A useful human handoff preserves the customer request, verified facts, missing details and reason for escalation, then assigns a real owner. The AI should stop when information, authority or a connected system is missing.
Key takeaways
- Define stop conditions before writing prompts.
- Distinguish live transfer, callback task and approval queue.
- Never claim a handoff succeeded until a case was created.
- Test routing and staff receipt, not just chatbot wording.
The problem and the first decision
A customer asks a website assistant for opening hours. It answers correctly. Another asks whether a service is available in their area. The assistant needs to check a policy. A third wants a refund that falls outside the normal rules. If the AI keeps chatting through every situation, it may sound helpful while quietly creating mistakes. A good receptionist agent is designed around two equally important abilities: responding within its role and recognising when to stop.
What is a human handoff in an AI receptionist?
A human handoff is the transfer of a customer request from an automated assistant to a responsible person or team, together with the information needed to continue. It should include the customer's actual request, verified answers, missing details, current status and reason for escalation. Saying “Someone will contact you” is not a complete handoff unless the business receives the case and has an owner who can act on it.
Five reasons an AI receptionist should stop
The answer is not supported — The business's approved knowledge does not contain the requested policy, price or availability. The agent should admit uncertainty rather than invent an answer.
The request needs human authority — Refund exceptions, contract commitments, disputes, negotiation and regulated or clinical decisions should follow the business's approved review process.
The customer asks for a person — An explicit human request should be respected. The agent can ask for enough contact context to connect the request, without trapping the customer in repeated prompts.
A tool or data source fails — If a calendar, CRM or order system is unavailable, the agent should mark the action unconfirmed, log the failure and arrange a follow-up.
The interaction becomes urgent or sensitive — Safety, abuse, suspected fraud, sensitive personal information or unusual distress may require specialised escalation. The agent should follow approved business procedures, not improvise.
An example of a useful handoff
A customer tells an appliance-service company: “The technician did not arrive, and I want to cancel.” A weak assistant replies with another generic FAQ and ends the chat. A better assistant identifies an existing-service complaint, checks only the approved records it may access, asks for the minimum details needed, and routes the case to the service manager. The handoff brief should say: Intent: Missed visit and cancellation request. Customer statement: “Technician did not arrive.”
Reference: Job number provided, verification status noted. Actions taken: No booking changed; no refund promised. Missing details: Whether staff attended the address. Owner: Service manager. Next action: Investigate and contact customer. This is illustrative, but it shows why a summary is more useful than a transcript dumped into an inbox.
Choose the right handoff type
Live transfer: A staffed agent is actually available and the channel supports a conversation switch. Callback or task: Staff are unavailable; the system creates a clearly assigned follow-up and communicates the expected process, without an invented response-time promise. Approval queue: The AI can collect the request but must not perform an action until an authorised person approves it. Specialist routing: The issue belongs with a particular team such as accounts, customer support or a clinician.
An AI must not say “I've transferred you” when no one is available or no case was successfully created.
Design boundaries before drafting prompts
Write a small decision table that the business owner can review. Question answered by current approved information → Agent may answer. Missing but non-sensitive detail → Ask one focused clarification. Human decision required → Collect context and escalate. External system not responding → Mark pending and alert. User requests person → Offer approved handoff channel. The wording of the bot matters, but permissions, system checks and queue ownership matter more.
Test the handoff experience end to end
Run conversations where the user changes intent, asks for a refund exception, repeats “human please,” supplies conflicting information, or encounters an unavailable calendar. Verify not only the response text but also the actual routing: did the correct task appear, with correct status and context? Did duplicate requests create repeated tickets? Did staff receive a notification? Have a person review cases labelled completed, pending and escalated separately.
Measure whether escalation is improving service
Track correct escalation rate, unresolved cases without an owner, time to human action, repeat explanations required, inaccurate confirmations and customer satisfaction. A high handoff rate is not automatically a failure; it may be appropriate for a narrow first pilot. The danger is a silent handoff that nobody receives.
How NORVOI can help
NORVOI's AI-agent work includes defining tasks, approved knowledge, human handoff points, tests and review. We can help map what the agent may answer and the exact workflow that follows when it cannot.
Frequently asked questions
When should a chatbot transfer to a human?
When information is uncertain, authority is missing, a system fails, the user requests a person, or the request falls outside approved rules.
What information should accompany a handoff?
The original request, verified facts, missing details, actions taken, what remains unconfirmed, and the assigned owner.
Can the AI promise a callback time?
Only if the business has a genuine, approved service commitment and a process capable of meeting it.
How do we know if handoffs work?
Test whether the right person actually receives and resolves the request—not merely whether the bot says it escalated.
A reliable stop is part of a good answer
The best AI receptionist isn't the one that never says “I can't confirm that.” It is the one that helps within clear boundaries and makes the next human step easier. Bring one example of a conversation where an AI receptionist should stop and a person should take over. We can help define the escalation trigger, the information staff must receive and how to confirm the handoff actually happened.
Have a similar task in your business? Start with the request, the information your team trusts and the point where a person should take over.