Every conversation about AI in business eventually arrives at the same cliff: does it act on its own? Asked that way, the question has only bad answers. “Never” buys you a smarter dashboard and leaves the 2 a.m. work on your desk. “Yes” hands your company’s name to a system whose mistakes you’ll discover after they’ve been made in it. The way off the cliff is to notice that the question is malformed. Autonomy isn’t a property of the system. It’s a setting on each routine, and the setting is yours.
One routine, four positions
Take the most boring routine at Atelier Mira: reminding customers about overdue invoices. Here is that one routine at each stop of the dial. Everything else (the reading, the reasoning, the drafting, tomorrow’s reconciliation) is identical at every stop. The only thing the dial changes is the license to act.
The invoice-reminder routine at Atelier Mira, at each stop of the dial. Same intelligence, different license to act.
One tap on Approve and SENA sends both, logs both, reconciles both tomorrow.
Approve → sent 09:02 → graded in Friday's ledger.
The dial is per-routine, not global. Payroll can live at “Suggests” forever while invoice reminders earn “Runs.” Trust is earned one reconciled morning at a time, and it turns back left in one click.
Rightward moves are earned, not granted
How does a routine get from “Drafts” to “Runs”? Not by a toggle in a settings page on day one, and not by a vendor’s confidence. It gets there on its ledger. The invoice-reminder routine at Mira ran at “Acts on approval” for eleven weeks. Its reconciliation record over that stretch: forty-one reminders sent, thirty-eight hits on the stated expectation, zero customer complaints, two misses both graded and explained. At some point, approving the forty-second identical reminder stops being oversight and starts being ceremony. That (a specific routine, with a specific record, in a specific corridor) is what earns a move right. And the corridor stays narrow: under €25k, customers with clean payment history, everything else still queues for a yes.
Trust that arrives as a default is a liability. Trust that arrives as a track record is an asset. The dial is just the interface between the two.
The asymmetry that makes it safe
Moving right takes weeks of graded evidence. Moving left takes one click, effective immediately, no questions asked. That asymmetry is the entire safety model, and it’s deliberate. The dial is also per routine, which is what saves you from the false choice of trusting “the AI” as a monolith: at Mira, invoice reminders run themselves, reorder proposals wait for a yes, and anything touching payroll or a person’s employment sits at “Suggests”, permanently, as policy. All three positions are correct at once, because they’re answers to different questions about different risks.
At every position, every action is logged with its evidence, attributed to the routine that took it, and reconciled the next morning. “Runs the routine” means SENA stops waiting for permission; it never means SENA stops reporting to you. The unannounced action does not exist on any setting.
The vendors shouting “fully autonomous” and the skeptics answering “never autonomous” are having the same argument, and both sides are wrong the same way: they treat autonomy as an ideology when it’s an operations decision. Where should the dial sit? Wherever this routine’s record says it should, this quarter, for you. Somewhere left of what the hype wants. Somewhere right of where your Sunday evenings currently are. And always, permanently, one click from wherever you’d rather have it.