Everyone is excited about autonomous AI agents. So are we — we build them. And that is precisely why we refuse to deploy some of them.
An autonomous AI agent is a system that observes, decides, and acts without human validation at each step. That is exactly what makes it powerful. And exactly what makes it a problem in the sectors we work with: law firms, finance, industrial companies under sensitive contracts. In these environments, a mistake doesn't cost time. It costs a client, a lawsuit, sometimes more.
The technical fact you cannot work around
No language model guarantees the absence of hallucinations. None.
The best models of 2026 are wrong less often than before, but they are still wrong — with total confidence, on verifiable facts. This is not a bug awaiting a patch: it is a property of the current technology. Models predict plausible text; plausibility is not truth.
As long as this property exists, giving an agent the ability to act alone in a context where the error is irreversible is a risk-management decision, not a technical one.
The right question: "on which decision do we remove the human?"
The question we get asked is almost always "can we have AI agents?". That's not the right one. The right question is: "on which precise decision are you willing to remove a human from the loop?"
The criterion that works in practice has two dimensions:
- Reversibility. Can the action be undone without damage? Sorting emails, yes. Sending a legal document to a client, no.
- Exposure. Does the action leave the company's perimeter? An internal summary reviewed before use exposes little. A message sent to a third party commits you.
Cross the two, and the map draws itself:
| Reversible action | Irreversible action | |
|---|---|---|
| Internal | Automate without hesitation (sorting, drafts, summaries) | Automate with logging (filing, archiving) |
| Outbound | Automate with after-the-fact review | Mandatory human validation (client sends, transactions, contractual commitments) |
The bottom-right cell is the one that matters. Drafting a client email: the agent. Sending it: a human. Preparing an accounting entry: the agent. Approving it: a human. Analyzing a shareholders' agreement: the agent. Delivering the conclusion to the client: a human.
Why "semi-autonomous" is not a watered-down compromise
The classic objection: "if a human validates everything, we lose the productivity gain." That's wrong, for a structural reason.
In most business processes, 90% of the time is in preparation — searching, reading, cross-checking, drafting — and 10% in the decision to send. A semi-autonomous agent absorbs the 90% and leaves the 10% to the human. The productivity gain is almost entire; the residual risk is almost zero.
This is the architecture we chose for Eridia's agents: they search, analyze, draft, prepare — but the moment something goes outbound, via email, Slack, or Teams, a human must validate the action. This is not a technical limitation; it's a design choice.
This framework has an often-underestimated secondary benefit: it is auditable. When a client, an insurer, or a regulator asks "who approved this send?", the answer exists, with a name and a timestamp. With the AI Act progressively entering into application, this decision traceability is moving from best practice to requirement — and it matches the French CNIL's recommendations on governing generative AI systems.
The three mistakes we see in the field
1. Autonomy by default. Deploying an agent with full rights, then restricting after the first incident. That's the wrong order: start closed, open perimeter by perimeter, with a track record justifying each opening.
2. The implicit perimeter. "The agent handles customer support" is not a perimeter. "The agent drafts level-1 replies, sends them alone if the client is not in dispute, otherwise hands off" is one. If the perimeter doesn't fit in three written sentences, it doesn't exist.
3. Trust by habit. The agent was right 200 times, so people stop reviewing. That is precisely the mechanism by which a hallucination ends up with a client. Human validation must be structural (enforced by the tool), not behavioral (left to vigilance).
Where to start
List your planned agent use cases, and for each ask the two questions — reversible? internal? Everything falling into the first three cells of the matrix can be launched quickly, with real gains within weeks. The last cell can wait until the technology guarantees what no model guarantees today.
And if you want to run this exercise with us on your real processes, we offer 30 minutes to map what can be delegated — and what never should be.
