Starting with AI in SMEs: begin with one boring process
Published on 7/20/2026
Most AI stories sound the same: time savings in percentages, costs cut in half, staff suddenly twice as productive. Nice numbers — but if you run an SME or manage a team, the real question is: where do I start, without making an expensive mistake?
The answer is less spectacular than the promises: start small, start with something boring.
Why big AI ambitions so often stall
I get the appeal of the big vision. You read about companies transforming entire departments with AI and you don't want to be left behind, so you ask a vendor for a proposal — and that proposal is broad and ambitious, because a vendor who says "start small" is a harder sell.
The problem: a broad AI rollout touches a lot at once. Workflows that span multiple teams, data that isn't in order, staff who are skeptical, integrations with systems that have worked the same way for years. The odds of this going smoothly are slim, and the odds of it costing more than planned are high.
What's worse: if it goes wrong, nobody can say exactly why. Was it the tool? The implementation? The data? The people? You learn nothing, you pay a lot, and your confidence in AI projects takes a long-term hit.
What makes a good starting point for AI in your business?
A boring process is a task that is:
- Repetitive — it comes back every day or every week, roughly the same each time.
- Well-defined — it has a clear start and a clear end; it doesn't depend on ten other things.
- Time-consuming but not deep-thinking work — someone does it, but the work doesn't really require much judgment.
- Verifiable — you can check the result without being a specialist.
Think of: categorizing and summarizing emails, routing incoming requests, turning quotes into structured data, drafting meeting notes from a transcript, answering FAQ questions based on an existing document.
None of these tasks sound revolutionary. That's exactly the point.
How do you pick that one process?
Use these questions as a filter. The right process scores well on most of them.
1. How often does it happen? Once a month is too rare. Daily or weekly is interesting — the gains add up fast.
2. How much time does it actually cost right now? Not the time it's supposed to take on paper, but the real time, including switching between tasks, searching, and restarting. This is often more than you'd think, and easy to estimate if you track it for a week.
3. Does it have a fixed input and a fixed output? "Email comes in → structured summary in CRM" is well-defined. "Improve customer communication" is not.
4. Can someone without specialist knowledge judge whether the result is correct? If checking the output requires bringing in an expert, your feedback loop becomes too heavy. A summary you can read and judge yourself is fine.
5. Is the current way of doing it painful but not business-critical? You want room to experiment. Don't start with the process that brings everything to a halt if it goes wrong. Start with the one where a mistake is annoying but recoverable.
Proving value at small scale: measure to know
Once you've found that one process, the next step is simple: automate it, measure it, and draw a conclusion.
Measure how much time it currently takes, and put the solution next to it. After two or four weeks, check whether it works — not based on a feeling, but based on something concrete: less time, fewer errors, less back-and-forth.
If it works, you now have two things: a working solution and a story. That story is your ticket to the next step. You no longer need to convince anyone with promises — you have a result.
If it doesn't work well, you've still learned something. Maybe the input was too irregular, maybe the AI lacked context, or maybe the process was less well-defined than it looked. That lesson costs you one contained first step, not an entire implementation project.
AI as a teammate, not a miracle cure
I often compare AI to a new team member. When you hire someone, you don't hand them the most complex, business-critical project on day one. You start with something concrete, give clear instructions, check the result, and give feedback. That's how you build trust — on both sides.
AI works exactly the same way: good input produces good output. A clear task with clear context gives you a usable result. A vague assignment gives you a vague answer. If you point AI at a process that's already unclear, you get unclarity at scale.
Start with a task you understand well yourself, so you can also judge whether the AI is doing it right. That's not a weakness in the approach — that's how you build a working implementation instead of an interesting experiment nobody uses.
What "no silver bullets" means in practice
I don't promise clients savings I can't deliver — not because I'm pessimistic about AI (I see it add value every day), but because unrealistic expectations do more damage than a modest start.
A modest start that works always beats an ambitious project that stalls. A working solution on a boring process is the best preparation for a bigger step: you know how AI responds to your data, your team has worked with it, and you've learned where the limits are.
That's the structural improvement I'm talking about. Not a spike, but a new normal.
Q&A: common doubts
"Isn't my company too small for AI?" The question isn't how big you are, but whether you have repetitive processes. Almost every company has those, small ones included.
"Our data isn't in order. Does that need to happen first?" Not always. Some processes — like summarizing emails or answering questions based on a document — work fine without a perfect data structure. That's actually a reason to start small: you discover what you actually need.
"What if my staff won't use it?" Adoption starts with a problem they find annoying too. Pick the process together with the people who currently do it. If they see it genuinely saves them work, you won't need to convince them.
"How do I know it's safe for our data?" That depends on what data you use and which tool you deploy. This is a real question that deserves an honest answer, not a sales pitch. If privacy is a hard requirement, there are options to keep things closer to home. That's a conversation I'm happy to have before you build anything.
Where to start now
If, after reading this, you're thinking "yes, but I still don't know which process to pick," that's exactly the right question. The answer is in your business, not in a generic list.
I offer a concrete first step for that: a pilot. In about two weeks I build one working solution for your own situation — no report, but a running thing you see work yourself — so you see results before committing to a bigger implementation project. Fixed price, fixed scope, and upfront we agree on paper what "done" means.
No report, no silver bullets, just something that works.
Want to know if your business has a boring but valuable process hiding somewhere? Book a free intake. We'll spend half an hour talking through where you stand and whether this is a fit for your business.
Once you have a first AI solution running, the next question comes up: how do you know the output is correct? Read: How do you check AI output when you're not an expert?