You can’t govern the AI you can’t see
By Xavier Franco, CEO de Elternativa · Grupo Zemsania
· 6 min read
Written with the help of AI and reviewed by Xavier Franco

In this article
There is one question that almost never gets a quick answer in a company: how many AI agents are there in yours today? Not the ones you approved, but all of them. The ones in the CRM, the ones in the ERP, the ones someone built with a low-code tool, the ones each supplier uses and the ones each person brings from home. That is called shadow AI, and it is the real starting point of AI governance.
Your teams already use AI, and often their own
There is no need to guess. Microsoft and LinkedIn’s 2024 Work Trend Index, based on a survey of 31,000 people in 31 countries, found that 75% of knowledge workers already use AI at work. And that 78% of those who use it bring their own tools.
The same report has another figure that I find even more revealing: 79% of leaders believe that adopting AI is critical to staying competitive, but 60% say their company has no vision or plan to implement it.
In other words: AI has already come in through the back door, while the company decides where it wants it to come in.
In practice: ask three different teams which AI tools they use in their day-to-day work. Compare their answer with the list of approved tools.
Why banning doesn’t work
The first reaction is usually to ban. I understand it, but it doesn’t work: if the tool helps people do their work better, they will keep using it, only without telling you. And then you lose exactly what you need: visibility.
Shadow AI is not just the employee with their chatbot. In an average company today, agents from different suppliers live side by side, each in its own system, with no common point:
- the assistants that come built into the CRM, the ERP or the service platform;
- the flows someone built with a low-code tool;
- the AI that each development supplier uses, each in its own way;
- and what each person uses on their own.
What each of them does ends up, at best, in logs that only a specialist understands.
In practice: don’t start with a ban policy. Start with a question: which AI touches our data and who put it there?
The three risks that really matter
When AI can’t be seen, three risks pile up that management ends up paying for:
- Data. Sensitive information leaving for tools that nobody has reviewed.
- Cost. Model consumption grows out of control, with no idea which agent or use case triggers it. The AI bill is already a new line in the budget, and nobody knows what it has delivered.
- Compliance. Without traces or evidence, proving that you comply with the EU Artificial Intelligence Act is not feasible.
None of the three is solved with a memo. They are solved with inventory, rules and evidence.
From shadow to inventory: how I would do it
My proposal is to move from banning to seeing, controlling and proving, in this order:
- Living inventory. A list of all your AI systems, with their agents, the knowledge bases they use and the integrations they touch. Living, because it changes every week.
- Identity per agent. Each agent with its own key, to know who does what and to stop one from impersonating another.
- Rules that apply themselves. Thresholds for cost, latency and sensitive data, with a simple criterion: first warn and, if needed, block.
- Evidence. Every operation of every agent is recorded, so you can explain what happened and prove it to whoever asks.
- A view for management. Cost, risk and usage in a single dashboard, to decide with data.
And something important: the rules have to be deterministic. If the control of AI is done by another AI, you don’t have control: you have another opinion.
In practice: pick the area where the most AI is used (customer service, sales or development) and start the inventory there. With one area well covered you learn more than with a policy for the whole company.
What all this has to do with the AI Act
A lot more than it seems. The EU Artificial Intelligence Act is already under way: according to the European Commission’s official timeline, the prohibitions and the AI literacy obligation have applied since 2 February 2025, and general application arrived on 2 August 2026.
AI literacy is exactly what shadow AI puts to the test. In its questions and answers the Commission clarifies that it does not require a certificate, but it does require training adapted to each person’s role, and that it is advisable to document what is done. Pointing to the manufacturer’s instructions is not enough.
And how are you going to train people on the AI they use if you don’t know which AI they use? The inventory is the first step of compliance, not an extra. Without knowing which systems you have, you can’t classify their risk, or train those who use them, or prove anything if you are asked.
In practice: cross-check your first AI inventory with the people who use each tool. That list is the basis of your AI literacy plan.
Where to start
- Ask and listen: which AI is used, in which teams and for what.
- Make the first inventory of agents and integrations, even if it is incomplete.
- Define three rules that are not negotiable: which data does not leave, which cost triggers a warning and which agent needs approval.
- Connect the telemetry of the agents you already have, without changing supplier.
- Review every month the inventory, the cost and the warnings with management.
You don’t need to have it all on day one. You need to stop not seeing it.
How we help you
At AI Governance we help you see and govern your company’s AI: a living inventory of agents, knowledge bases and integrations, rules for cost, latency and sensitive data, and the evidence that the EU Artificial Intelligence Act asks for, built from real data.
We do it with Regentic.AI, of which we are the exclusive distributors in Europe. It connects through open standards such as OpenTelemetry, works on any framework and model, applies the rules with a deterministic engine and issues sealed evidence reports.
This article is informational and does not replace your organisation’s legal advice.
How many AI agents are there in your company today?
In a demo we show you how Regentic.AI inventories your agents, applies cost and sensitive-data rules and leaves evidence of what they do.

