Draft and approve: make AI drafts quick to check
Key takeaways
- A person in the loop only helps if checking is quick. Show the draft, the facts it used and the one decision on one screen.
- Pull the facts from your own systems, not from the AI that wrote the draft, so a mistake can’t vouch for itself.
- Make rejecting as easy as approving, and keep a record of what you changed.
Short version: human in the loop means the AI prepares the work and a person approves it before it goes anywhere. It only works if checking is quick. Put three things on one screen: the exact draft, the facts it was built from, and the one decision you need to make. Then approving takes seconds, and people keep doing it properly instead of clicking yes without reading.
What is human in the loop AI?
It’s a plain idea with a technical-sounding name. The AI does the legwork, such as reading the message, looking up the order and writing a reply, and a person decides whether it goes out. Nothing reaches a customer, and no money moves, until someone has said yes.
For a small business it’s the sensible middle ground between letting the AI act alone and not using it at all. It’s the draft pile in act, draft or never.
Why approval turns into rubber-stamping
The weak point isn’t the AI. It’s the person approving. If every draft means opening the order system in another tab, finding the customer’s last three emails and remembering the returns window, checking takes longer than writing the reply yourself. So people stop checking. They skim the first line and click approve.
At that point you have the slowness of a person in the loop with none of the safety. The fix is to do the looking-up for the approver, not to tell them to try harder.
Put three things on one screen
- The draft. Exactly what will be sent, word for word, as the customer will see it. Not a summary of it.
- The facts it used. The order, what was bought, when it was delivered, the returns window it relied on, and the customer’s message. If the draft says “your order arrived on Tuesday”, the delivery record showing Tuesday should sit right beside it.
- The one decision. A plain question with a yes or no answer: Send a replacement planter? Refund this order in full? If a draft needs three decisions, it’s really three drafts.
With all three in view, the approver isn’t redoing the AI’s work. They’re checking it against the facts, which is quicker and harder to get wrong.
A human in the loop example
Say you sell garden planters online, and a customer writes: “The planter came with a chip on the rim. Can you send another?”
A slow draft is just a reply in a box: “So sorry about that, a replacement is on its way.” To check it, you have to go and find out whether this is a real order, whether it’s recent, and whether there’s a photo.
A quick draft looks like this:
- Decision: send a replacement planter, yes or no?
- Facts: order found, one planter, delivered last week, photo of the chip attached, no earlier claims from this customer.
- Draft reply: the exact message, saying sorry and confirming the replacement once you approve.
You look at the photo, glance at the date and click yes. The reply that goes out says what you decided, not what the AI guessed.
Why is it important to review AI generated content?
Because AI writes confidently whether or not it’s right. It can get a date wrong, promise something you don’t offer, or answer a slightly different question from the one the customer asked, all in a friendly tone that reads perfectly well. The tone is not the thing to check. The facts and the promise are.
That’s why the facts belong next to the draft. A reviewer can’t catch a wrong date they can’t see.
Take the facts from your systems, not from the draft
The facts on the approval screen should be pulled from your own systems, such as the order record, the delivery tracking and your written returns policy, not written by the AI. If the AI writes both the draft and the facts that support it, a mistake in one will happily agree with the other.
The same goes for numbers. Totals and dates should come from your systems, and there’s more on that in Don’t let AI do your sums.
Small habits that keep approvals fast
- Make rejecting as easy as approving. One click to reject, with a short reason. Otherwise people approve things they’d rather not, just to clear the queue.
- Let the approver edit before sending. Often the draft is nearly right. Fixing one sentence shouldn’t mean starting again.
- Flag what is unusual. A new customer, a large order or a second claim in a month deserves a visible note at the top, so it gets a slower look.
- Keep a record. Note what was approved, what was changed and what was rejected. A few weeks of that tells you which jobs the AI gets right every time.
That record is also how you decide, one job at a time, when a job no longer needs a person to approve every draft.
Want to know which jobs in your business are worth handing to AI? Grafto’s AI Tools Report works it out from a short questionnaire.
Frequently asked questions
What does human in the loop mean?
It means an AI prepares the work, such as a reply or a refund request, and a person approves it before anything is sent or any money moves. The AI saves the legwork; the person keeps the decision.
Why is it important to review AI generated content before using it?
AI writes confidently even when it’s wrong. It can get a date wrong, promise something you don’t offer or answer the wrong question. Checking the facts and the promise before it goes out catches that.
How do I make AI draft replies quicker to approve?
Put the exact draft, the facts it was based on and the single decision you need to make on one screen, with the facts pulled from your own systems. Make rejecting one click, and let the approver edit a line before sending.
Questions, or want a tool we don't have yet? Email hello@grafto.co.uk — a real person replies.