If you run an SMB and wonder where to start with AI, this article is for you. I'm not here to tell you everyone must adopt AI now — but I can share what I've seen work, and what systematically goes wrong.
I spent ten years in Shanghai, where AI integration in SMBs is eighteen months ahead of Europe. It's not a question of intelligence or means — it's a question of approach.
Why "now" is no longer a marketing slogan
In 2020, saying "we must adopt AI now" was wishful thinking. In 2026, it's just observation. LLM costs dropped by a factor of 50 between 2023 and 2026. An agent that would have cost €8,000/month to run in 2023 today costs between €150 and €400.
Concretely:
- Claude Haiku 4.5 or Gemini Flash 2.5 handle most SMB cases at negligible costs (a few cents per conversation).
- n8n, Make, Zapier have all integrated native AI connectors — no need to code 80% of glue layers anymore.
- Open source models (Llama, Qwen, Mistral) can now be self-hosted on a modest machine for ultra-sensitive cases.
Result: the brake is no longer technical or financial. It's strategic.
The number 1 mistake I see in SMBs
"We're going to do a big AI project that will transform everything."
No. Definitely not. What I see working, with managers who get real ROI from AI, is the opposite: an approach in small batches, measured, iterative.
Attack plan I use almost systematically
| Week | Step | What's produced |
|---|---|---|
| W1 | Task audit | List of the 10 tasks eating the most time |
| W2 | Prioritisation | Ranking by (hourly gain × frequency) / complexity |
| W3-W4 | First build | One automation shipped to production, with metrics |
| W5-W8 | Expansion | 2-3 additional automations, consolidated measurements |
| W9+ | Targeted AI agent | If and only if previous builds held |
Eight to ten weeks to prove it works on your case, before putting a euro on a real complex AI project. It's defensive, and that's exactly why it works.
The 4 AI builds that always work in SMBs
In the order I deploy them:
1. Automating recurring administrative tasks
Invoicing, follow-ups, onboarding, reporting. Zero AI at first — just well-oiled workflow with n8n or Make, what I deliver on my automation engagements. AI comes next, on micro-decisions (categorise an email, extract a field from a PDF, route a ticket).
Typical ROI: 5 to 15 hours per week, on a team of 5.
2. Automatic document extraction
Contracts, supplier invoices, received quotes, CVs, scanned delivery notes. Today's multimodal AI reads these documents with 95 to 99% precision on structured fields.
Typical ROI: 10 to 40 hours per month for an assistant who used to re-key by hand.
3. Consolidated dashboards
Nothing "AI" there, but it's the foundation that makes everything else useful. Without a consolidated view of your data, you won't know if your automation does its job.
Typical ROI: your Sunday evening.
4. The level-1 support agent
The first real AI agent to deploy, once the three previous builds have cleaned up your processes. Because without clean processes behind it, an agent gets stuck at the first messy case.
Typical ROI: 40 to 70% of level-1 tickets handled without human intervention.
What systematically goes wrong
The "we'll do it like Uber / Amazon / BNP" project — you don't have their volume, their data, or their data science team. Copying doesn't work.
The "we'll replace X people" project — the internal announcement kills the project before it starts. No one will help you map the tasks of their own job.
The "we signed with [all-in-one AI SaaS vendor]" project — 80% of all-in-one AI platforms are either marketing or lock-in. Make sure you can leave without losing your data and business logic.
The project with no measurement — if you don't measure time saved, errors avoided, cost avoided, you'll never know if your AI works. Neither will your team.
What to start with this week
Three things:
- List the 10 tasks eating the most time in your team. Not the most visible — the most repetitive.
- For each, estimate the time spent per week and the current error margin.
- Rank by (time × frequency) / automation difficulty.
The top 3 of this list is your entry point. Not a €50k project — a €2-4k build running in 3 to 5 weeks.
If you want me to look at your list together, grab 30 minutes. I'll tell you honestly which are worth it, and which are traps.
To go further
- Related service: Workflow automation
- Related articles: How to automate processes in an SMB · 5 daily tasks to automate
