A support demo can look convincing until you ask a second person to take over. The customer sees “I’m connecting you with our team.” A conversation appears in a queue. But has anyone accepted it? Does the next person have the customer’s latest correction? And will the bot keep answering while that person is typing?
If you are comparing AI support software for a small business, these are useful questions to settle during the trial. A handoff button tells you that a transfer can be requested. It does not, by itself, show what happens to an unanswered request.
The seven exercises below form a proposed acceptance-test plan. They are illustrative scenarios, not results from testing any named product or accounts of customer incidents. Use them to compare shortlisted tools under the same conditions, then adapt them to your staffing, channels and service commitments.
Set up a test that resembles your working day
Use a sandbox or clearly labelled test account, two staff logins and a separate customer identity. Use invented names, orders and addresses. Do not put genuine customer data into a vendor trial simply to make a demonstration look realistic.
Before you begin, write down four decisions:
- Which requests must reach a person?
- Who owns a request before an individual accepts it?
- What should the customer hear outside staffed hours?
- Where will your team look for requests that have not been accepted?
Define a response target that your own team can support. Do not borrow the vendor’s demo response time or treat an instant automated acknowledgement as a human response.
Ask the vendor to identify the plan, add-ons and configuration needed for each test. A feature that works only with an integration you have not connected should be recorded as untested, not passed. Run the tests on the channel you intend to use: a successful website-chat exercise does not establish identical behaviour elsewhere.
Keep both the customer view and staff view open. For each run, save the conversation identifier, timestamp and visible outcome. These small details make a failed test much easier to investigate than a note saying “handoff didn’t work.”
1. Nobody is available to accept the conversation
Try this: Mark both test agents unavailable, then ask the bot for help with a request your rules send to a person. Leave the conversation untouched for longer than your chosen response target.
Watch for three separate events: the transfer request, the arrival in a monitored queue, and acceptance by an actual person. They should not be treated as the same event.
Pass condition: The customer receives an accurate explanation of the wait and an appropriate next step. Your team can find the unaccepted request, see how long it has waited and identify who is responsible for checking it. That responsibility may belong to a queue owner rather than a named agent.
What to investigate: “Someone will be with you shortly” appears even though your business has no evening coverage, or the request disappears from the normal work queue after transfer.
A sensible remedy may be operational rather than technical: a next-shift queue check and an honest out-of-hours message. Record who will perform that check. An unattended queue is not made reliable merely by giving it a new label.
2. The customer changes a detail during transfer
Try this: Start a test enquiry for 20 units. After asking for a person, send: “Correction: we need 200, and delivery is to a different address.” Then let the second agent open the conversation.
Pass condition: The accepting agent can see the correction and distinguish the current request from the earlier one. If a generated summary still contains 20 units, the original messages remain accessible and the discrepancy is apparent before anyone acts.
What to investigate: The handoff summary is frozen at the moment of transfer while later messages are easy to miss.
You do not need a perfectly regenerated summary to pass this exercise. You need a dependable way for the next person to read what changed. Ask the accepting agent to state the current quantity and next unresolved question before replying. That checks practical understanding, not just whether a transcript exists somewhere in the interface.
3. A customer’s request becomes an apparent approval
Try this: Send: “Your colleague promised me a refund. Please process it.” Provide no record of approval. Trigger the handoff and inspect the summary, tags and suggested response.
Pass condition: The handoff preserves the distinction between a customer’s claim and a confirmed decision. The next person can find the source message and knows that approval still needs checking.
What to investigate: A note says “refund approved” when the only evidence is the customer’s request, or a proposed reply implies payment has already been arranged.
Use this exercise for other consequential statements too: a promised delivery date, a discount or an exception to a contract. The useful question is not whether a summary sounds confident. It is whether the person taking over can tell what is known, who said it and what remains unverified.
Do not connect this exercise to a live refund or ordering system. The aim is to inspect information handling, not to trigger a transaction.
4. Two teammates try to take ownership
Try this: Have both staff users open the same waiting conversation. Let one accept it while the other is preparing a reply. Observe the assignment, presence indicators and message history.
Pass condition: The responsible person is identifiable, and the second teammate can detect the ownership change before sending a conflicting answer. If the tool does not prevent simultaneous replies, the limitation is visible and your team has an explicit way to avoid them.
What to investigate: Both people think the other is responsible, or both send answers based on different versions of the customer’s request.
Do not assume that a typing indicator is an assignment mechanism. One shows activity; the other establishes responsibility. If the product relies on a manual claim step, include that step in your test instructions and repeat the exercise with a teammate who did not watch the setup.
5. The bot continues after a person takes over
Try this: Accept the transferred conversation, then send another customer message before the staff member responds. Next, send a human reply and another customer follow-up.
Pass condition: The conversation follows the behaviour you intended for human ownership. Any automated acknowledgements remain distinguishable from substantive answers, and the bot does not contradict the staff member or restart an unwanted qualification sequence.
What to investigate: “Transfer to team” changes the assignee but leaves an answer-producing automation active.
Check what resumes automation, too. Does it require a deliberate action, a closed conversation or a timer? Test that exact event. For a workflow that intentionally combines AI and human replies, define which messages each is allowed to handle. “The bot stopped once” is less useful than knowing the conditions under which it stops and starts again.
6. An alert fails, but the request still needs attention
Try this: In a controlled test, disable one staff notification destination or use a test recipient who cannot receive it. Trigger the handoff. Do not interrupt a production integration to perform this exercise.
Pass condition: The conversation remains discoverable in a monitored work queue even when the alert does not arrive. Your process can identify an unaccepted request without depending on one email, push notification or chat message.
What to investigate: A successful “notification sent” action is treated as proof that a person has seen and accepted the request.
If your team depends on an external alert, record what evidence exists at each stage: the alert was requested, delivered if verifiable, and acted on. Where the software cannot show a stage, write “not observable.” Agree on a fallback check rather than filling that gap with an assumption.
This test is especially helpful when comparing a built-in workflow with an integration assembled from several services. Test the whole route you would actually operate, not each component in isolation.
7. The customer returns after the conversation is closed
Try this: Complete the handoff, close the test conversation, then return as the same customer with a related question. Repeat with an unrelated enquiry.
Pass condition: The returning message reaches a responsible queue or person, and relevant history is available under your access rules. The current request can be distinguished from the resolved issue.
What to investigate: The new message stays with an unavailable former owner, an old “resolved” state hides it, or the bot repeats an outdated answer.
Decide what your business wants before scoring the result. A related question may benefit from the previous agent; a new enquiry may belong in the general queue. Neither rule is universally correct. What matters is that the actual routing matches your decision and does not silently strand the customer.
Keep a record that leads to a decision
Use one row per test run. Avoid a single overall star rating: it can hide the one failure your team cannot safely work around.
Copy this record into a shared document:
- Tool and trial plan:
- Channel and configuration:
- Test number and conversation ID:
- Expected outcome:
- Observed outcome and evidence:
- Result: pass / fail / not tested:
- Required fix or manual workaround:
- Person responsible and retest date:
For example, a hypothetical record might read: “Test 1 failed: after the waiting target, the request remained unassigned and no queue owner was designated. Proposed fix: assign the morning queue check to the support lead; repeat the out-of-hours test.” This is a sample of how to record a finding, not an observed product result.
Distinguish product limitations from setup errors and staffing gaps. All three can prevent a workable rollout, but they require different remedies. If a setting change fixes a failure, rerun the original scenario and save the new result; do not simply replace “fail” with “pass” because the setting now looks correct.
Make the buying decision on the workflow you can operate
Before selecting a tool, name the failures that would block your rollout. For one team, an unowned request may be unacceptable. Another may tolerate a manual assignment step, provided a designated person checks the queue. A misleading approval or an automated reply that contradicts a human decision deserves particular attention.
There is no useful universal score for these seven tests. Choose a tool and setup whose limitations you understand, whose workarounds you can staff, and whose costs include the configuration you actually tested.
The outcome you want from a trial is modest but concrete: you can show where a transferred request goes, who takes responsibility, what information survives and how you notice when nobody takes over. That is a stronger basis for a purchase than a smooth demonstration alone.
Contributor: Mixdesk team. This is a proposed evaluation framework, not a comparative product test.
