On this page
AI implementation in your business does not start with a tool, but with a task. Pick work that costs time every week, measure how it goes now, test AI for four to six weeks on real data with a few people, then decide on what you measured: roll out, adjust or stop. A person keeps approving where it matters, and your team learns to work with it before you roll out wider; in the EU the AI Act has required that since 2 February 2025. A pilot costs little: in our example about 600 euros in licences and time, against 1,200 euros of time saved in six weeks. That way AI does not become a trial that dies after a month, but a fixed part of your work. Belgian firms are ahead here: 34.5 percent of companies with ten or more staff used AI in 2025, see AI adoption in the EU.
If you want a partner for this, see AI for business.
Where AI implementation usually goes wrong
Most small businesses start with AI the way they once started with social media: everyone tries something, nobody knows what it returns. A few patterns keep coming back:
- Everyone gets a licence, but nobody knows what for.
- The pilot runs on sample texts instead of real files.
- There is no baseline, so nobody can say whether it returns anything.
- Customer data ends up in a free tool.
- Nobody owns it, and after three months one person still uses it.
That is no reason to wait. It is a reason to handle it like any other project: a goal, a measurement, an owner and a decision.
Six steps from idea to everyday use
- Pick a task that costs timeNot "something with AI", but a task your team does every week: writing quotes, sorting emails, writing reports, retyping documents. The more often it comes back, the more a pilot teaches you.
- Measure how it goes nowHow many hours a week does the task take, how many errors or corrections, how long does a customer wait for an answer? Without that baseline you cannot prove anything afterwards.
- Run a small pilot on real dataFour to six weeks, a few people, real files. Use a business plan of a tool or a custom solution, never a free public tool for customer data.
- Keep a person approvingAI prepares, a person decides. Certainly for anything that goes to a customer or decides about money. Note during the pilot what you had to change: that says more than any demo.
- Roll out and trainIf it works, extend it to more people and teach them to work with it: what the tool does well, where it makes mistakes and which data never goes in. In the EU that is also how you meet the AI literacy duty in Article 4 of the AI Act.
- Measure again and keep buildingRepeat the measurement after a few months. If the gain holds, take the next task. If it fades, find out why before you buy more licences.
- Week 0Baseline
- Weeks 1 to 6Pilot on real quotes
- After week 6Measure and decide
Pilot card
Task in the worked exampleTwelve quotes a week
- Time per quoteBaseline45 minDuring the pilot20 min
- Time per weekBaseline9 hoursDuring the pilot4 hours
- Corrections after sendingBaseline3 a weekDuring the pilot2 a week
What do you need to know to decide?
- Does it save enough time to carry the cost?
- Do errors stay the same or fall?
- Does the team want to continue?
- Does a person still approve where it matters?
- Roll out
- Adjust
- Stop
Worked example: a six-week pilot on quotes
An installer writes twelve quotes a week, each 45 minutes of work. Three people test a business assistant for six weeks that writes a first draft from their own template and price list; a person checks every quote. We count 40 euros of wage cost per hour and ChatGPT Business at 21 euros per user per month, billed yearly. The times are an example; measure your own. The measurement itself is on the pilot card above: from 45 to 20 minutes per quote, 5 hours a week less. This is how the pilot adds up:
| Item | Calculation | Amount |
|---|---|---|
| Licences | 3 people x 21 euros x 2 months | 126 euros |
| Training | 3 people x 2 hours x 40 euros | 240 euros |
| Owner's guidance | 6 hours x 40 euros | 240 euros |
| Cost of the pilot | 606 euros | |
| Time saved in six weeks | 5 hours x 6 weeks x 40 euros | 1,200 euros |
| Time saved per year after rollout | 5 hours x 46 weeks x 40 euros | 9,200 euros |
If an hour of your team costs 50 euros, the yearly gain after rollout is 11,500 euros, and the pilot pays for itself sooner. Plug in your own hourly cost: the larger it is, the more an hour saved is worth.
In this example the pilot pays for itself within the six weeks. More important than the amount is the decision you can make afterwards: you know how much time it saves, whether errors fell and whether your team wants to continue. For quotes, a fixed structure helps the assistant: start from our quote template. What a custom solution costs per use, work out with the AI cost calculator.
Three routes to bring in AI
| Route | Example | Cost | Start | Fits when |
|---|---|---|---|---|
| An off-the-shelf assistant | ChatGPT Business, Claude Team | About 18 to 21 euros per user per month | This week | The task is writing, summarising or analysing |
| The AI in your software | Copilot in Microsoft 365, AI in your CRM or accounting | Copilot Business 18.20 euros per user per month billed yearly, on top of Microsoft 365 | This month | Your data already lives in that suite |
| A custom solution | An assistant on your documents, a flow that processes enquiries | Build plus usage per token | A few weeks | AI must work with your own data and processes |
Which tool fits which task: AI tools for business; the two biggest side by side: ChatGPT or Copilot.
AI use cases that often pay off first
The best first task is boring, frequent and easy to check.
| Task | Why it makes a good first one | What you measure |
|---|---|---|
| Quotes and texts | Comes back every week, a person reviews anyway | Time per quote, corrections |
| Sorting emails and preparing replies | High volume, clear types of question | Time to answer, emails per hour |
| Reading documents | Invoices and order forms as PDF, easy to check | Retyping errors, time per document |
| Customer questions | The same questions again and again, with a person on complaints and money | Share resolved without a person |
| Reports | Figures from your systems, AI only writes the explanation | Time per report |
For document flows and multi-step processes, an automation platform or a custom build often fits better than a chat assistant; see Make, Zapier or n8n.
Rules and data first, then AI
AI works as well as the data it rests on. A business that keeps its planning in three Excel files and a WhatsApp group mostly gets faster confusion from AI. That is why an AI project often starts with something that is not AI: your data in one place, with clear rules. How to take that step: Excel to web app.
For DA Logistics we built such a foundation. The platform tracks fuel under the Belgian Kader B regime to the litre, shows planning conflicts and lets drivers submit leave and receipts themselves in three languages. Because everything sits structured in one system, it is exactly the foundation AI can build on: a summary of deviations, a planning proposal, an answer to a driver's question.
That order is also the cheapest. A rule that is always right is cheaper and more reliable than a language model that is usually right. Use AI where a rule is not enough: where text has to be read, weighed or written. If you want AI connected to your own systems, that is AI automation. If AI has to choose its own steps, read AI agents explained.
Training your team has been a duty since 2025
In the EU, Article 4 of the AI Act on AI literacy has applied since 2 February 2025. After the amendment of July 2026 it asks businesses that use AI to take measures to support their staff's AI literacy. It is not a duty to reach a set level or a certificate, but it is a task to teach your people what they work with. Since 2 August 2026 people must also know when they are talking to an AI system. How to put that into a short AI policy: the EU AI Act for business; whether you need one: do I need an AI policy.
Support exists in many places, but it differs per country and changes often. Across the EU, European Digital Innovation Hubs offer "test before invest", training and advice on financing for small businesses, half funded by the Digital Europe programme. Check what your own region offers before you start.
We test AI on our own work first.
AMOSI is our own marketing software for small businesses; it went live in July 2026. One of its building blocks is a website scan that collects facts about a company from its own site, so a person starts from what is written down instead of a blank page.
Building it taught us where AI saves time and where a person has to decide. We draw that line again in every client pilot, from the same baseline on paper.
Questions about AI implementation
How do you use AI in your business?
Start with a task that costs time every week, measure how long it takes now, test AI for four to six weeks on real data and keep a person approving. If it works, roll it out and teach your team to work with it.
What are good AI use cases in business?
Tasks that come back often and are easy to check: quotes and texts, sorting emails and preparing replies, reading documents, answering customer questions and explaining reports.
How long does an AI implementation take?
A pilot on one task is best run for four to six weeks, long enough to compare before and after. How long rollout takes depends on how many people and systems are involved.
Do my staff need AI training?
In the EU, the AI Act asks you to take measures to support your staff's AI literacy; Article 4 has applied since February 2025. There is no required level or certificate, but whoever works with AI must know what the tool does and what must not go in.
What does AI implementation cost?
A pilot with an off-the-shelf tool costs little: in our example about 600 euros in licences, training and guidance for three people over six weeks. If you have something custom built, you pay for the build and then for usage. Count your team's time to learn and approve in both cases.
What if the pilot returns nothing?
Then you stop, or adjust: a different task, a different tool or better instructions. That is a result too, and a cheap one, because you did not buy licences for the whole team.