How Much Autonomy Should AI Agents Have? The Reversibility Rule

Six · 12 min read · September 15, 2026

How Much Autonomy Should AI Agents Have? The Reversibility Rule

Say you've handed an AI agent your overdue invoices. You want it to work out who hasn't paid, write each customer a reminder and keep the billing records up to date, so nobody on your team has to spend Friday afternoon chasing payments.

Writing a reminder and sending one look like the same job. They aren't. If the agent drafts a reminder with the wrong amount, someone catches it and fixes it before it goes anywhere. If it sends that reminder to a customer who paid last week, the customer has already read it. You can apologize. You can't make them unread it.

That difference gets you most of the way to answering how much autonomy an AI agent should have. There isn't one setting for the whole agent. Each action gets its own decision, and the ceiling on that decision is how well you could recover if the agent got that action wrong.

Autonomy belongs to an action, not an agent

Look at the invoice work as a list of the things the agent actually does. It reads the billing system and each customer's payment history. It drafts reminders. It marks invoices as chased and sets the next follow-up date. It sends the emails. And when a customer writes back asking for more time, someone has to decide whether to offer a payment plan or waive a late fee.

Those are five different decisions, and they sort into three kinds. Reading the billing system and drafting a reminder only prepare work, and nothing has changed until someone acts on the draft. Marking an invoice as chased changes your own records, which you can check and change back. Sending the reminder and offering a payment plan reach a customer, who reads the email or holds you to the terms you offered. An agent you trust to draft reminders hasn't shown you anything yet about whether it should be offering payment plans.

The Government of Canada's guide on the use of agentic AI draws the same line. It recommends "read-only by default" (view and draft) where possible. Its examples of actions that change a system's state are sending, publishing, approving, spending and updating records. For those, it recommends a review or confirmation step, unless the expected impact is demonstrably low and the action is trivially reversible.

Being able to send email isn't the same as being allowed to. The agent's tools may let it send, but that doesn't mean anyone in your business decided it should. That decision has to come from a person with the authority to make it, and it has to be enforced somewhere the agent can't change, like the permissions on the accounts mailbox. Approving one email is a smaller decision again: it covers that email, not the next one. And whoever owns your receivables is still accountable for what goes out, whoever pressed send.

OWASP's guidance on excessive agency has an example that shows why this matters. A developer wants an email assistant that only needs to read messages, but the plugin they choose can also send them. OWASP's fixes are plain: use a tool that only reads, connect with read-only access, or have a person review and press send on every draft the agent writes. It also recommends putting the authorization check in the downstream system, rather than relying on the model to decide whether an action is allowed.

The Reversibility Rule

We call this the Reversibility Rule. Before an agent gets more room on an action, ask four questions about that action:

  1. Is the scope narrow and written down?
  2. Can you see what actually happened?
  3. If it goes wrong, can you undo the action and repair what it caused?
  4. Has the work held up under conditions like the ones it will face next?

The answers set a ceiling. They tell you the most autonomy an action could reasonably have. They don't give permission. A reminder email might come through all four questions well and still be off limits, because your contract with a customer says collections go through their account manager, or because nobody with the authority has signed off.

Tying autonomy to reversibility isn't a new idea. It's already in Canada's test for when an action needs a review step. The four questions are a way to apply it one action at a time.

Keep the weak answers visible. A strong answer to one question doesn't make up for a weak answer to another. If you can undo an action perfectly but can't see when it went wrong, you won't know to undo it.

Is the scope narrow and written down?

Write the scope as limits a system can enforce. For the invoice agent, that might be invoices more than 30 days overdue, from your billing system only, using the standard reminder templates, sent from the accounts mailbox, for this quarter. Then the exclusions: customers in a billing dispute, accounts an account manager has flagged, anyone already on a payment plan.

If you don't know whether a limit is actually enforced, write "unknown" next to it. "The agent shouldn't email disputed accounts" is a hope until something stops it from doing so.

Access doesn't carry over, either. Billing access the agent was given for chasing invoices isn't permission to use it for issuing refunds.

Can you see what actually happened?

You need to see the result, not the agent's account of it. "Sent 40 reminders" is a summary. What you want is the list: which customers got which message, what each invoice looked like before and after, which emails bounced, who replied, and anything the agent skipped and why.

Keep that list somewhere the agent can't edit. Canada's guide asks for everything an agent does to be logged in a system the agent can't change. Then make sure someone reads it. OWASP puts logging and monitoring among the steps that won't prevent excessive agency but can limit the damage. For the invoice agent, that means someone in finance looking at the list while there's still time to pause the next batch.

Be honest about what the record can't show. If no customer has complained, that doesn't mean every email was right.

Can you undo the action and repair what it caused?

There are two questions here, and they're easy to treat as one.

The first is whether you can undo the action itself. If the agent marks the wrong invoice as chased, you change the field back. If it starts sending reminders to the wrong list, you stop it. That's the technical side: roll back, retry, pause, switch off.

The second is whether you can repair what the action caused. A sent email can't be unsent. Changing the invoice status back doesn't change what the customer read. If they forwarded it to their finance team, or paid twice, or decided they'd rather not keep working with you, you're into correcting, apologizing and fixing things by hand. If the agent offered a payment plan in writing, you may not be able to take it back at all.

Undoing the action can be easy while repairing the consequence is hard. When that's true, the harder answer sets the ceiling. Canada's guide covers the same ground: where a full undo isn't possible, it recommends previews, human approvals and clear ways to correct unexpected results, such as withdrawing, adjusting or notifying the people affected.

Write down both answers for each action, and say whether each one has been tested or is assumed. "We can roll it back" means more if you've tried it.

Has the work held up?

Track record counts, but only under conditions like the ones the agent will face next. If you've reviewed two months of drafted reminders and the corrections have mostly stopped, that tells you something about drafting reminders for this customer base, from this billing system, with these templates. It tells you much less if you switch billing systems next month or start chasing large accounts with their own contract terms.

The track record is evidence for whoever decides, so it should include the failures and near misses along with the clean runs. NIST's Generative AI Profile asks for performance to be demonstrated under conditions similar to where the system will be deployed, and it warns against extrapolating from "narrow, nonsystematic, and anecdotal assessments."

Write the grant down

Before you widen what an agent can do on an action, write down what you're granting. Keep one record per action. Where you don't know something, write "unknown" instead of leaving it blank, because the unknowns are often what should decide the route.

Here's what that looks like for one invoice action, sending the standard overdue reminder.

Field What the record says
Action Send the standard overdue reminder.
Context Invoices in your billing system, emailed from the accounts mailbox to the billing contact on file.
Intended outcome Each customer more than 30 days overdue gets one reminder with the right amount and invoice number, and the invoice shows the date it was chased.
Scope Invoices more than 30 days overdue, standard templates only, this quarter. Not disputed accounts, accounts an account manager has flagged or anyone on a payment plan. Whether those exclusions are enforced by the billing system or only written in the agent's instructions: unknown.
Evidence The sent list (who got which message, each invoice before and after, bounces, replies and skips), kept in a log the agent can't edit.
Commitment point The moment the email leaves the accounts mailbox.
Undoing the action Pause the send queue and set the invoice status back. Tested once, on a test list.
Repairing the consequence A correction email or a call from the account manager. A double payment needs a refund. Nothing fixes how a customer feels about a reminder they shouldn't have had.
Track record Two months of reviewed drafts, with corrections mostly stopped. Nothing yet on sending without a person checking first.
Route Act after approval of each action: someone in finance checks each reminder before it goes.
Authority Granted by your finance lead. The mailbox only sends approved templates. Whether the agent can change the amount inside a template: unknown.
Owner Your finance lead.
Stop and escalation A reminder to an excluded account, more bounces than a normal week, or a customer complaint. Pause sending and go to the finance lead.
Revocation The finance lead removes the agent's send access in the mail system.
Review date End of the quarter. If nobody reviews it, the grant ends.
What reopens it A new billing system, new templates, a new customer segment or a change to the model behind the agent.
Other limits Customers whose contracts send collections through their account manager. Any security rule about what's allowed to send from the accounts mailbox: unknown until someone checks.

That's a lot of fields for one email. Most of them are questions you'd want answered before letting a new hire send collection emails in your company's name. And the three unknowns, whether the exclusions are enforced, whether the agent can change an amount and what the mailbox security rule allows, are why sending stays on approval for now.

Pick a route for each action

Each action sits on one of five routes. They aren't a ladder the agent climbs. One agent can be on different routes for different actions at the same time, and doing well on one action doesn't move another.

Route What it means In the invoice example
Observe or read only Looks without changing anything. Reading the billing system and payment history.
Recommend or draft only Prepares work that someone else commits. Drafting reminders while you're still learning how it writes them.
Act after approval of each action A person reviews this specific action and its evidence before it happens. Sending each reminder after someone checks it.
Act inside a bounded grant Acts on its own within enforced scope, with monitoring, a way to stop it and an expiry. Updating invoice status and follow-up dates.
A person does it, or nobody does Consequence, recovery, authority or another limit rules out the agent doing it. Offering payment plans or waiving fees.

Approval only works if the person approving has the authority, the information and the time to say no. If someone is approving forty reminders in a row without reading them, the approval step isn't doing its job.

Your finance lead picks the route for each action and writes down why. Nothing should move an action to a new route on its own: not a good month, and not the fact that the agent has been running for a while.

When to look at a grant again

Every grant should be provisional, owned and set to expire. Say the invoice agent's grant to update records runs for the quarter, with your finance lead as owner. At the end of the quarter, they look at the record and decide again. If nobody looks, the grant ends. It doesn't roll over.

Some things should reopen it sooner. A reminder going to a disputed account, an exception nobody expected, or anyone pressing stop should send the action back for review. Once the problem is fixed, the action doesn't resume on its own. A person decides that it should.

Change should reopen it too. Canada's guide recommends reassessing risks and controls whenever an agent's tools, data sources, permissions or scope change. If you move billing systems, add a new customer segment, change the model behind the agent or rewrite the templates, the track record you built belongs to the old setup. It's still worth reading. It just can't carry the grant across by itself.

Where autonomy stops

Some limits sit outside the four questions. For the invoice agent, one is the contract that sends a customer's collections through their account manager. Another is a customer in a billing dispute, where a reminder can make things worse whatever it says. There may also be a security rule about what's allowed to send email from the accounts mailbox. A narrow scope, a clean rollback or a long track record doesn't settle any of them.

When one of those limits is unresolved, the owner keeps the action tighter and writes down which limit decided it. If nobody has checked the security rule yet, sending stays on approval, and the record says so. The next review then knows what would have to change.

The same person has to be able to stop the agent. NIST's AI RMF Core calls for mechanisms, with responsibilities assigned and understood, to "supersede, disengage, or deactivate" AI systems whose performance or outcomes are inconsistent with their intended use. For the invoice agent, that means your finance lead can switch off sending in the mail system without asking the agent.

Start with one workflow

Pick one workflow where an agent's mistake would cost you something. Break it into actions, each one a verb and an object. Mark where preparing turns into committing. Fill in the record for each action, unknowns included, and have the owner choose a route.

In the invoice example, reading, drafting and updating your own records stay inside your business, so they can move quickly once the scope and evidence are in place. Sending reaches customers, so it waits on approval until the record supports more. The payment plans stay with a person. A single autonomy setting for the whole agent would have been either too tight for the drafting or too loose for the payment plans.

Then put the review date in the calendar.

Supanova asks you the same kind of question about your own atoms. You choose how often your workforce checks in: every step, every task, or once per project. Start new work on the closer check-ins, and give it more room once you've seen it hold up.

If you're working out which of your team's Friday-afternoon jobs an autonomous workforce could take on, start at supanova.team.

Sources

AI Agent AutonomyRisk ManagementAI OperationsAI GovernanceHuman Oversight