On this page
What human in the loop AI means
Human in the loop AI is a system where a person checks the AI's work at the points that matter, before it reaches anyone outside the team.
In my work that comes down to one rule: the AI drafts and prepares, and a named person decides and sends. The AI does the reading, the sorting and the first drafts, which is where most of the hours go. The person keeps the calls that need judgment: what to tell a customer, what to charge, who to hire, how hard to chase a late payment.
The other design is an AI that acts alone: it replies, books, pays or posts with nobody looking. That is fine for a few small jobs inside the business. For anything a customer reads or anything that moves money, it means the first confident mistake reaches the customer before anyone on your team has seen it.
How my own desk runs
My own practice runs on an AI system I built in Claude Code. Each morning it reads my calendars, inboxes and project files and hands me one ordered list: the replies I owe, already drafted in my voice, the people waiting on me, and what is about to slip.
It sorts the drafts before I see them.
Each one is either ready to send or needs a check first. A draft that quotes a date, a figure or a promise goes in the check-first pile, with the thing to verify written next to it.
It never sends.
I read each draft, change what is wrong and press Send myself.
It checks its work against my files.
Prices against my price list, claims against my proof points, tone against my voice guide. Anything that does not match comes to me flagged.
It remembers.
Decisions, open items and who is waiting on whom are written to files, so the next morning starts where this one ended.
The AI Chief of Staff Starter Kit is that workspace as a template you fill in with your own business. The Claude skills for small business are the routines inside it, and each one drafts and stops for you to send.
Where the person goes
Put the person at the last step before the work leaves the business, or before it changes something other people rely on. In the systems I build, that is four places.
- Person sends
Anything a customer reads.
Replies, quotes, follow-ups, the final notice on a late invoice. The AI drafts them, and a person reads and sends.
- Person decides
Anything that moves money.
Refunds, credits, payments and wires. At a private lender, every outbound wire has a two-person verification by callback.
- Person decides
Anything that changes a record others rely on.
A price, the schedule, a customer's account. The system can propose the change, and a person makes it.
- Person decides
Anything new.
The first time a situation comes up, it goes to a person. Once there is a written rule for it, the system can prepare it the same way next time.
- Runs alone
Work only your team sees.
Sorting the inbox, logging a call, building the morning list, reminding someone a task is due. A mistake there costs a minute, and someone on the team sees it anyway.
One kind of customer message can go out without a person pressing Send: a message you approved in advance, like the text a missed caller gets back. You made that decision once, when you signed off the template.
Who does what, job by job.
Six routine jobs in a growing business, and where the line sits between the system and the person on your team. A made-up team, so the names are roles.
| The job | The system | The person |
|---|---|---|
| A new inquiry | Drafts the reply in your voice | SLSales lead reads it and presses Send |
| A missed call | Texts back from a template you approved and puts the callback on a list | CCoordinator confirms the callback |
| Invoices past due | Sends the first friendly reminder on schedule | YYou OK the final notice |
| The weekly numbers | Builds the report | OLOperations lead reviews it before it reaches you |
| A new hire | Sends the paperwork and the first-week checklist | YYou make the hire |
| A price change | Nothing | YYou always |
One job stays entirely yours. In the other five, the system does the routine part and a named person makes the call.
I write these limits down for each system before it goes live: what it may draft, what it may do alone, and what always comes to a person.
How I set one up
- Find the job that waits on you most. The Operations Leak Audit ranks them, and the first system gets built for the top one.
- Write the procedure before the system. A named owner, the steps and what done looks like. The free SOP template has the fields.
- Write the limits. What the system may draft, what it may do alone and what always comes to a person.
- Teach it your voice. Every draft starts from a brand voice guide, so the person edits facts instead of rewriting the tone.
- Make approval one step. The draft lands where the person already works, with the thing to check written next to it.
- Watch what the person changes. When the same fix keeps coming up, change the draft or the rule. The person stays on anything a customer reads.
When the AI has to answer live
Some jobs cannot wait for a person to press Send. A phone receptionist answering a call is the obvious one. There the person's check moves earlier: you test the system on hard calls before any customer hears it, you write down what it may promise, and someone reads what it said while it settles in.
My AI receptionist guide has the test calls I run before one goes live and how I grade them.
Questions
What is human in the loop AI?
An AI system where a person reviews and approves the work at the points that matter, before it reaches a customer, moves money or changes a record. In the systems I build, the AI drafts and prepares, and a named person decides and sends.
Does a person in the loop slow everything down?
Very little. The slow part of most routine work is the reading, the sorting and the first draft, and the AI does all three. Reading a finished draft and pressing Send is the quick part.
Does the AI ever send anything on its own?
Only messages you approved in advance, like the text a missed caller gets back, which comes from a template you signed off. Everything else a customer reads waits for a person.
Which jobs can run without a person?
Work that only your own team sees: sorting the inbox, logging calls, building the morning list and internal reminders. A mistake there is cheap, and someone on the team sees it anyway.
What tools do you build it in?
The tools your team already uses for email, chat, scheduling and the books, connected where they need to talk to each other. My own desk runs on Claude Code. The written limits on what the system may do matter more than the tool.
Do I need to be technical?
No. You need to know which decisions are yours. I build the system, write its limits with you and train the person who will run it.
How do I start?
With the Operations Leak Audit: a 45-minute working session on one ordinary week of your business. You leave with a one-page map of where the work waits on you and the first system I would build.
Where this fits in my work
I build AI systems and automations for growing businesses in any industry, and I work beside founders as their fractional COO. The systems take the routine work off your desk, and the decisions stay with you.
