Practical AI for small and mid-sized businesses

Most of the AI coverage aimed at business owners is written for enterprises with a data team and a seven-figure budget. If you run a 10-to-200-person company in Finland or elsewhere in Europe, that advice doesn't map to your week. You don't need a strategy deck. You need one or two things to work by next quarter.

The good news: the AI projects that quietly succeed in small and mid-sized businesses share a pattern. They are narrow. They attach to a workflow you already run. And a human is still in the loop while trust is being earned.

Start where the pain is boring

The best first AI project is almost never the one on the pitch deck. It's the one your team already complains about on Friday afternoons.

Look for tasks that are:

• Repetitive — the same shape of work, many times a week.
• Text-heavy — email, PDFs, forms, notes, transcripts.
• Non-final — a draft, a suggestion, or a triage step, not the last word to a customer.

Three examples we see land in mid-sized companies:

1. Support triage. Incoming emails are read, categorised, and either drafted a reply or routed to the right person. The human still sends.
2. Document extraction. Invoices, delivery notes, or contracts arrive as PDFs. Key fields go into the ERP without someone retyping them.
3. Internal search. Employees ask a chat interface questions about internal handbooks, product specs, or past project notes, instead of hunting through SharePoint.

None of these are glamorous. All of them return time to people every single day.

What to skip in year one

A short list of things that sound like great AI projects and almost never are, for a company under 200 people:

• A custom foundation model. You will not train GPT. You will call an API.
• A company-wide chatbot on your website. The maintenance cost is higher than the value for most SMBs. Start internal.
• A "single AI platform" that promises to do everything. In practice you end up paying for ten features and using one.
• Predictive analytics on data you don't yet collect cleanly. Fix the data pipeline first; the model is the easy part.

If a vendor's demo requires your data to already be perfect, that's a signal, not a feature.

Scope the first project like a two-week experiment

The failure mode we see most often is scope. Someone gets excited, the project grows arms and legs, and six months later there's nothing to show. Cap the first one hard:

• One workflow. One team. One measurable before/after.
• A two-week build, then two weeks of shadow mode where the AI runs alongside the human and its output is reviewed.
• A written kill criterion — "if accuracy on the last 50 tickets is below X, we stop and reconsider."

Boring project management, applied to AI, is what separates the pilots that ship from the ones that stall.

Keep a human in the loop until the numbers earn trust

The most common reason an AI rollout gets pulled is not that the model was bad. It's that someone lost trust — a customer got a wrong answer, a supplier got a wrong number — and there was no review step to catch it.

Design for that from day one:

• The AI drafts, a person approves.
• Every decision is logged with the input, the output, and who signed off.
• Confidence is exposed. If the model is unsure, it says so, and the case is escalated.

Over time you can loosen the loop on the categories where the numbers are boringly good. You do not have to loosen it everywhere at once.

A realistic first-year outcome

If you pick a narrow workflow, ship a two-week build, keep a human reviewing, and measure honestly, a mid-sized company can reasonably expect one or two workflows that save each involved employee a few hours a week by month six. That's it. That's the win.

It is not a headline. It is a real, compounding change in how the business runs — and it's the foundation that makes the next project easier, because the plumbing, the review process, and the internal trust are already in place.

The companies that get to a serious AI capability in three years almost all start here. The ones chasing the headline almost all stall.