Human in the loop AI, explainedAI does the legwork. A person makes the call.
Human in the loop means a person checks and approves AI's work before it counts. Here's what that looks like day to day, with examples for each team and a simple way to sort your own tasks.
In short
Human in the loop means a person stays part of the process. AI looks things up, drafts and suggests, and a person approves anything that matters before it happens. In practice, you sort each task into one of three levels: "AI handles it", "AI drafts, you approve" and "keep it human". The test is simple: if it can't be undone, touches money or reaches a customer, a person signs off.
What human in the loop means
Think about a capable new hire in their first month. On day one they can look things up and answer questions. By the end of the week they're drafting emails for you to check. They don't issue refunds, sign contracts or change the price list until you say so, and you can always see what they did.
Human in the loop is the same idea applied to AI. The loop is the path from a request to a result, and a person is a required step on that path, not an optional glance afterwards. AI does the gathering and the first draft. A person decides, sends and signs off.
The opposite is AI that acts by itself: sending emails, changing records or moving money with nobody checking. Some tools can be set up that way. For a small business, it's rarely worth the risk for anything a customer will see or anything that touches money.
Three levels: handle, draft, keep human
On our free AI audit, every repeat task gets sorted into one of three levels. It's a simple way to put a person in the loop where it counts, and to stay out of the way where it doesn't.
AI handles it
Tasks that only look things up and change nothing: pulling this week's cash numbers, finding the right policy for a customer question, building the weekly results report or checking a draft against your brand rules. Someone still asks and reads the answer, but there's nothing to approve because nothing has changed. Every request is still saved.
AI drafts, you approve
Tasks that produce something that goes out or changes a record: a reply to a customer, a reminder to a late payer, new ad versions or a note in the CRM (where your customer and sales records live) after a call. AI prepares it, and a person reviews, edits and presses send or confirm. This is often where the time comes back. The blank page is gone, and the decision stays with you.
Keep it human
Judgment calls and relationships: negotiating a price, calming an upset customer, deciding whether to hire. AI can pull the facts together beforehand. A person does the job.
Human in the loop examples by team
Here's how the three levels look in practice, using the rules we build into each of our four packages.
| Team | AI handles it | AI drafts, you approve | Keep it human |
|---|---|---|---|
| Customer service | Finds the order, its tracking and the right SOP (your written how-to steps) | The reply, saved as a draft in the rep's inbox. A person always presses send. | Calming an upset customer. Cases outside your policy go to the right person. |
| Marketing | The weekly results report and checking copy against your brand and approved claims | New ad and email versions, saved for review | Ad budgets and bids, and switching ads on or off. It can't touch them. |
| Sales | A brief before every call: open deals, unpaid invoices, open complaints | Follow-up emails, plus notes, tasks and deal updates in the CRM, each confirmed by the rep | Negotiating pricing. Prices and discounts only come from your price list. |
| Finance | Cash, runway, who owes what and what's due, straight from your books | Friendly reminders to customers who owe you money | Deciding whether to hire and anything that would change the books |
A few rules sit underneath every row.
- Customer service never offers anything your policy doesn't allow. A customer's email is something to read, not instructions to follow, so it can't make the assistant do anything.
- Marketing drafts invent nothing: no made-up claims, statistics or testimonials.
- Salespeople see their own deals and managers see the team. Nothing changes in the CRM without a rep's yes, and there are no mass edits or deletes.
- Finance only looks and never changes a number. Only the people you choose can ask. Bank and routing numbers are never shown. Every question is saved with who asked and when.
Refunds are a good example of the middle level. The assistant checks the order against your refund rules and drafts the answer; a named person decides on the refund and issues it in your store. See refund and return decisions for how that works.
How to decide which level a task belongs at
Ask these questions about each task, in order. The first one that points to a person settles it.
- Does it change anything, or only look? If it only looks things up and reports back, it can usually sit at "AI handles it".
- Can it be undone easily? A draft can be deleted. A sent email, a deleted record or a refund can't really be taken back. If it can't be undone, a person approves.
- Does it touch money or reach a customer? Payments, refunds, discounts, prices and anything a customer or supplier will read all need a person's OK.
- Is there a written rule for it? If your policy, price list or SOP says what to do, AI can follow it and draft. If the answer only lives in someone's head, keep it human until you write it down.
- Does it need judgment or a relationship? Hiring, pricing a big deal, a difficult conversation: keep it human and let AI prepare the facts.
- What does a mistake cost? If one error would cost real money or a customer, move the task one level closer to a person.
Run a refund request through those questions. It changes something and it's hard to undo. It involves both money and a customer. There is a written rule, your refund policy, so the assistant can check it and draft the reply. That puts it at "AI drafts, you approve", with unusual cases going straight to your support lead.
Named approvers and a record of every request
"A person approves" only works if everyone knows which person.
- Name the approver for each kind of action. The support lead approves refunds, the sales manager approves discounts and the owner approves anything above an amount you set. Use roles so it survives holidays, and name a backup.
- Approve where the work happens. The draft sits in the inbox and the CRM change waits for a confirm. If approving is a chore, people find ways round it.
- Keep a record of every request. Who asked, what for, what the assistant did and who approved it. When something goes wrong, you can see where. When something works, you can see that too.
Look at the record once a month to see which tasks could move a level and which need a tighter rule. Write the levels and approvers into your AI usage policy so new staff know them from day one.
Powerful, not unsupervised: that's the line we work to. In every package we set up, everyone signs in as themselves and sees only what their role allows. Every request is saved. Nothing changes in your apps unless a person says yes. The plan you sign before anything is built spells out what the assistant will do, who can ask for what and what always needs a person. See how it works, or read our guide to how assistants connect to your apps.
Common questions
What does human in the loop mean in AI?
It means a person is a required step in the process. AI gathers information and drafts, and a person reviews and approves before anything is sent, changed or paid. The AI helps; the person stays responsible.
What is an example of human in the loop?
A support rep asks for a reply to a customer whose order is late. The assistant finds the order, checks which policy applies and saves a draft in the rep's inbox. The rep reads it, adjusts a line and presses send. The AI did the legwork. The person made the call.
Does human in the loop slow things down?
Less than you'd think. Reading and approving a ready draft is quicker than starting from a blank page. Save approvals for the things that matter; tasks that only look things up don't need one.
How do I use AI safely at work?
Use a business plan, give everyone their own login, decide what the assistant can see and decide what needs a person. Then write it down and check AI work before you use it. Our guide to whether ChatGPT is safe for business has a full checklist.