Human-in-the-loop AI workflow: asking a person, then carrying on with the answer
How to build an automated flow that stops at the one decision only a person can make, asks them on WhatsApp, Slack or email, and finishes the job with their answer.
Part of the guide AI employees
A human-in-the-loop AI workflow is an automated flow that does the routine work by itself, stops at the point where a person's judgment matters, asks that person a clear question, waits for the reply and then finishes the job using what they said. The skill is in choosing where to stop and in asking a question someone can answer in ten seconds from their phone.
This piece is for operations, sales and field teams who have automated part of a process and keep finding one step that still needs a human: picking which leads to visit, deciding how many jobs the crew can take, choosing between three sites. It covers the difference between asking and approving, how to word the question, what happens when nobody replies, a made-up example and how to set it up as an AI employee.
Asking is not the same as approving
Most tools offer an approval step: the flow pauses, someone looks at what is about to go out and says yes or no. That suits a message to a customer or a post on social media, where the work is finished and only needs a sign-off. An approval cannot carry information back, though. It cannot tell the flow which of five shops to visit or how many appointments fit into Thursday.
A question can. In xMap Flows, the Ask a person block takes a free answer, one of the choices you offer, a number, or a yes or no, checks the reply fits, and hands it to the blocks after it as answer, together with who answered and when. A Filter, a Sort and limit or an xMap LLM block can then act on it. Approval stays the right tool for the final yes before sending; the two often sit in the same flow.
Where to put the person
Put the question where the flow lacks something only your team knows, and as late as possible, so the person sees a short, finished shortlist rather than raw data. Good places:
- After ranking, before effort. The flow has found and scored twenty prospects; a manager picks which few get a visit or a written proposal.
- Before a limit that changes weekly. Crew size, van space, a budget for the week: a number only the supervisor knows on Monday morning.
- When the data is ambiguous. Two places with similar names at one address; the area manager knows which one is real.
- When a client must choose. An agency shows a client three rivals and asks which worries them most, then writes the brief about that one.
Avoid asking about everything. If the answer is the same nine times out of ten, write it as a rule instead and keep the question for the tenth.
Writing a question people answer quickly
- Ask for one thing. "Which of these should Ana visit this week?" is answerable; "thoughts on this list?" is not.
- Offer choices when you can. On WhatsApp up to a few choices arrive as buttons, which is the fastest reply there is.
- Show the list inside the question. The block can number the items for the person, so they can reply "2 and 5" or tap a button.
- Pick the channel they already watch: WhatsApp from xMap's number, a Slack channel (the first reply in the thread counts), or email with a link to a page where they answer, which suits clients without an account.
- Decide what happens on silence: stop the run, or carry on with a default answer after a wait of one hour to fourteen days.
A worked example
This example is made up. A catering supply firm in Mexico City sends one rep, Ana, to visit restaurants each Thursday. The sales lead used to pick the visits by scrolling a spreadsheet on Wednesday night. Now an AI employee does the groundwork:
- Every Wednesday at 16:00 a Schedule starts the flow.
- Nearby places finds restaurants within 5 km of the warehouse that opened in the last 6 months, independents only, with a phone or website.
- Only new drops the ones already offered in earlier weeks, and Sort and limit keeps the eight best rated.
- Ask a person sends one WhatsApp question per restaurant to the sales lead: "Visit {{name}} this week?", yes or no, waiting until 09:00 Thursday, carrying on with No if unanswered.
- A Filter keeps the Yes answers, Route planner puts them in visiting order from the warehouse, and Ana gets an email with the route and a map.
The sales lead now spends a minute tapping Yes or No on the bus instead of half an hour in a spreadsheet, and the restaurants he skipped come back only if the flow is set to treat them as new again after a while.
Mistakes to avoid
- Asking too early. A question about fifty unranked places gets ignored. Filter and rank first.
- No default. If the person is on holiday, the run waits until its limit and then stops. Choose a sensible default for routine questions.
- Free text where a choice would do. Free answers are harder to act on; a choice or a number feeds straight into a Filter or Sort.
- Asking someone outside the company about internal data. People outside your organization are asked by email only, and only by a run that sends for real.
- Hiding the answer. Put answered_by in the final email so everyone knows whose call it was.
Running it as an AI employee
Templates such as New outlets, graded and routed to each rep already do the finding, grading and routing; add Ask a person between the grading and the route and the employee checks in with a person before planning the day. For outbound work, Full outbound engine shows the companion pattern, an approval before anything is sent. The workflow automation guide covers how triggers, logic and actions fit together, and the help article on asking a person in the middle of a run lists every setting.
Once it runs, the person asked also gets a notification in the app and can answer there; the flow's owner is told when the answer arrives, and the run carries on within moments.
Questions
Does the person need an xMap Flows account to answer?
No. Someone outside your organization answers through the link in an email. Teammates can answer by email, on WhatsApp from xMap's number once their number is confirmed, or in Slack.
What if the reply does not fit the question?
The answer is checked against what you asked for: a word where a number was expected is asked again.
Can I ask about each item separately?
Yes. Ask once for the whole run, or once per item for up to 20 items; each item then carries on with its own answer.
How long will the flow wait?
From one hour to fourteen days, as you set it. After that it stops or carries on with your default answer.
Is asking a person charged?
No, the Ask a person block itself is free; the other blocks in the flow are charged as usual.
Templates that do this
Read next
More on ai employees and automation · Every article · Templates