The moment that decides whether you have meaningful control isn't when the AI acts. It's when someone asks you to prove it did what it was supposed to, and you have to reach for the evidence. A regulator, an auditor, a customer's risk team, a court. None of them will ask whether you intended to stay in control. They'll ask you to show it. And if you can't produce the artifact, then as far as they're concerned, and as far as the truth is concerned, you didn't have the control.
That's the whole reframe, and it's harder than it sounds. Meaningful control is not a claim you make. It's evidence you can produce. "We keep a human in the loop" is a sentence. It carries exactly as much weight as "trust me" until you can show who that human was, what they could actually do, and that they did it.
What a regulator counts as proof of control
The law already works this way, which is worth sitting with. The EU AI Act doesn't ask providers of high-risk systems to promise oversight. It makes oversight a design requirement: the system has to be built so a human can monitor it, understand it, and step in. Read what that is actually saying. The regulator isn't asking you to feel in control. It's asking you to have built the system so that control is demonstrable. It's the same instinct a good engineer already has.
So here's what proof looks like, concretely, whether the party asking is a regulator or just you being honest with yourself.
Boundaries set in advance. A dated record of what the agent was and wasn't allowed to do, written before it acted. Limits you describe after an incident are a story you're telling. Only limits set down in advance are control, because only those could have constrained the behavior in the first place.
A human who could really intervene. Not a name on an org chart. A person with the authority and a working mechanism to stop or override the system, and the competence to know when. A reviewer who can't actually act is a rubber stamp, and a rubber stamp is not oversight. It's the appearance of it.
A decision trail. What the system did, why, and on whose behalf, recorded per action and reconstructable later. You cannot prove control over actions you can't reconstruct. "The system decided" is not an answer. It's the absence of one.
Evidence the control was used, not just available. This is the one people miss. Oversight that exists on paper but never fires looks identical, from the outside, to no oversight at all. The proof isn't the policy that permits intervention. It's the record of the interventions that happened.
Reproducibility. The ability to show, after the fact, why the system produced the decision it did. "The model chose" will not survive a serious question. If you can't re-derive the decision from its inputs and its rules, you can't defend it, and you can't really claim you controlled it.
None of this is new to anyone who's shipped in a regulated or safety-critical world. Proof was always the deliverable. "It worked" never cleared the bar. "Here's how we know it worked, and here's the record" did. A regulator is just an outside party demanding the evidence you should already keep for your own sake.
The trap is the retrofit, and it's fatal. You cannot manufacture this proof after the request lands. Evidence assembled after an incident isn't evidence of control, it's a reconstruction, and a competent regulator can tell the two apart. So the work isn't writing a control narrative when someone asks. It's designing the system so the narrative writes itself, at runtime, while the thing runs.
Get that right and the question you dread changes character. It stops being a scramble to explain yourself and becomes a query you can answer, because you built the answer in before anyone asked. Meaningful control was never what you meant to do. It's what you can show you did.