AI & Technology

Where Human Approval Belongs in AI Agent Workflows

By Hisham Asghar

One of the first agents our team at Octacer put into a live business workflow had six tools. A month later the team wanted thirty. Every new tool restarted the same argument in the team channel: does a person need to approve this, or can the agent just do it?

Teams tend to land on one of two bad answers. Some gate everything, and within a few weeks reviewers are approving requests the way people accept cookie banners. Others gate nothing and find out what the agent can do from an angry customer.

What follows is how we decide where sign-off goes, how to stop approvals becoming the bottleneck, and what to log so you can reconstruct a decision months later.

Start With the Action

People ask whether an agent is “ready” to run on its own. We don’t think that question has an answer. The same agent can be perfectly safe drafting a reply and dangerous the moment it can issue a refund.

So we go tool by tool and ask a few blunt questions. Can this be undone, and how easily? Does it reach anyone outside the company? Does it move money or change someone’s access?

The question people skip is how many records one call can touch. A single wrong update is a support ticket. The same mistake applied to an entire CRM segment is an incident with a postmortem.

The OWASP Gen AI Security Project’s entry on Excessive Agency describes this problem well. It traces the damage to agents having too much functionality, too many permissions, or too much autonomy, and it lists human approval of high-impact actions among its mitigations. We’d stress the order, though. Strip tools and permissions first, then use approvals for whatever risk is left.

What We Gate by Default

Anything that goes out under a person’s or the company’s name is at the top of our list. Customer emails, chat replies, social posts. After that comes money above a small limit, whether it’s a refund, a credit, or a discount.

Deletes and bulk edits get a gate, and so do changes to user permissions or account status. Production configuration too. Contract terms and employee records sound obvious, but we’ve seen them slip through because the tool was labeled something harmless like “update document.”

What Usually Runs Without One

Reading and searching data the agent is already allowed to see. Drafting text a person will review and send. Tagging, routing, summarizing, creating an internal task, or a single-field update that’s easy to reverse.

These still get logged. And the line moves with context. A tag that triggers an automated email to a customer isn’t just a tag anymore, so follow what an action sets off downstream before calling it safe.

Thresholds Beat Yes-or-No Rules

The pattern we keep coming back to is a limit. The agent can refund small amounts on its own, and anything over the limit waits for a person. It can update a few records per call, while bulk changes need approval.

Limits let you start strict and loosen with evidence. If reviewers have approved nearly every small refund for a couple of months, and spot checks agree with them, raise the limit. If they keep rejecting one kind of request, that’s useful too, and it usually points at the prompt or the data rather than the reviewer.

Keeping Approvals Fast

Slow approvals kill adoption faster than bad output does. If a reviewer has to open three tabs to work out what the agent wants, they’ll ignore the queue or rubber-stamp it.

Where the request shows up matters more than people expect. Send it to the chat channel, inbox, or ticket queue the reviewer already lives in. A separate approval dashboard sounds tidy, and in practice nobody opens it.

Then there’s what the reviewer sees. They need the exact email text or the exact field change, the record it applies to, a one-line reason from the agent, and links to the data it used. Nobody should have to read a full transcript to approve a refund.

One small feature makes a big difference, and that’s an edit button. A lot of the time the agent is mostly right and the reviewer wants to change one sentence. With only approve and reject, you get rejections that throw away good work.

Routing is the other half. Refunds should go to someone who can actually approve refunds, and access changes to whoever owns that system. A single generic “AI reviewer” turns into a bottleneck and, worse, a person approving things outside their expertise.

Finally, decide what happens when nobody answers. For nearly everything, the safe default is to do nothing and flag it. An agent should never proceed because a timer ran out.

When Approval Turns Into Theater

The awkward truth is that people who approve the same kind of request a hundred times start trusting the system more than they should. The EU AI Act names this directly. Article 14 on human oversight calls it automation bias and expects the people overseeing high-risk systems to stay aware of it and to be able to override or stop the system.

Your use case may not fall under that law. The behavior shows up anyway.

What helps is checking approved actions every week, not only the rejected ones. Some teams also seed the queue now and then with an item that should clearly be rejected, as an open training exercise, to see whether it gets caught. And watch approval rates by action type, because a gate that approves almost everything for months is either unnecessary or not being read.

Log What the Reviewer Saw

When something goes wrong, you’ll want three answers fast. What did the agent do? Who approved it? What exactly were they looking at when they clicked?

That third one gets missed. A log line that says “approved by a reviewer at 14:02” is useless if you can’t see the version of the email they approved, since agents regenerate text and the final send might differ.

On Octacer projects, we log the full proposed payload for every gated action, along with the inputs and sources the agent used and its stated reason. Add the reviewer’s identity, decision, and timestamp, any edits they made, and what actually executed along with the target system’s response. Keep ungated actions in the same log so the sequence makes sense, and put it somewhere the agent can’t write to.

The NIST AI Risk Management Framework is a good reference if you want a broader structure. Its four functions, Govern, Map, Measure, and Manage, are a sensible way to organize this work. The approval log feeds the last two directly, since it’s your evidence for which gates to tighten and which to drop.

Revisit the Gates

We’d review approval rules monthly at first. Look for gates that never reject, actions that caused trouble without a gate, and tools added since the last review that nobody classified.

That last case is the dangerous one. Every new tool should arrive with an explicit decision about whether it’s gated, ungated, or limited. If nobody made the call, it defaults to gated.

The Trade-Off

We’ll be honest about the cost. Every gate adds delay, and part of the value of an agent is that it works at 2 a.m. when nobody’s around to approve anything. You’re trading speed for safety, and you won’t get the balance right on day one.

Start stricter than feels necessary and let the logs argue you down. Loosening a gate is an easy conversation. Explaining why an agent emailed the wrong price list to a few hundred customers is not.

Related Articles

Back to top button