Human in the loop: when AI may act alone and when a person confirms
“AI does not send, delete or pay without a person confirming” sounds good, but in practice you have to decide exactly what gets confirmed, by whom, and for how long. Here is the table I use.
Actions carry different weight
“Do we trust AI” is the wrong question. AI that reads and summarises can be wrong and nothing happens: someone reads a poor summary and checks the source. AI that sends an email to a customer or transfers money is wrong with consequences. So I set the rules by type of action, not by how “intelligent” the model is.
The decision table
| Action | Example | At the start | After a trial period | Never alone |
|---|---|---|---|---|
| Read | Find, summarise, compare, answer from documents | Alone, citing the source | Alone | — |
| Draft | An email reply, a note, a product description, a proposed category | Alone; the draft waits for a person | Alone | — |
| Send | An email to a customer, a chat message, a publication | A person approves each one | Exceptions: narrow categories with a high agreement rate | Proposals, disputes, first contact |
| Change | A CRM field, a ticket status, a calendar slot | A person approves each one | Exceptions: one field, reversible, logged | Customer contact details, prices, contracts |
| Delete | A record, a file, a contact, a message | A person, always | A person, always; AI may propose a list | Everything. Archive instead of delete |
| Pay | A transfer, an order, a limit change | A person, always | A person, always; AI prepares | Everything |
The two right-hand columns matter more than they look. “Never alone” is a list that does not change after a successful trial. Deletions and payments stay with a person not because AI gets them wrong more often, but because mistakes there are irreversible or expensive. AI can prepare the list of records to delete and the batch of transfers. A person clicks.
What “a person approves” looks like in practice
Approval has to be convenient, otherwise people start clicking “yes” without reading, and then there is no human in the loop, only a delay. A few rules:
- The draft appears where the person already works: in the inbox, in the CRM, on the team channel. Not in a separate tool.
- The source sits next to the draft: the document passage, the record, the table. Checking takes seconds.
- Rejection has a “why” field. Those answers are how the instructions get improved.
- The queue has an owner and a response time. A draft that waits three days is worse than no draft.
How to graduate from “confirm everything” to “confirm exceptions”
- Measure agreement. For the first month every AI decision is approved or corrected. We count the percentage of cases where the person changed nothing, separately for each category. Not “90% overall”, but “appointment confirmations 99%, complaints 70%”.
- Pick categories. Those with very high agreement and a low cost of error (confirming an appointment, answering about opening hours, routing a ticket to a department) are candidates for exceptions-only. Complaints are not, even at 95%.
- Define the exception. What still goes to a person: model confidence below a threshold, an amount above a threshold, a new customer, the words “complaint”, “lawyer”, “cancel”, anything outside the topic list. The exception is written down, not “if something looks odd”.
- Switch it on for one category. Not for all the candidates at once. Two weeks of observation with spot checks: a person reads every tenth message sent.
- Keep a brake. One switch that returns to “confirm everything”. Available to a manager, no developer needed. Used after any serious mistake until it is explained.
A worked example
A hypothetical example: a services company with an inbox of customer questions. Month one: AI sorts the emails and writes drafts, a person approves everything. Sorting agreement 96%, draft agreement in the “appointment confirmation” category 98%, in the “price enquiry” category 60%. Month two: sorting runs alone, except for emails where the model has low confidence; appointment confirmations go out alone if the customer is in the CRM and the slot exists in the calendar; everything else is still approved by a person. After a quarter the “alone” list has five categories and a person reads the log once a week. Price enquiries stay with a person for good, because that is where a mistake costs money.
If you are building an automation in which AI makes one decision and you want an exceptions queue from day one, this is how an automation with one AI step goes.