Skip to content
BitByteCore ConnectBook a demo
Open the product menu

Guide · AI agents · For the owner whose team answers the same questions every day

Meta's AI agent on your WhatsApp support line: what Meta says it may do alone and what goes to a person

Meta has published a guide to putting its AI agent for business first in line on a WhatsApp support number. Meta's point is that most support messages are the same questions every day, and the agent is worth having if it answers those consistently and gets the rest to a person quickly.

Published 6 October 2026

What Meta says a support agent is for

Meta separates support from selling. Nobody is browsing, the customer usually has a problem, and success is measured by how many cases are resolved and how quickly the rest reach a person, not by sales. Meta lists what the agent handles:

Policy questions
Answered from the business's own written policies, not from general knowledge.
The customer's orders
Looked up from the conversation, without asking for an order number first. This needs a connection to the business's order system; without one, the agent can give only general answers.
Starting a return
Including the reason and the return label, when the item is within the return policy.
A ticket for a person
Raised with the conversation attached, when the agent cannot resolve the case.
Handover
To a person for complaints, damage, and anything that involves deciding on a refund.

What the agent answers from

An instruction with a named list of handover cases
Meta calls this list the most important part. 'Hand over when appropriate' is not something an agent can follow consistently; a list such as damage, missing item, payment dispute, refund decision, legal or safety issue, or asking twice for a person behaves predictably.
FAQs from the business's own top reasons for contact
Meta suggests taking the most common reasons customers got in touch last quarter and writing one entry per reason, in the words customers actually used. One question per entry: an entry covering returns, exchanges and refunds together matches all three poorly.
A plain yes to 'can I speak to a person?'
With the hours. Meta warns that a customer who asks for a person and gets deflected escalates in tone, and the case is harder for whoever picks it up.
Documents
Terms of sale, a guarantee booklet, care and assembly guides: anything too long for an FAQ entry.
The help pages of the website
Kept to the help pages. Meta notes that reading a whole shop site pulls in thousands of product pages and dilutes the help content. The agent reads a snapshot, so pages that change need reading again.

What the agent may do on its own

Meta draws the line at actions that are easy to undo, and says to draw it in what the agent is technically able to do, not only in its instructions.

Yes: start a return within policy
Reversible, and cheap to get wrong.
No: refunds, goodwill credits, cancelling after dispatch
Not easily undone. These go to a person, through a ticket or a handover.
The business's system decides what is returnable
Left to compare a delivery date with the policy, Meta says the agent will occasionally get it wrong, and tends to err in the customer's favour. The order system should say whether each item is still within the return window.

Looking up an order without asking for a number, and what an agent should and should not say about a late delivery, are covered in the order and basket guide.

Handover, nights, and the numbers to watch

Meta says escalation only works if there is somewhere to escalate to. The agent needs a handoff setting, and the conversation has to actually move: while the team holds it, the agent stops replying and staff answer in the same WhatsApp chat, with the ticket or order reference carried over. How that works when two systems share one number is in the conversation routing guide.

  1. Decide what happens out of hours. Handing over at 02:00 into an empty queue is worse than saying when the team is back and raising a ticket so the case is not lost.
  2. Watch two numbers together. The share of conversations handed over, and the time from handover to the first reply from a person. A high resolution rate is only good if the cases that need a person reach one quickly.
  3. Watch ticket creation hardest. Other failures are visible because the customer complains. A failed ticket is not: the agent has already promised a colleague will be in touch, and nobody finds out until the customer writes again, angrier.
  4. Test what should not happen. Meta singles out a refund demand, a direct request for a person, a question the policies do not answer, and an angry repeat contact. A support agent is judged on its worst answers, not its average one.

Questions

Will Meta's agent make up answers about our policies?

Meta's guide sets it up to answer only from the business's own written policies, and to hand over what they do not cover. Meta says the failures that matter are confident answers to questions the agent had no basis to answer, so test those before going live.

Can the agent give refunds?

Meta advises against it. A refund decision is not easily undone, so it goes to a person. The agent can start a return that is within policy.

What should the agent do when a customer asks for a person?

Say yes plainly, give the hours, and hand over. Meta says deflecting that question is the fastest way to turn a solvable case into a complaint.

How do we know the agent is helping?

Meta suggests tracking the share of conversations handed to a person and the time to the first human reply, alongside how many cases the agent resolves.

Sources

Want handed-over cases to reach someone fast?

We set up a business's WhatsApp number on a system its team runs, so when an AI agent hands a customer over, a person can pick the conversation up in one place the business owns. A short call is the easiest way to see how that would work for you.

Book a demo