In short
Governing AI in an SME means deciding three things. What the system holds for you, so open items leave your head. Where to put a person who checks, without becoming the bottleneck. Who owns the company's context, so the AI does not start from scratch each time. If these three answers are vague, AI is still a tool and not a system.
Key points
- The constraint on whoever decides is almost never time: it is the mental capacity taken up by open items and checks.
- Governing means choosing what the system holds and what a person decides, before you think about prompts.
- Control goes in three moments: before (what it can touch), during (where it stops and asks), after (who verifies).
- The company's context must belong to the company, not to a single chat or a single tool.
- Before prompts you need four answers: which data goes in, which outputs are validated, who owns the process, how value is measured.
In conversations with people who run an SME, three problems keep coming back. The first: you don’t lack time, you lack headspace. Too many open items to hold together, and your head stays busy remembering them.
The second: every result has to be checked, and the checking falls on the same people. The third: every new conversation with AI starts from zero, and the context has to be rebuilt by hand.
These are three different problems with one cause. AI came into the company as a tool, not as a system.
How do I take out of my head what AI can hold for me?
A general manager who opens ChatGPT to “write a summary” is using a Ferrari engine to go and buy bread. But the real problem is not power. It is that nobody has decided what the system should hold and what should stay with a person.
Governing AI starts here. You choose what the system holds: open items, deadlines, context, drafts. Then what a person decides: priorities, exceptions, replies to clients.
The organisational behaviour that creates chaos
When AI enters a company without a system, three things tend to happen. Each person uses a different tool. Nobody clearly tells a draft from support material or from a critical output. Management senses the chaos, but does not measure it.
This picture fits the European context. The European Commission links adoption, skills and trust in the same path. Google, in its official guide to optimising for generative search, repeats that quality, structure and reliability matter more than any “AI-first” shortcut.
Where do I put the control without becoming the bottleneck?
Checking every result is right. But if all the control sits in one person, the system stops where that person stops.
Fabio Armellini describes how he works with AI like this:
“I do, I verify and I check” - Fabio Armellini, call of 22 October 2026 (translated from Italian)
The answer is not to check less. Control goes in three precise points. Before: what the AI can touch. During: where it stops and asks for confirmation. After: who verifies the output before it goes out. The detail is in the article on where to put the human when AI works for you.
Why does AI start from scratch, and who owns the context?
Every new conversation starts without remembering the rules, the clients, the decisions already made. So the context has to be rebuilt by hand each time.
“between 6 and 8 hours a week go on context transfer […] practically one day a week” - Pietro Landri, diagnostic session of 28 September 2026 (estimate declared by the client)
Context should not live in a chat or in a tool. It should belong to the company, with a person who keeps it up to date. The first step is knowing which tools are in use: we cover it in the AI tools register.
What does it mean to “govern” AI instead of “using” it?
Governing AI means building a system where input, validation and responsibility are defined before the prompts. In practice:
- Data goes in once, from Excel, ERP or CRM, with no manual retyping.
- The process is validated: the AI does not decide for you, it gives you verified options.
- The output is measurable: you know how many hours you have recovered and where.
- The team adopts it, because the system is calibrated on the people and not generic.
The 4 questions that separate casual use from real governance
Before you ask “which tool do we use?”, you should be able to answer these questions:
- which data can enter the system
- which outputs require human validation
- who owns the process
- how you measure the value produced
If any of these four questions stays vague, AI is not yet a system. It is just an accelerator of variability.
The simplest test
If I asked you today to explain in 10 minutes where AI comes in and who checks what, could you answer without improvising? Could you say which decisions it speeds up and which errors it avoids?
If the answer is no, you don’t need a better prompt. You need architecture.
How do I know the system is really working?
For us, the first sign is not a more polished output. It is the moment the team stops asking “which prompt do I use?” and starts asking “which decision do we want to speed up?”.
Don’t look only at the number of prompts used. Look at these signs. Less time spent chasing data. More uniformity in outputs. Less rework before a meeting. More clarity on who validates what.
To see them you need a starting measure and one after the first work cycle. Without the first, the second tells you nothing.
What is the difference between AIEH and a prompt engineering course?
A course teaches you how AI works. We build how your team uses it, on real processes, with real data, live.
The difference is between knowing that GPS exists and having a navigator who knows the streets of your city.
We don’t sell training. We build a decision system: if it doesn’t produce results, it hasn’t worked, and we know how to measure that.
How do I know if I’m still in chaos?
You can tell in a few minutes. The same requests are redone from zero every week. There is no clear line between internal drafts and outputs to validate. Nobody can say how many hours you are really recovering. Each person on the team uses AI to their own standard.
If you recognise at least 2 of these signs, the problem is not the tool. It is the architecture.
Typical mistake to avoid
The mistake is not using the wrong tool. It is starting from prompts, skipping the process design, not defining ownership and validation, and measuring enthusiasm instead of value.
Where is it best to start?
The right sequence is not to buy more tools. It is this:
- pick a single high-friction process
- map input, output and bottlenecks
- define the exact point where AI helps
- assign a clear human validation
- measure the result against the starting point
For a more practical version of this step, continue with AI policy for SMEs: an operational checklist to avoid chaos. Or go to the For Companies page.
How to get started?
The starting point is the SME diagnostic: about 7 minutes to see where work gets stuck. You get a first operational diagnosis, the estimated cost of the status quo and the most sensible next step.
From there the path is calibrated on your profile. No commitment.
Primary sources and references
Frequently asked questions
What does it mean to govern AI in an SME?
Deciding in advance which data goes in, which outputs need a person to check them, who owns the process and how value is measured. Everything else, from prompts to tools, comes after.
How do I avoid becoming the bottleneck for every check?
Put the checks where they are needed, not everywhere. First define what the AI can touch. During the work, make it stop only on actions that are hard to undo. After, verify the outputs that leave the company.
Why does the AI start from scratch in every conversation?
Because the context stays in the single chat or tool. It has to live in a place that belongs to the company, with a person who keeps it up to date. Then each new job starts from rules and information already defined.