Every team building an AI agent hits the same fork. The model is good. Now, how much do you let it do without asking?
Most teams answer that with an accuracy number. "It's right 95% of the time, so let it run." I think that's the wrong question. Accuracy tells you how often the AI is right. It says nothing about what happens the 5% of the time it's wrong.
The question I actually ask before granting an AI action autonomy is simpler: can this be undone? That's the reversibility test, and it decides more about whether a product is safe to ship than any benchmark score.
Accuracy is a rate. Consequence is a floor.
An accuracy figure is an average across many attempts. It hides the thing that actually hurts you: the cost of a single wrong action, on its own, right now.
Sending an email is 95% fine and 5% embarrassing. Sending a wire transfer is 95% fine and 5% catastrophic. Same accuracy, completely different products. The model didn't change. The reversibility of the action did.
So before I look at how often the AI is right, I look at what one wrong action costs and whether that cost can be walked back. A high accuracy rate on an irreversible action is still a loaded gun. It just has fewer bullets.
The three tiers
I sort every action an AI might take into three buckets. The bucket, not the accuracy, decides how much autonomy it gets.
Tier 1 — Reversible. The action can be undone quickly, cheaply, and by the user themselves. Drafting text. Tagging a record. Reordering a list. Suggesting a reply. If the AI gets it wrong, the user hits undo and moves on. These actions should run automatically. Asking permission here just adds friction to something that was never dangerous. Make the undo obvious and let the AI work.
Tier 2 — Reversible with cost. The action can be walked back, but it takes effort, time, or a small amount of cleanup. Sending an internal message. Moving a deal stage. Scheduling something on someone's calendar. Publishing a draft to a private space. These can run on their own, but only if the reversal path is designed and visible — a clear undo window, an activity log, a way to see what happened and roll it back. The AI acts; the user can always find and fix the action after the fact.
Tier 3 — Irreversible or externally visible. The action can't be cleanly undone, or it's already been seen by someone outside the user's control. Sending an external email. Moving money. Deleting data with no recovery. Posting publicly. Making a commitment to a customer. These get a human gate, every time, no matter how accurate the model is. The AI prepares the action, shows its work, and waits for a person to confirm.
The trap is treating tier as a function of confidence. Teams say "the model is confident, let it send the wire." Confidence is irrelevant to reversibility. A confidently wrong irreversible action is the single worst outcome an AI product can produce.
Reversibility is a design lever, not a fixed property
Here's the part most teams miss: which tier an action lands in is often something you control.
An irreversible action can frequently be redesigned into a reversible one. Instead of the AI sending the email, it drafts and queues it with a 30-second cancel window. Instead of deleting, it soft-deletes to a recoverable state. Instead of committing to the customer, it stages the commitment for a one-tap human approval. You just moved an action from tier 3 to tier 2, and now the AI can do far more of the work while a person still guards the point of no return.
This is the single most valuable move in agent design. Don't ask "is the model good enough to do this irreversible thing." Ask "can I make this thing reversible, so the model being merely good is enough." Most of the autonomy you want comes from adding an undo, not from improving the model.
This is why every AI action should be observable, auditable, and off-switchable. Those aren't compliance checkboxes. They're what turns an irreversible-feeling action into a recoverable one — and recoverable actions are the ones you can safely automate.
How I run the test
Three questions, in order, for every action before it ships with autonomy:
- If this fires wrong, can it be undone? No — tier 3, human gate. Yes — keep going.
- How fast and how easily, and by whom? Instantly, by the user, at no cost — tier 1, fully automatic. With some effort or delay — tier 2, automatic but with a visible, designed reversal path.
- Can I make an irreversible action reversible? A cancel window, a soft state, a staged approval. If yes, do that before you decide the tier. Redesigning for reversibility beats gating for permission almost every time.
Run those three and the autonomy question answers itself, without a single argument about benchmark numbers.
The best AI products aren't the ones with the highest accuracy. They're the ones where being wrong is cheap — where every action the AI takes on its own is one a person can walk back, and the only actions behind a gate are the ones that truly can't be undone. Design for reversibility and you can give the AI far more to do, precisely because it's safe to let it try.
FAQ
Isn't a human gate on every important action just slow? Only if you gate the wrong things. The point of the test is to gate few actions — the genuinely irreversible ones — and let everything else run. Most teams do the opposite: they gate everything out of fear, or nothing out of ambition. Sorting by reversibility gives you speed where it's safe and caution only where it's earned.
What if I can't tell whether an action is reversible? Then treat it as irreversible until proven otherwise. Uncertainty about reversibility is itself a reason to add a human gate. The cost of gating a reversible action by mistake is a little friction. The cost of auto-running an irreversible one is the whole product's credibility.
Doesn't a cancel window just add latency to everything? Only to the actions that need it, and only briefly. A 15–30 second undo window on an outbound action is invisible in normal use and priceless the one time it matters. You're not slowing the product down; you're buying a cheap insurance policy against the expensive mistake.
How does this relate to accuracy testing and evals? They answer different questions. Evals tell you how often the AI is right, which decides whether the feature is worth shipping at all. The reversibility test tells you how much autonomy that feature should get once it ships. You need both — a great eval score does not earn an irreversible action the right to run unattended.
I'm Ravi Jadav, Chief Product Officer and Co-Founder at Sunbots Innovations and Co-Founder at Xwits Developers. Get in touch.