How to start with AI: a practical guide for SMBs

Most AI projects in SMBs run aground at the same moment: right after the strategy session. A consultant delivers an 80-page report, you read through it, and then everyone goes back to the daily grind. Three months later somebody asks at a Friday drink whether "that AI thing" is still running anywhere. The answer is usually no.

This piece is about how it actually works. No consultancy circus, no strategy without execution, no six-month platform shopping. Instead: a concrete pilot within two weeks, measured in time and money, on one clearly scoped process. Only then do you scale.

This approach isn't for everyone. If you're building a data warehouse for 200 employees you genuinely need a strategy phase. But for 90% of Dutch SMBs, and according to CBS data that's the overwhelming majority, this is the path that works.

Step 1: first understand what AI can and cannot do

Before you pick a tool or call a vendor, make sure you broadly understand what the current generation of AI does. Not to become an expert, just to avoid building a chatbot when you needed a spreadsheet, or deploying an agent when an email template would have done the job.

Knowing the difference between an LLM and an AI agent is the minimum. A language model like ChatGPT or Claude answers a question. An agent executes tasks across multiple steps. People who don't have that distinction clear in their head usually buy the wrong thing.

What AI does well right now: summarising text, classifying, rewriting, extracting structured data from unstructured input, generating code, taking simple rule-based decisions. What it doesn't do well: arithmetic without tools, exact recall without a memory layer, guarantees on factual accuracy, and replacing existing software where deterministic logic is required. Keep that list in the back of your head whenever someone proposes "AI" as the answer.

Read the explainer on what AI is for SMBs if you haven't yet. Forty-five minutes of reading saves you three months of going the wrong direction.

Step 2: find your three time-sinks

No abstract exercise, no post-it workshop. Take a week, set a timer on your phone, and three times a day note what you or your team is actually doing. Or even simpler: ask the three most productive people on your team where their time goes.

You're looking for tasks with digital input and digital output. Drafting a quote from a client PDF is a good candidate. Counting physical inventory isn't. The task also needs to be repetitive with variation. Not identical, because then a macro is enough, but the same pattern. Think answering emails, rewriting content for different channels, entering invoice data, categorising support tickets. And it needs to be quantifiable in time. "This takes me 20 minutes on average, I do it 30 times a week." Without that number you can't measure later whether it worked.

According to the Stanford AI Index Report 2025, the average productivity gain from LLM-based tools on repetitive knowledge work sits between 25% and 40%, with peaks up to 70% on writing tasks. So you need a time-sink of at least 4 to 5 hours a week before the effort pays itself back.

Step 3: pick one pilot

One. Not three at once, not "we'll start broad". One. From your three candidates pick the one with the highest time investment and the lowest complexity. Often that's not the sexiest option, and that's exactly right.

In the pilots I've watched work in the Dutch SMB market over the past two years: categorising incoming emails and pre-drafting replies in Outlook or Gmail, rewriting product descriptions for webshops from a supplier feed, auto-summarising meetings with action points per person, drafting quotes from an intake call and a price list, or handling first-line customer service for the standard questions that turn out to be 80% of the volume.

What you should not pick as a first pilot: a chatbot on your site that needs to "do everything", a sales prediction model on data you still need to collect, or an agent that takes over "the entire customer process". Those are phase-3 projects, and they're how you torch your team's belief in AI in one go.

If you can't decide between candidates, take the AI scan. It gives you a ranking in 5 minutes of where the highest return per hour sits for your business.

Step 4: put two weeks on the clock

This is where most projects still run aground. Not because it's technically hard, but because it needs to feel urgent. So block two weeks in the calendar, no longer.

Week 1 is build. Day one, lock down the happy path with ten examples of input and desired output. Day two and three, build the first version. For 80% of pilots in this category you don't need to write any code. ChatGPT with custom instructions, Claude Projects, a Zapier flow or a Make scenario gets you very far. Only when the pilot works and scaling is on the table do you look at custom integrations. Day four and five, test with the actual user, you or a colleague, not with an ideal demo. Note where it breaks.

Week 2 is polish. Day six to eight, fix the three most common errors. Not every edge case, just the three that cause 80% of the problems. Day nine, deploy it into the real workflow for one person. Day ten, measure, which is step 5.

No platform comparisons, no RFPs, no vendor beauty contests. If the pilot works, in phase 2 you'll look at whether it should sit on a more serious platform. The OpenAI documentation and Anthropic's prompt engineering guide are free and good enough for the first version.

Step 5: measure what it delivers in time/money

Without measurement you don't have a pilot, you have a hobby. The measurement doesn't need to be complex, but it does need to be hard.

Three numbers, recorded before and after. Time per task: stopwatch, not estimate. Take five representative cases and time how long the old method takes, then time the same five steps with the new workflow. Error rate: how many of those five outputs need correction? This matters more than time, because a process that's 90% faster but wrong half the time costs you net time. And user satisfaction. Not a formal NPS, just: would you want to keep working with this tomorrow, yes or no? If the answer is no, the time savings don't matter.

Translate it into money. A process going from 30 minutes to 8 minutes, run 30 times a week, is 11 hours saved per week. At an internal hourly rate of 60 euros that's 660 euros per week, or 34,000 euros a year for one pilot. Against 15 to 30 euros a month in tooling, payback is usually within a week.

According to a 2024 McKinsey study, 23% of companies starting with AI achieve measurable revenue or cost impact in the first year. The difference between that 23% and the rest is almost always this point: do they measure concretely, or not.

Step 6: document and train your team

Pilot works, numbers check out. Now comes the part 80% of companies skip: making sure it still works in six months when you're on holiday.

Write down what the pilot does and doesn't do. Not as an ISO document, just as one A4 page in Notion or Confluence. What's the input, what's the output, where is the human in the loop, and when do you escalate to manual. Then train at least two people. One bottleneck isn't a success, it's a single point of failure. DataDream's AI training programmes are built specifically for this, but you can also do it yourself with two hours of explanation and a week of shadowing.

Finally, check your obligations under Article 4 of the AI Act. As of February 2025 this is in force: anyone working with AI systems in your organisation needs to be sufficiently AI-literate. That sounds heavy but isn't. You need to be able to demonstrate that the people using it understand what it does. Read the AI Act explainer or take the AI Act check to see what applies to your situation.

Documentation and training are often the difference between "fun pilot from last year" and "standard part of how we work".

Step 7: scale or stop

At the end of the pilot you have three options: continue and expand, replace with something else, or stop. None of those three are wrong. What is wrong is staying in the floating middle ground where you're no longer enthusiastic but haven't officially stopped either.

Decide on numbers, not feelings. Time saved of 25% or more with high user satisfaction: scale. Add the second task from your time-sink list and start over at step 3. Time saved between 10 and 25% with mixed satisfaction: two more weeks of polish, no longer. If it's still not there after that, stop. Time saved under 10% or low satisfaction: stop, learn, pick a different pilot. It's not failure, it's information.

One scaling rule: only add a second pilot once the first has run without your attention for at least 4 weeks. Otherwise you build a pile of half-working systems that all demand your time.

What this isn't

For clarity, these are the patterns this guide explicitly pushes back against.

No "AI strategy first". An AI strategy without pilots is a PowerPoint without a customer. Strategy follows from patterns you see in pilots, not the other way around. If a consultant says you first need a two-year vision, thank them and start at step 2 above.

No "platform first". Microsoft Copilot, Google Gemini, your own LLM on Azure are all fine choices, but only in phase 2. In phase 1 it doesn't matter. A pilot that works on ChatGPT also works on Copilot, and vice versa.

No "big bang implementation". Switching the whole company to AI workflows in one go is how you lose everyone. One pilot, one team, one month. Then the next.

Common mistakes

In the pilots I've guided the same handful of errors keep showing up. The scope is too large. "We want to automate our entire customer service" isn't a pilot, it's a programme. Start with the three questions that drive 60% of the volume.

No baseline gets measured. Without a before-measurement you don't know whether you delivered anything, and "it feels faster" is not a measurement. People lock into a platform too early. Signing a yearly contract with an AI vendor before you have a working pilot is almost always more expensive than two months of your own experimentation. The user gets skipped. The person who does the work daily knows better where the friction is than the boardroom. If that person isn't building along, the pilot fails. And the exit criterion is missing. Agreeing in advance when you stop is just as important as agreeing when you scale, otherwise every half-working pilot keeps swimming around until somebody quietly retires it.

When to bring in outside help

Most of the above you can do yourself. External help genuinely makes a difference at a few specific moments.

If you don't have time to focus for two weeks, an external partner is sometimes the only way to enforce a real deadline. If the pilot touches customer data, financial data or personnel data, you don't need to pay top dollar for a GDPR consultancy, but you do want someone who can check whether the setup is legally sound. If you want to scale to more than 5 processes, thinking about architecture starts to make sense, and an AI strategy phase becomes a valid investment. Only then, not earlier. And if your team isn't AI-literate under Article 4 of the AI Act, a one-day training programme gets you over the legal threshold and saves you fines and hassle later.

Even with help, stay owner of the problem. The external party does it faster, you understand the business better. The combination works, the handover rarely does.

Conclusion

Starting with AI isn't a matter of strategy, it's a matter of one two-week pilot on one concrete time-sink, with hard numbers before and after. Do that once well, and your second pilot takes half the energy. Do it ten times and AI isn't a project anymore, it's just how you work.

The biggest mistake I see in SMBs isn't that people pick the wrong tool, it's that they spend months picking. Two weeks of concrete action delivers more than six months of comparing.

Want to know which pilot has the highest return per hour for your business? Take the AI scan, 5 minutes, no sales pitch afterwards. Or get in touch if you'd rather spar directly about your three time-sinks.