Why Every AI Approval Has Two Clocks

Six · 9 min read · September 22, 2026

Why Every AI Approval Has Two Clocks

Say you run customer operations at a software company. On Monday, an AI agent drafts an email telling your existing customers that support hours are changing next month. It asks for your approval before sending. You read the copy, check who's on the list and approve it for Wednesday morning.

Then the send waits. Legal asks to see it first, and the agent holds until they reply. They come back on Friday with no changes. By then, a few other things have moved. Someone corrected a typo in the subject line. The list refreshed overnight and now includes everyone who signed up for the newsletter, not only paying customers. And your approval is still sitting there, marked approved.

Nobody withdrew it. But you approved one email to your customers. What's about to go out is a slightly different email to a much larger list, two days late. Is that still the thing you said yes to?

An approval is a decision about a specific piece of work under specific conditions. It can stay in force after those conditions change, and the word "approved" looks the same whether they have or not. So we think of every approval as having two clocks. One tracks whether your decision is still in force. The other tracks whether the work in front of you still matches what you reviewed.

"Approved" is doing several jobs at once

When an agent is ready to act, three separate questions are hiding behind the green check.

The first is whether the agent can technically do it. Does it still have access to the email system, and is the account allowed to send to this list? That's a question about systems and credentials. NIST's Zero Trust Architecture guidance treats it that way: access to each resource "is granted on a per-session basis," and getting into one resource doesn't automatically get the agent into another. The email system can let the agent send without anyone having decided the email should go.

The second is whether your decision is still in force. Has anyone withdrawn it or replaced it? Has it already been used? Are you still the person who gets to make it?

The third is whether the work about to happen still fits the decision. Is it the same email, going to the same people, at a time when the reason for sending still holds?

Outside all three are rules no approval can waive: privacy law, contract terms, your company's own policy on what goes to customers. An agent can have access, a valid approval and an unchanged email, and still be blocked by one of those.

Federal internal-control standards keep the first two questions apart. The U.S. Government Accountability Office's Green Book, the 2025 revision of the internal-control standards for federal agencies, lists "access restrictions to and accountability for resources and records" and "authorization of transactions" as different control activities. It was written for government, but the same split shows up in the support-hours email. The agent has had access to the send button all week. What it needs before Friday's send is a yes from someone who still has the authority to give one.

A saved approval keeps the decision, not the reasons

Agent frameworks are built to hold an approval while work waits. OpenAI's Agents SDK documentation describes a run that pauses when a tool call needs approval and saves its state, including "approval decisions, pending tool calls, and tool arguments." That state "is designed to be durable," so it can sit in a database while Legal reads the email. When the run resumes, it "continues where it left off."

That's the tooling doing its job, and it's also how a decision ends up in storage while the list refreshes. The same page advises keeping a version marker with any run that may wait a while, so it doesn't break when the models, prompts or tool definitions change. Nothing in the saved state tells the agent whether you'd still approve the email.

Resuming isn't renewing. Age alone doesn't make an approval stale, either. A month-old approval for work that hasn't changed can be perfectly sound. An approval from an hour ago can be out of date if the list changed in between.

The authorization clock: is the decision still in force?

The authorization clock tracks the life of the decision itself. Your approval for the support-hours email is active until something ends it. You approved one send, so sending the email uses it up. If the agent wants to send a reminder the following week, that's a new decision. If you'd set an end date, the approval would lapse on that date. You can withdraw it at any point, or replace it with a newer one. And say you moved to another team on Thursday and handed customer email to a colleague. Your approval would still be sitting there, but nobody could be sure it should carry the send until your colleague looked at it and said so.

Withdrawing an approval can stop work that hasn't happened yet, but it can't unsend an email. And when a decision is withdrawn, replaced or used up, keep the record of it. If Monday's approval disappears when someone approves the larger list on Friday, it will look as though the larger list was the plan all along.

The freshness clock: does the work still match?

The freshness clock compares what's about to happen with what you reviewed: the work itself, the facts behind it, who it affects, the timing and who owns the decision.

"The same" means the same as far as your decision is concerned, not identical down to the character. Fixing the typo changed the email but not what you decided. Sending the identical email to every newsletter subscriber changes the decision without changing a word. Comparing the text, or even an exact fingerprint of the file, can't settle freshness on its own.

You can also decide ahead of time which changes don't need you, and security and medical-device guidance already allows for this. NIST's configuration management guide describes a baseline as a set of specifications "formally reviewed and agreed on at a given point in time," and lets an organization name changes that are "preapproved," such as vendor security patches. Those changes are still tested and documented before they go in.

The FDA takes a similar approach for AI-enabled medical device software. Under its guidance on predetermined change control plans, a manufacturer describes planned modifications up front, and once the plan is authorized, the changes it describes can be made "without necessitating additional marketing submissions" for each one.

Neither was written for a customer email, but the idea carries over. When you approved the send, you could have said that typo and formatting fixes are fine, and that anything changing what the email says, who gets it or when it goes comes back to you. Then the typo is covered, and the new list plainly isn't.

What counts as a material change?

A change is material when it crosses a boundary you set, or when the person who owns the decision would want to know about it before the work goes ahead. No universal threshold or timer settles it. A review date on the calendar tells you when to look again, but it wouldn't have caught the list refreshing while the email waited for Legal.

If nobody set a boundary, ask whoever owns the decision now whether the email should still go. If you don't ask, the stored approval answers for you, and it always answers yes.

The Approval Freshness Test

Before approved work runs or resumes, ask five questions, and ask all of them. If any answer has changed, a person should look before the work goes ahead. Five unchanged answers still don't override a technical block, an end date on the approval, or a legal or policy rule.

Is it the same work?

Is this the action, output and version the approver reviewed? In the example, the subject line changed. As far as the decision goes, it's the same email.

Are the facts the same?

Have the inputs, assumptions or instructions behind the decision moved? Suppose leadership pushed the support-hours change back two weeks while the email waited. Every word would be unchanged, and the email would be wrong.

Is the consequence the same?

Has the audience, cost, sensitivity or reversibility changed? Going from paying customers to the whole newsletter list is the big change here. It reaches more people, including people who aren't customers yet, and it's an email you can't take back.

Is the timing still right?

Does the reason for acting now still hold? Customers were meant to get a month's notice. If Legal had taken three weeks instead of a few days, the same email would reach them with about a week to go.

Is it still the right owner?

Does the decision still belong to someone with the authority to make it? If you've handed customer email to a colleague, they're the one to answer.

Five ways to respond

The questions tell you whether something moved. Deciding what happens next takes judgment, and there are five sensible answers.

Continue, when nothing crossed a boundary, the decision is still in force, and nothing technical or legal is in the way.

Note it and continue, when a change falls inside what was agreed in advance. Record what changed, why it fits, who checked and when. The typo goes here.

Recheck, when something material moved or nobody can tell. Show the current owner what changed and let them decide. The new list goes here, and the recheck could end several ways: send to the original customers only, ask for a new approval for the larger list, or stop.

Ask for a new approval, when the work is meaningfully different from what was reviewed. Keep the old approval on record next to the new one.

Stop, when the decision was withdrawn, replaced, used up or expired, or when access or a rule blocks the action. Stopping halts what hasn't happened yet. It can't undo what already has.

More clicks don't mean more control

It's tempting to fix all this by asking for approval more often. That adds clicks without necessarily adding judgment. A click shows that someone made a decision. It doesn't show that they understood what they were approving, or that they could have stopped it.

The appendix on human-AI interaction in NIST's AI Risk Management Framework says human roles in oversight "need to be clearly defined and differentiated." Whether the people doing the approving are actually "empowered and incentivized to challenge AI system output" is a question it leaves open for further study.

So match the control to the work. The support-hours email can't be recalled once it goes, so it needs your yes on the send itself, with typo and formatting fixes agreed in advance so they don't come back to you. An internal draft you can fix later could simply be reviewed afterward. Watching the work as it runs would have shown you the list changing, but it wouldn't decide whether the email should still go. Whatever the control, the person who owns the decision needs to see what changed and be able to stop the send.

A fresh approval can still be wrong

Freshness only tells you the decision still fits the work, not that it was a good decision. An email can match exactly what you approved and still say the wrong thing to the right people. The review that made the decision sound in the first place still has to happen.

Know what each approval was for

Part of the problem is that one click often carries more than one decision. When you approved the support-hours email, you accepted the draft and you authorized the send. Accepting the draft says the writing is right. Authorizing the send says this email can go to these people at this time. There's a third kind, too: approving work to start, which lets the agent begin. A yes to one of these doesn't carry over to the others.

Tie each approval to the smallest thing the approver actually looked at, and keep a short record of it:

A record like that is what lets the next person piece the decision back together without asking you.

Try it on one approval

Pick one approval that's waiting right now, or one your team reuses every week. Write down what kind of decision it was and exactly what was approved. Then ask the five questions. If you can't answer one, leave it open and take it to whoever owns the decision today. Keep the old approval in the record, so today's answer doesn't rewrite what was decided before.

Say a Supanova atom had drafted the support-hours email. You choose how often your workforce checks in: every step, every task, or once per project. If you're working out which calls, like Friday's send, should come back to you, and where the work should keep moving, start at supanova.team.

Sources

AI ApprovalsApproval FreshnessAI GovernanceAI AgentsHuman Oversight