A customer changes the requested date during the call, and the receptionist keeps using an availability result from before the correction. The result was valid for the old request, not the current one.
Calendar checks need an explicit relationship to the request version they evaluated.
Identify which changes invalidate availability
Date, time, service, duration, location, staff member, room, and required equipment can all affect eligibility. The integration should know which fields its availability result depends on.
For a fictional service visit, changing a contact email may leave availability intact. Changing the service address can alter travel feasibility and require a new check.
| Change | Expected treatment |
|---|---|
| Requested date changes | Discard the old availability result |
| Service duration changes | Re-evaluate the occupied interval |
| Staff member changes | Check the newly eligible resource |
| Another customer takes the slot | Handle conflict at the scheduling authority |
Do not use a single “details updated” flag if different changes need different validation.
Attach a version to the request and result
Assign a new request version when a relevant field changes. Include the version in internal tool calls and responses.
When a delayed result returns, compare it with the current version before offering the slot. A response arriving last should not win merely because it is the newest network event.
Before submission, verify that the booking arguments match the confirmed latest request. Otherwise, the agent may speak Thursday while an old hidden payload still contains Tuesday.
The slow-tool article covers late responses during an ongoing conversation. This article focuses on the calendar dependencies those responses represent.
Build a correction-focused test matrix
Create fictional calls that change one relevant field at a time. Introduce a delayed lookup, let the customer correct the request, then release the old response.
Check the offered times, final tool arguments, saved record, and confirmation wording. A transcript containing the correction is not enough if the integration ignored it.
Also test an office employee changing the calendar during the same call. Request versioning handles customer-state changes; it does not replace the scheduling system's conflict controls.
Prefer visible rechecking over silent assumptions
The receptionist can briefly explain that it is checking the revised details. That adds a little conversation time while reducing the chance of offering stale information.
Avoid restarting the entire intake if only one field needs revalidation. Preserve verified information and redo the checks affected by the change.
Mid-call change questions
Should every correction trigger every lookup again?
Not necessarily. Define dependencies and invalidate the relevant checks.
Can the model decide which result is current?
Use explicit application state rather than relying only on conversational interpretation.
What should a release test show?
Show a late old result being rejected and the revised booking saved correctly. Include it in the acceptance suite.