Blog · Human in the loop

Human approval for AI: five moments a person must say yes

AI can prepare almost anything. Here are the five moments where a person must approve before it continues, and who is responsible when it goes wrong.

· 7 min read · Written by Butler @ yalla ya! · Reviewed by Ido Kobelkowsky

Cover: the words Human approval for AI, a pause symbol and five numbered check marks

An AI assistant can now write the code, find the customers whose payment failed, draft the apology and prepare the refund, all before you finish your coffee. The useful question is no longer whether it can. It is where it has to stop and wait for a person. Human approval for AI does not mean reading everything the machine does. It means naming in advance the few moments where there is no way back, and putting a person with a name at each of them.

This guide gives you those moments, what the person should look at before saying yes, and who answers when something goes wrong anyway.

Where does human approval for AI really matter?

Short answer: where the action cannot be undone in minutes, or where it leaves your business and reaches someone else. Everything else can run on its own and report afterwards.

We use three questions. Can this be taken back quickly and completely? Does it touch money, customer data or people outside the company? If it goes wrong, will someone have to explain it to a customer? A “no” to the first, or a “yes” to either of the others, is enough to stop and ask.

The security community says the same in its own words. OWASP, the open project behind the best-known lists of software risks, calls the problem “excessive agency”: an AI system that was given too many functions, too many permissions or too much autonomy. One of its recommendations is to have a human approve high-impact actions before they are taken.

The opposite mistake is just as real. Ask a person to approve everything and they stop reading. The European AI Act even has a name for the habit of trusting what the system produced because it is usually right: “automation bias”. So the list has to be short. Ours has five lines.

AI prepares the change Can it be undone in minutes? AI continues and reports after AI stops and an alert asks A named person says yes or no Written down: who, what, when Yes No These wait for a yes 1. Going live 2. Payments and refunds 3. Deleting data 4. Messages to customers 5. Access and keys
The rule in one picture: AI continues alone only when the action can be undone. Otherwise it stops and a named person decides.

The five moments where AI waits for a yes

1. Going live

A change on a test copy of your site costs nothing if it is wrong. The same change in front of customers can stop sales in the middle of a campaign. AI can build the change, test it and show it to you. A person decides that it goes live, and when: not on Friday evening, not during the sale. Before saying yes, ask two things: have I seen it working, and how do we go back if it fails? That is the deal on the websites and stores we build with AI: you see the progress and you approve what goes live.

2. Payments, refunds and prices

Money moves in one direction more easily than in the other. Stripe, a payment processor many online stores use, is a good example. Its documentation on refunds says that the processing fees of the original payment are not returned when you refund, that only some card refunds can be cancelled and only for a short time, and that the customer may see the money 5 to 10 business days later. A refund sent by mistake is not fixed with an undo button.

So AI can find the renewals that failed last night, work out why and prepare the retry. Charging, refunding, changing a price or switching on a discount waits for a person who sees the exact amounts and the exact customers.

3. Deleting data

Deleting is the purest case: the only undo is a backup. Before anyone approves a cleanup of old orders, duplicate contacts in a CRM or files on a server, the question is not “do we still need this?” but “when was a backup last restored as a test?” A backup nobody has tried to restore is a hope, not a way back. This is why, on the servers we run for clients, backups are tested by restoring them. A deletion proposed by AI stays a proposal until a person approves it.

4. Messages to customers

An e-mail that has been sent cannot be called back, and a message to your whole list is thousands of chances to be wrong: the wrong name, the wrong price, a promise the business cannot keep. Our rule for messages that go out in a client’s name is simple: AI drafts, and a person edits and approves before anything is sent. An answer to one customer in a chat can follow rules you agreed in advance. A message to everyone waits for a yes.

5. Access and keys

The last moment is the quietest: who can get in. Passwords, user accounts, permissions, and keys, the secret codes that let one program act inside another account, such as your payment processor or your server. An assistant that can give itself or anyone else more access can do everything else on this list without asking. So new users, wider permissions and new keys are approved by a person, and the accounts stay in your name. It is part of how we keep a site or an app secure.

The five approvals in one table

MomentWhat AI preparesWhat the person checks before yesWhy it waits
Going liveThe change, tested on a copyThat it works, and the way backCustomers see it at once
Payments, refunds, pricesThe list, the amounts, the reasonExact amounts and exact customersMoney and fees do not come back by themselves
Deleting dataWhat to delete and whyA backup that was restored as a testNo backup, no undo
Messages to customersThe draft and the list of recipientsEvery word, and who receives itSent is sent
Access and keysThe request and its reasonWho gets what, and for how longAccess opens every other door

Who is responsible when AI gets it wrong?

Short answer: the business, through the person who said yes. Not the tool.

The law is moving the same way. The European AI Act (Article 14) requires that high-risk AI systems be built so that people can oversee them, decide not to use their output, override it, and stop the system. That article is written for high-risk systems, not for a store’s assistant, and this is not legal advice: if you work in a regulated field, ask a lawyer. But the direction is clear. A person must be able to stop the machine, and a person answers for what it did.

For an approval to mean something, three things have to be true. The approver has a name: “the team” approves nothing. They see exactly what will happen: these twelve refunds for this amount, not “process the refunds”. And it is written down: who said yes, to what, and when. On our own site we say it in one line: AI does the work, people are responsible for it.

This is also where general AI tools fall short. Chat assistants and coding assistants do what they are asked and stop at the code. As we wrote in what none of the AI tools for business connects, they do not link your business numbers, your decisions and your code, so they cannot know which action needs a yes from whom. We built Butler, the workspace we set up for each client, around that connection: an alert asks a person to act, a routine reports instead of acting silently, and every warning and every action is written down.

Where the person came in

We run about 20 projects in Butler: stores, CRMs, mobile apps, servers. The thread through all of them is the one in this guide. AI prepares, and a person decides what goes live, what moves money and who gets access. Where exactly to draw the line is still judgement, not a formula, and we keep adjusting it: ask too often and people stop reading, ask too rarely and you find out afterwards.

This article is an example. It was written with AI, under rules the owner of yalla ya! set in advance: a guide like this one may be published, with the owner told the same minute, while any story from a client’s project waits for the owner’s approval.

What to do this week

Write your own five lines. Next to each one put a name, and what that person must see before saying yes. If you cannot name the person, AI should not be doing that action yet. Then check that your tools really stop at those lines, and do not just promise to.

If you want a second pair of eyes, bring the list to a free 30-minute consultation and project review. We will go through where your site, store or app lets AI continue alone, and where it should wait.

Keep reading

All articles
Yalla, let's go

Tell us what you wish AI would do for you. We'll take it from there.

Or just tell us what is slowing you down. Book a free 30-minute consultation and project review, and we will show you what an AI-assisted team would look like for your business.

Skip to content