Human in the Loop: How to Approve What Your AI Employee Does (2026)

Key Takeaways

Human in the loop is a sentence, not a system. Adding "show me before sending" to a task is all it takes to turn any action into a reviewable draft.
Sort actions by blast radius, not by effort. Anything that leaves your company (emails, customer replies, ad changes, posts) earns a review. Internal summaries and research can run free.
Use the trust ladder. Start a new workflow with drafts only, approve a few rounds, then loosen the rule once the output is boringly correct.
Review happens where you already work. The draft lands in your Slack thread, you reply with a fix or a go-ahead, and the AI employee takes it from there.
Recurring tasks can carry a standing checkpoint. "Every Friday, draft the client update and wait for my OK" runs forever without ever sending on its own.
Oversight is what makes delegation possible. The teams that hand over the most work are the ones with clear checkpoints, because they never have to wonder what happened.

The fear is specific: it did something I did not approve

Nobody worries that their AI employee will summarize a channel badly. The worry is sharper than that: it emails a customer something wrong, changes an ad budget, or posts a half-baked update to the whole company, and you find out after the fact. That worry is healthy. An AI employee that can act in your real tools deserves the same thing a capable new hire gets in their first month: a review step on anything that matters.

The good news is that the review step is not a settings page or an approval platform you have to build. With an AI employee that lives in Slack, human in the loop is mostly a habit of phrasing plus a couple of standing rules. Stanford's AI Index has tracked a steady rise in publicly reported AI incidents year over year, and the pattern behind many of them is the same: a system acting with more freedom than the stakes justified. This post is about matching freedom to stakes, one sentence at a time.

What does human in the loop actually mean here?

Answer first: it means consequential actions pause for your explicit go-ahead, while everything else runs without you. The AI employee does the work, produces the finished thing, and holds it as a draft until a person says send.

That definition matters because "human in the loop" often gets treated as all or nothing. Either you approve every step, which is slower than doing the work yourself, or you approve nothing and hope. Neither is how you would manage a person. You would let a new hire read everything, draft anything, and you would look at outbound work before it ships, then stop looking once they have proven out. Anthropic's engineering guidance on building effective agents lands in the same place: agents are most dependable on clearly bounded tasks with a human reviewing the consequential steps, not turned loose end to end.

So the practical question is not "should a human be in the loop" but "which actions belong in the loop." That comes down to blast radius.

Sort actions by blast radius

A simple test: if this action goes wrong, who sees it? Use that to sort everything you delegate into three buckets.

No review needed: work that stays between you and the AI employee. Research, summaries, analysis, a digest of yesterday's support emails, a table of last month's Stripe numbers. Wrong is cheap here, you just ask again.
Review the first few times: internal but visible work. A weekly post to #team-updates, a Notion page the whole company reads, a Jira ticket filed on an engineer's behalf. Mistakes are embarrassing, not damaging.
Always review: anything that leaves the building or moves money. Customer emails from Gmail, replies to leads in HubSpot, budget changes in Google Ads, invoices, anything a customer or the public will see.

Most people find that the third bucket is smaller than they feared and the first bucket is bigger. The bulk of delegated work is reading and writing drafts, which is exactly the work that needs no oversight at all. The loop only needs to exist where the stakes are real.

The one-sentence checkpoint

Here is the entire mechanism. Compare:

@Viktor follow up with the leads that went quiet in HubSpot this week.

@Viktor draft follow-ups for HubSpot leads with no reply in 7 days. Post the drafts here and wait for my OK before sending anything.

The second version is human in the loop, implemented in one sentence. The drafts arrive in your thread, you read them in thirty seconds, reply "send all" or "fix the tone on the second one," and it proceeds. Nothing left your workspace until you said so.

Useful phrasings to keep in your pocket:

"Show me before sending."
"Draft only, do not send."
"Wait for my OK before making any changes."
"Post the plan first, act after I approve."

You can also set the boundary below the task level. If you never want a tool touched without you, connect it read-only, which we covered in How to Control What Your AI Employee Can Access. Read-only access plus draft-for-approval phrasing covers almost every oversight need a small team has.

Review where you already work

The reason this works in practice is that the review step costs almost nothing. The draft is not in a queue in some approval tool you forgot the login for. It is a message in the Slack thread where you asked for the work, and approving it is a reply.

That thread is also where you improve the work, not just gate it. If the draft is close but not right, say what is wrong: "shorter, and do not mention the delay." The AI employee revises in place and shows you again. This draft-review-revise loop is the same rhythm you have with a person, and it is why the checkpoint makes delegation faster rather than slower: you spend your time on judgment, not production. If your first drafts keep missing, the fix is usually in the brief, and we wrote up the format that works in How to Write Tasks Your AI Employee Actually Nails.

Climb the trust ladder

Approval on everything forever is not the goal. The goal is to earn your way to autonomy the same way a new hire does.

Weeks one and two: drafts only. Every outbound action pauses for review. You are learning what its judgment looks like.
Once the drafts are boringly correct: approve by exception. "Send unless a deal is worth flagging to me first." You review the edge cases, not the routine.
For the proven, low-stakes routines: let it run. The Monday metrics post that has been right eleven weeks straight does not need a twelfth review.

The signal to climb a rung is repetition without surprises. The signal to climb back down is any surprise at all. Because corrections you make in the thread persist as memory, each round of review teaches the AI employee your standards, which is what makes rung three reachable instead of theoretical.

Standing checkpoints for recurring work

Recurring tasks are where a checkpoint earns its keep, because the task runs when you are not watching. Build the approval into the schedule itself:

@Viktor every Friday at 3pm, draft the weekly client update from #project-alpha and the HubSpot deal notes. Post the draft here and wait for my approval before emailing it.

This runs every week forever, and it will never send without you. You get the draft at 3pm, you approve it at 3:02, done. The same pattern fits ad reporting, invoice reminders, and outbound follow-ups: the collection and drafting are automatic, the go-ahead is yours. For how to set these up well, see How to Set Up a Recurring Task for Your AI Employee.

Frequently Asked Questions

Does reviewing everything defeat the point of an AI employee?

No, because you only review the small slice of work with real blast radius. The research, summaries, and drafting that make up most delegated work run without you. Reading a finished draft takes a fraction of the time producing it would have, so you keep the leverage and add the control.

How do I make sure it never sends anything without me?

Two layers. In the task, say "draft only, wait for my OK." At the tool level, connect the account read-only so sending is not possible at all. Use the second one for any tool where a mistake would reach customers.

What should always require approval?

Anything that leaves your company or moves money: customer emails, replies to leads, social posts, ad budget changes, invoices, payments. Internal, reversible work like summaries and research does not need a gate.

Can approval be part of a recurring task?

Yes. Phrase the schedule with the checkpoint included, like "every Friday, draft the update and wait for my approval before sending." The task runs on schedule forever and still never acts on the outside world without a human reply.

When can I stop reviewing a task?

When the drafts have been correct enough times that reviewing them feels like a formality, and the downside of a miss is small. Loosen gradually, keep approval on high-stakes actions permanently, and tighten again the moment something surprises you.

What happens if I reject a draft?

You reply in the thread with what to change, and the AI employee revises and shows you again. Rejections are not wasted work, they are training: the correction sticks as memory, so the next draft starts from your standard.

Control is the feature, not the tax

An AI employee you cannot supervise is one you will never hand real work to. The loop is what makes the handoff safe: sort actions by blast radius, put a one-sentence checkpoint on the ones that matter, review in the thread where you already work, and promote proven routines to run on their own. Teams that work this way delegate more, not less, because every checkpoint they keep is a worry they no longer carry.

Add Viktor to your workspace and keep every send behind your OK