# How can I let AI act for my business and still check its work?

Human in the loop AI means the software drafts and a person decides. Here is how drafts wait for your yes, each with its reason beside it.

Published: 2026-10-12  
Canonical: https://felesh.ai/blog/how-can-i-let-ai-act-for-my-business-and-still-check-its-work

## What human in the loop AI means

You have seen the pitch. Software that drafts the reply, updates the record and tells you afterwards. The demo looks effortless. Then you imagine it running for a month while you are busy with the work itself, and you wonder what has been said in your name.

The phrase "human in the loop AI" sounds technical, but the idea is plain. The software does the drafting and the first pass. A person decides whether the draft stands.

Picture the ordinary scene. A customer writes with a complaint and a question about timing. You know the answer, but you are on your way out of the office. That is exactly the moment you want to see the words before they count.

So the loop is not a limitation. It is the point. The software carries the effort, and you keep the judgement.

The answer is to let the software draft and keep the yes for yourself.

## Why checking is the real work

Unchecked automation fails in ordinary ways. The first is a wrong promise. A customer is pleased until it turns out to be wrong.

Tone is the second. A short reply that suits one customer reads as cold to another. The software has not had the conversation. It knows the words on the page, and you know the person.

The third is the record. A date, a note or a next step changes, and later you cannot tell why the plan moved. You trust the record less, and then you stop using it.

Checking is not a chore added to the work. It is the work. Whether a promise is right, whether a tone fits, whether a note is accurate: that is the part only you can do. Software that removes the check removes your judgement along with it.

## Drafts that wait for your yes

Here is how it works in Felesh. You finish a conversation with a customer, on the phone or in person. You type one sentence about what happened. Felesh takes that sentence and drafts three things: a note about the conversation, the promise with its date if you gave one, and the next step.

Each draft appears with its reason beside it. You read the draft, you read the reason, and then you decide.

The draft waits for your yes.

Small things are handled differently. Adding a task to a list does not call for judgement, since it only records something to do. So a task is added directly, with no approval step.

The rhythm of the day is short. You talk to people. You type a sentence about each conversation. You look at what Felesh has drafted, approve what is right and leave what is not. The drafting is done for you. The deciding is still yours.

## Seeing why a draft was written

Checking is easier when you can see why a draft says what it says. That is why each one carries its reason.

If the reason does not match what you meant, you know before you say yes.

You can also ask questions in your own words. "What did we agree with the supplier last week?" Felesh answers from your own records. You can check the answer against what you remember.

This works because Felesh answers from what it holds: what you have typed, what you have taught it and the contacts you have brought in. You can bring contacts and companies in from one Excel or CSV file. The import is reviewed, so you see the column mapping and the rows before anything is saved.

## What it will not decide for you

There are questions Felesh will not guess at. In the customer inbox it answers from what you have taught it, through documents you upload and instructions you write. If a message falls outside that material, Felesh does not make something up to be helpful. The message goes to a waiting list for you, with the reason it could not answer. Nobody is notified, and when a person joins the conversation, Felesh stays quiet.

You would rather be told "this one needs you" than be handed a confident answer built on nothing. It is also the same principle at work again: the software handles what it can support, and the person handles the rest.

Felesh does not run the business. It keeps the business known, and it does the work you approve. Strategy, promises and the judgement about what matters stay with you. That is the arrangement that makes sense when you are the one who answers for the outcome.

## What to look for in any tool

If you are comparing tools, ask these four questions of each.

- Does it show you the draft before anything counts?
- Does it say why it wrote what it wrote?
- When it does not know, does it say so, or does it fill the gap?
- Do you decide, or does it decide and tell you later?

A tool that answers well to all four lets you hand over real work and still check it. A tool that cannot show its reasoning asks you to trust it, and trust is not a control.

## Start with one draft

The easiest way to see the loop is to try it on one conversation. After your next call, type one sentence about what happened. Felesh drafts the note, the promise and the next step, each with its reason, and waits for your yes.

See how drafts and approvals work, and how the inbox [admits what it does not know](/features#no-made-up-answers).
