Microsoft's full Copilot costs $21 per user per month for organisations of up to 300 users, while Copilot Chat is free for Microsoft 365 customers. Claude Team costs $20 per user per month billed annually. The real cost is the rollout: choosing the task, setting rules and training people.
How many companies use AI?
According to Eurostat, 19.95% of EU enterprises in 2025: 17% of small, 30.36% of medium-sized and 55.03% of large companies.
Where does it deliver the most?
Research points to routine text work and less experienced staff. In one customer support study, the average gain was 14%, and 35% for newcomers.
Will AI replace our staff?
In SMEs, typically not. It mostly cuts the time spent on routine tasks so people can focus on more important work. Where companies try to replace whole teams, quality usually decides how far that can go.
What is the biggest risk?
Uncontrolled use: staff pasting company data into personal accounts without rules and using answers without checking them. A company account, a short policy and some training handle most of this.
Sources
- Eurostat – Use of artificial intelligence in enterprises
- NBER Reporter – The Economics of Generative AI (Brynjolfsson, Li)
- Harvard Business School – Navigating the Jagged Technological Frontier
- Microsoft – Copilot Plans and Pricing for Business
- Google Workspace Updates – The best of Google AI, now included in Workspace Business and Enterprise plans
- Microsoft Learn – Data, Privacy, and Security for Microsoft Copilot
- Anthropic – Claude plans and pricing
- Anthropic Privacy Center – Is my data used for model training?
