An approval gate on every step teaches people to click yes without reading. Put one where a mistake cannot be taken back, give the reviewer what they need to decide, and let them send the work back with a reason.
The first instinct with an AI agent is to make a person approve everything it does. It feels safe. Within a month the approvals queue is long, the reviewers are busy, and every request gets the same quick yes. A gate that approves everything protects nothing. The useful question is where a gate goes and what the person at it needs to see.
A run is many steps. Most of them read: look up the customer, pull the policy, check the calendar, draft a reply. A mistake in any of those is caught by the next step, or by the person who reads the result. Some steps write: send the email, issue the refund, change the record, submit the form. A mistake there leaves the building.
Put the gate in front of the step that cannot be taken back, and only that step. Everything before it can run at machine speed. The reviewer sees a decision, not a process.
A step that is reversible, internal and small rarely needs a person. A step that fails all three always does. Most real gates sit in between and depend on the data, which is why the rule belongs in policy rather than in the workflow's wiring.
A gate is only as good as the decision it supports. The reviewer needs:
If the reviewer has to open three systems to check the work, the gate becomes a formality.
Approve and reject are not enough. The most common real answer is “nearly, but not like this.” If the only options are yes and no, the reviewer either approves something they would have changed or rejects work that was almost right and starts over.
A third answer, send it back with a reason, fixes that. The reviewer writes what was wrong (“use the renewal date, not the start date”) and the work is redone with that guidance, under the same record. The correction becomes part of the history, and a correction that keeps coming up is a sign the instructions should change.
Every gate leaves evidence: what was proposed, who decided, what they decided and why. That record is what an auditor asks for, and it is how you know whether a gate still earns its place. A gate that is always approved unchanged is a candidate to remove. A gate that is often sent back is telling you the work upstream needs fixing.
In AIOS, each step carries a risk level the workflow's author sets, policy rules decide which steps wait for an expert's sign-off, and the reviewer can approve, reject or send the step back with guidance. Every decision is written to a record that cannot be edited. More on how that fits together: AI agent governance, how to audit an AI agent and the agent lifespan problem.
Keep reading
AIOS learns your processes and carries out the work across your existing systems—from customer follow-ups to invoice collection. Your team handles approvals and exceptions.