BLOG
A WhatsApp chatbot should pass a conversation to a person when it cannot safely or usefully finish the task. The handoff needs more than a “talk to an agent” button: the staff member must receive the customer’s request, relevant steps already taken, and a clear owner. Otherwise automation makes the customer repeat everything. Design the exit before automating the greeting, and test it when the team is busy or offline.
Decide when the bot must call a person
Set explicit triggers: the customer asks for a person, the question is outside the approved knowledge base, the bot fails twice, the request concerns a complaint or cancellation, or the action needs authorization. Sensitive medical, legal, financial, or safety matters require additional professional rules. Do not wait for a generative model to invent a threshold during the conversation.
Some requests can be routed by a simple menu or keyword, while others need interpretation. In either case, leave a visible human route. Do not tell a customer they have been transferred until a staff queue or assigned person actually receives the case.
Send a useful summary, not an unfiltered dump
The agent needs the customer’s name if provided, the reason for contact, the product or service discussed, relevant order or appointment ID, answers already given, and any pending decision. The customer should not have to recount those facts. At the same time, avoid copying an entire conversation or private details into every ticket when only a short, accurate summary is needed.
Make the summary readable and editable. An AI-produced summary can omit a negation, mistake a date, or add an assumption. Show the source conversation to the authorized agent when necessary and let them correct the record. Define which system stores the case and who may see it.
| Handoff element | Minimum useful detail | Common failure |
|---|---|---|
| Reason | Customer’s actual request | Generic “needs help” label |
| Context | Relevant answers and IDs | Customer repeats information |
| Owner | Named team or queue | No one accepts the case |
| Status | Waiting, accepted, resolved | False transfer confirmation |
| Fallback | Offline expectation and contact path | Conversation stalls overnight |
Make the transfer a four-step process
First, detect the trigger and stop the bot from continuing to answer as though it owns the case. Second, create a short summary and attach the necessary customer context. Third, assign the case to a monitored queue or person, recording the time and channel. Fourth, tell the customer what happens next using a realistic response expectation.
Track acceptance separately from assignment. A ticket placed in a queue is not the same as a human having read it. If the assigned person is unavailable, return the case to a monitored queue and notify the team. Keep a status the customer can understand without exposing internal workflow details.
Define the offline fallback
What happens after hours, during a system outage, or when the queue is full? State when the team will respond and capture only the information needed to continue later. Do not imply 24-hour human coverage if it does not exist. For urgent or regulated topics, use the organization’s approved route rather than improvising advice in a bot.
Have staff test the fallback from an external number. Check whether the case is visible the next morning and whether the customer receives duplicate automatic messages. A silent queue is not a handoff.
Test three real conversation patterns
Use fictional data to simulate a standard request that the bot can finish, a request outside its scope, and a complaint that asks for a person. In the second and third cases, verify the summary, assignment, agent notification, and customer expectation. Add a case where the customer changes their request midway. The bot should not keep pushing an outdated route.
Ask the agent whether they could respond without asking the customer to restate the issue. Time the interval from transfer to first human reply. Review wrong summaries and abandoned cases, not just successful automation counts. Fix the routing before expanding to more topics.
Choose the right technical layer
WhatsApp Business App includes tools such as away messages and quick replies, but those alone do not provide every CRM assignment or automated bot-to-agent workflow. A shared inbox, official platform integration, or CRM may be needed depending on team size and channels. Confirm which feature exists in the chosen product and plan.
Do not choose an integration based only on its chat demo. Check whether it preserves conversation IDs, passes context, enforces role-based access, handles duplicates, and offers export or audit records. Review Meta’s Business App information and the official platform documentation for the specific path you will use.
Measure the human outcome
Track how many cases were passed to staff, how many were accepted, how long the customer waited, how often they repeated information, and whether the original request was resolved. A high “bot containment” rate can hide customers who gave up. Sample whole conversations to see whether the route felt clear.
Give staff a way to report a bad transfer and edit the knowledge base. Review exception rules regularly. When the bot changes, test the human exit again; a small prompt or menu change can make it harder to reach the team.
Set a service-level target the team can actually meet and review it by hour of day. If requests arrive when nobody is assigned, adjust staffing, hours, or the automatic expectation. Better wording cannot replace an absent owner. Close resolved cases explicitly so they do not keep returning to the bot or trigger unnecessary reminders.
Frequently asked questions
Does transfer require AI? No. Explicit menu options and rules can handle many handoffs.
Does the free WhatsApp Business App provide a full bot handoff? Its native messages and replies are useful, but verify any shared queue or CRM transfer in the actual product and plan.
Should the agent get the whole history? Give them the authorized, relevant context and a way to inspect the full conversation when necessary.
How do I know it works? Verify assignment, agent acceptance, first response, customer repetition, and final resolution in test conversations.
