
An AI-to-human handoff works when a person can act on the case without asking the customer to start over. Write down the trigger, the queue that owns it, the information the next person needs and the fallback when nobody is available. Confirm the case in the receiving inbox before promising a transfer.
Use the editable handoff template to run through one real workflow with synthetic records. The sample below is a double-charge report that requires a billing teammate.
Decide when to transfer
Use a concrete event or policy condition for each rule. Start with explicit requests for a person and actions the agent lacks authority to perform. Define the point where repeated attempts stop being useful. Choose the thresholds using the cases you see in support, then test the edges. Intercom escalation guidance
| Trigger | Receiving owner | What the AI can do before handoff |
|---|---|---|
| Customer asks for a person | The current case owner or the general queue | Confirm the request and preserve the conversation. |
| Customer reports two charges for one purchase | Billing | Capture the symptom and reference after the applicable verification. Leave charge decisions to Billing. |
| A write times out and its result is unknown | The queue that owns the affected system | Record the attempt. Verify the record if a safe lookup is available. Otherwise transfer as uncertain. |
| An account lookup cannot verify the customer | The account-support queue, using your approved verification route | Explain how to verify securely. Keep account details out of the message. |
| Two approved troubleshooting attempts fail | The team that can reproduce the problem | Record both attempts and the result each produced. |
In BerryDesk, Escalate to Human pauses AI replies and moves the conversation to the shared inbox. The routing settings can select an available teammate, use a verified owner or try the eligible pool. When nobody is eligible, the case stays Unassigned for the team to pick up. Check these outcomes in the channel your customers use.
Give the next person a case they can work
A handoff note should fit above the fold in the receiving inbox. Start with the customer’s goal and the next action, then add the facts that change the decision. Include a link to the existing conversation and an internal record reference the receiver can open. Record verification status without copying credentials, full payment details or sensitive attachments into the note.
Here is a filled example with synthetic data.
| Field | Example |
|---|---|
| Customer goal | Check whether invoice INV-DEMO-418 was charged twice. |
| Trigger and destination | Possible duplicate charge. Billing queue. |
| Verified facts | Customer verified through the signed-in account. The invoice shows one captured payment. The customer sees a second card entry. |
| Unknowns | The second entry’s status and whether it maps to the same payment. |
| Actions already taken | Looked up the invoice. No refund or other write was attempted. |
| Next action | Billing opens the authorized payment record, checks the second entry and replies in the original thread. |
A useful customer acknowledgement for that case is “I found one completed payment for this invoice. I’ve sent your report to Billing to check the second entry. They’ll reply here.” Send that only after the transfer is recorded and the receiving team can reply in that channel. If the sender cannot confirm delivery, say who the customer can contact and give a realistic next update based on your support hours.
Rehearse the case with the receiving team
- Send the request from the real entry channel. Use a logged-in test shopper and a synthetic invoice. Ask for a person in one run and describe only the duplicate-charge symptom in another.
- Open the receiving inbox. Ask a billing teammate to find the case without a link from the tester. Check the queue, owner, action history and existing conversation.
- Answer as a human. Verify the reply reaches the same customer and channel. Send one follow-up from the customer and check which owner receives it. Confirm the AI stays paused while a human owns the case.
- Repeat with nobody eligible. Change eligibility in staging and repeat. Check that an assigned team can see the unassigned case and owns the next check. Confirm the acknowledgement reflects what actually happened.
- Repeat with a delivery failure. Force a controlled integration error. Check that the agent avoids a false transfer confirmation and that a teammate can reconcile that attempt with the receiving queue.
If there is a business-hours queue, repeat once just before closing and once after hours. Have the receiving person read the note and open the referenced record. Mark the handoff incomplete if they must ask the customer for information the system already verified.
Measure the transfer and the first useful reply
Track the handoff attempt, the case appearing in the correct queue, the first human reply and resolution as distinct events. Count each conversation once per handoff episode.
- Handoff delivery rate. Cases found in the correct queue divided by handoff attempts. An acknowledgement alone supplies no delivery evidence.
- Usable handoff rate. Reviewed delivered cases with the required context divided by delivered cases reviewed.
- Time to first human reply. Measure from arrival in the queue to the first useful human reply. Report business-hours and after-hours cases separately if their service targets differ.
Review any case where a customer was promised a transfer and nobody could find the case. If a write might have succeeded before a handoff, resolve the uncertain record before attempting the write again. Run the transfer alongside the 12 staging cases for AI support agents, then use the implementation playbook to plan the rollout.
Rehearse a handoff with your support team
- Route a staging conversation into your shared inbox
- Check ownership, the customer acknowledgement and the next human reply
Set up in minutes



