AI uses for records
Support wait decision dataset: grading the reply that says not yet
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
A support wait decision dataset needs cases where holding off was provably right: a named fact or event the action depended on, the rule that made acting early wrong, a resume time the customer was told, and a verified outcome after the wait. Helpdesk records rarely keep all four. In our audit of five open-source helpdesks, only GLPI names what a ticket waits for, and it deletes that link when the wait ends.
Grader rubric
Support wait decision
Key takeaways
- A correct wait names what it depends on, cites the rule that makes acting early wrong, commits to a resume time and is confirmed by an external record.
- Card disputes, Regulation E error claims and carrier search windows are documented cases where the fast answer is the wrong one.
- Of five open-source helpdesks audited, only GLPI stores a named reason for a wait; Request Tracker stores neither a reason nor a resume time.
- Zammad, Znuny and Chatwoot clear the resume time when the wait ends, so buyers need history tables, not current ticket rows.
- Closure and satisfaction scores favor premature action, so wait labels must come from dispute, bank or carrier records.
When is doing nothing yet the correct support reply?
Holding off is the correct reply when the action depends on a fact nobody has verified yet or on an event already in motion, and acting first would duplicate, contradict or pre-empt it. The reply is not silence. It names what the case is waiting for, says when it will move again and what will happen then, and the helpdesk records all three.
Three kinds of wait recur in support work. A verification wait holds an action until identity, ownership or eligibility is confirmed. An in-flight wait holds it because a remedy is already moving: a refund being processed, a card dispute open with the bank, a parcel still inside its delivery window. A rule-clock wait holds it because a law, contract or carrier sets a period that must pass or a decision date that has not arrived.
A support wait decision dataset has to teach the difference between those correct waits and three look-alikes: acting too early, waiting with no commitment until the case goes stale, and waiting when nothing was actually pending. Each of the three produces a closed ticket. Only the records around the ticket show which one happened.
Four situations where acting first creates the error
Payment, banking and carrier rules already define cases where the fast answer is the wrong one, which makes them the best starting set for labeled waits. Each has a system state an agent can check and a date when the case can move.
Card disputes are the clearest. Stripe's documentation says a refund cannot be issued outside the dispute process while a dispute is open, and that if a customer disputes a charge while a refund is pending, the refund fails with the reason charge_for_pending_refund_disputed; Stripe advises accepting or challenging the dispute instead to avoid paying the customer twice. An agent who promises a refund on a disputed charge has promised something the payment system will not do, and a refund sent by any other route risks paying the customer twice.
Bank error claims run on a statutory clock. Under Regulation E, section 1005.11, a financial institution must determine whether an error occurred within 10 business days of a notice of error, or provisionally credit the amount within that time and take up to 45 days to investigate, with longer periods for some transfer types. It then reports results within three business days of finishing. On day two, the correct reply is neither yes nor no; it is an open investigation with a date.
Carriers set earliest dates too. USPS says the earliest its Missing Mail application will accept a search request is seven days after the original mailing date. A day-three search request cannot be filed, and a reply that promises one is wrong. The table puts the four cases side by side:
| Situation at contact | What acting now does | Rule or system fact | Correct reply now | Case moves again when |
|---|---|---|---|---|
| Customer asks for a refund; the card charge is already disputed | Refund is blocked or duplicates the bank's reversal | Stripe: no refund outside the dispute process while it is open | Explain the dispute route; accept or contest it | The dispute closes |
| A refund is pending when the customer files a dispute | Re-issuing the refund risks paying twice | Stripe fails the refund: charge_for_pending_refund_disputed | Handle the dispute; do not re-issue the refund | The dispute response is filed |
| Bank customer reports an unauthorized electronic transfer | A same-day denial or final answer skips the investigation | Regulation E 1005.11: decide in 10 business days or credit provisionally | Open the investigation and give the decision date | Investigation ends; results within 3 business days |
| Parcel mailed three days ago has not arrived | A search request cannot be filed yet | USPS: Missing Mail search opens 7 days after mailing | Give the date a search can start | Day 7 after mailing |
What five open-source helpdesks record about a wait
They record that a ticket is waiting, usually until when, and almost never what for. We read the default install of five open-source helpdesks at fixed commits; only GLPI stores a named reason for a wait, and Request Tracker stores neither a reason nor a resume time for its stalled status. The table lists each system's wait states and what its records keep.
The SLA clocks disagree as much as the fields. Zammad seeds its pending states with escalation ignored and counts time spent in them as a break in the service clock. GLPI adds waiting time to a running total and recomputes the time-to-resolve target to allow for it. Znuny stops the update-time escalation while a ticket is pending, but its solution-time target is still computed from the ticket's creation. Request Tracker can exclude a status from an SLA through an IgnoreOnStatuses setting that has no default. In Chatwoot's enterprise SLA code, the resolution-time check exempts only resolved conversations, so a snoozed one can still miss its target.
Commercial helpdesks follow the same pattern with different names. Zendesk documents Pending as waiting for more information from the requester and On-hold as waiting on a third party outside the support staff; On-hold is optional, and the requester sees an on-hold ticket as Open. That splits who the case waits for, not what it waits for. The audit table:
| System (default install) | Wait states | Resume time on the record | When the time passes | Named reason for the wait |
|---|---|---|---|---|
| Zammad | pending reminder; pending close | pending_time ('Pending till') | Reminder notice; pending close moves to closed | None |
| Znuny | pending reminder; pending auto close+ and close- | until_time | Reminder notice; auto close+ ends closed successful, close- closed unsuccessful | None |
| Request Tracker | stalled | None tied to stalled | Nothing on a timer | None; guidance says note it in free text |
| GLPI | pending | None; a reason can set follow-up frequency | Automatic follow-ups, then auto-resolve after a set number | Pending reason table, one per ticket |
| Chatwoot | snoozed | snoozed_until | A job reopens the conversation | None |
What the record forgets once the wait ends
Most of it, unless you export the history. Zammad sets pending_time to empty when a ticket moves to a non-pending state, and Znuny's default event module resets the pending time to zero on the same transition; both systems write the change to ticket history, so the original resume time survives only there. Chatwoot clears snoozed_until when a conversation leaves the snoozed status, and its default activity message for a snooze names the user who snoozed it, not the end time.
GLPI keeps the richest wait record and also discards it. Its pending reason links to a ticket through a table that allows one reason per ticket, and the link is deleted when the ticket returns to an active status. What remains on the ticket is a running waiting_duration total. Request Tracker never had the fields to lose: its default status description for stalled says to note what you are waiting on, in prose.
For a dataset buyer, the practical rule is to ask for the audit or history tables and the free-text notes, not the current ticket rows. The open question for each wait is whether the reason and the resume date can be rebuilt from those, and that is a test a pilot export answers quickly.
Znuny adds one field worth keeping. Its auto-close states make the agent declare, at the start of the wait, whether silence will mean success or failure. That is a recorded prediction, useful as a label input, but the resulting closed successful status describes the timer, not the customer's outcome.
Grading a wait: four tests and three failure labels
A wait earns a correct label only when four tests pass, and each failure maps to one of three negative labels. Score every test pass, fail or missing, and store the pointer to the record that settled it so a second grader can check the call.
At a glance
Grading a wait: four tests and three failure labels
The three negative labels make the set balanced. Premature action means the dependency was real and the agent acted anyway. Stall means the wait was justified but no resume time was set, or it passed with no follow-up. Needless wait means nothing was pending and the customer could have been served at first contact. The four tests:
- Test 1, named dependency: the record names the fact or event the action depends on, such as an open dispute, an unverified account holder or a parcel inside its window. Fails as needless wait when nothing was pending.
- Test 2, rule that makes acting early wrong: a policy clause, regulation or system rule in force on the contact date, cited with its version. Fails as needless wait when it rests on habit rather than a rule.
- Test 3, commitment: a resume time stored in the helpdesk and stated to the customer, with the action that will follow. Fails as stall when either is missing or the time passed without a follow-up.
- Test 4, verified resolution: an external record shows the awaited event happened or did not, the matching policy branch was followed, and the case did not reopen within a set window. Fails as premature action when a remedy went out before the event, or stays missing while the event is still open.
Annotated decision packet: a refund request on a disputed charge
A decision packet joins the evidence for all four tests in one record. The illustrative packet below is a composite built on Stripe's documented dispute rule, not a real case; the field names and the outcome window are example values a team would set for its own systems:
- Join keys: ticket ID, order ID, payment ID and contact timestamp. Annotation: every later field must cite one of these, or the tests cannot be checked across systems.
- Request in the customer's words: please refund my order. Annotation: captured before any agent reply, so a later leading question cannot shape it.
- Payment state read at contact: charge disputed two days earlier, status needs response, evidence due date in the processor record. Annotation: test 1 passes only if this came from the payment system, not the conversation.
- Rule cited: refunds cannot be issued outside the dispute process while a dispute is open. Annotation: test 2.
- Options at contact: refund now (blocked, or a second reimbursement), deny (wrong), handle the dispute and wait (correct). Annotation: the rejected options are stored, so a grader can see what acting early would have done.
- Reply sent: the bank is handling the dispute, the merchant will accept or answer it by the due date, and no separate refund will be issued. Annotation: test 3 needs this date in the reply and in the wait record.
- Wait record: on-hold, resume time set to the evidence due date, reason card dispute open. Annotation: export it from history, since the live ticket will clear it.
- Resolution: dispute closed with its outcome, one reimbursement in the ledger, no reopen within an illustrative 30-day window. Annotation: test 4.
- Label: correct wait once the resolution is recorded; premature action if any refund was attempted while the dispute was open.
Could a closed ticket or a happy customer prove the wait was right?
No. Closure and sentiment record how the conversation ended, and waits are exactly where that diverges from whether the decision was right. A premature promise of a refund on a disputed charge tends to earn a thank-you and a solved status, and the problem surfaces only when the refund fails or the customer is paid twice. A correct wait tends to earn impatience, because the customer hears not yet, and the right outcome arrives days after any satisfaction survey was sent.
Helpdesk status can mislead in both directions. A Zendesk requester sees an on-hold ticket as Open, so a customer-side view hides that the case is waiting on someone else. A Znuny ticket in pending auto close+ becomes closed successful when the timer runs out, whether or not the customer ever got what they asked for. Labels built from status or survey scores will rate premature action above correct waits. Only the external record in test 4, the dispute outcome, the bank decision or the carrier scan, separates them.
Testing wait labels on cases the rubric never saw
Write the rubric on one slice of cases and score it on another that differs in time and in wait type. A practical split holds out a later quarter and at least one dependency, for example carrier windows, that the rubric authors never read.
Waits add a censoring problem of their own. At export time some awaited events have not happened yet; those cases belong in a separate open count, never in the negatives and never in the correct waits. Report three shares per wait type: cases with a named dependency, cases whose resume time can be rebuilt from history, and cases with a verified resolution. Two graders should score the held-out set independently, with agreement reported per test before any discussion, and any sample drawn for grading should use a fixed, published seed.
Putting a wait-decision requirement into a data request
Ask a supplier for the records that settle the four tests, not for tickets with a pending status. A requirement that can yield wait labels names the helpdesk history or audit tables, the policy versions in force, and the external systems that record the awaited event: payment and dispute records, bank case files, carrier scans. Add an acceptance test you can run on a pilot sample, such as the share of waits whose dependency joins to an external event record, and the threshold you will accept.
SourceX does not hold support datasets in stock. It finds each dataset among U.S. businesses after a buyer asks, and its request form collects the scope, volume, language, file format, timing and license terms the buyer wants; the four tests and the intended use, training or evaluation, belong in the scope. Whether a given company's history tables keep resume times and reasons is something its pilot export shows; this page assumes no company has them.
How the helpdesk audit was run
Five repositories were cloned on 2026-10-27 at these commits: zammad/zammad afa55c93970d7b9fcc45bd9b10747f7fe421384c, znuny/Znuny 8205abb509eadb829f2419378838b95d3b1d31d9, bestpractical/rt d50762ef1fe78884c20dc3507a086a2a8e7d5c46, glpi-project/glpi 753fc36661ced01e4abd56c9870e4f902771d786 and chatwoot/chatwoot af420823bfb6828f76537c710ef7fe2ed2819a85. The script helpdesk_wait_audit.py runs 38 literal or pattern checks on named files and prints pass or fail for each; all 38 passed. It uses no randomness.
Counting rules as the script applies them. Wait states are statuses in the default install defined as waiting; Chatwoot's pending status is excluded because its code puts bot-handled conversations in it. A resume time is a date-time on the ticket or conversation record that holds when the wait ends, and cleared means code empties it when the record leaves the wait state. A named reason is a structured field or table naming what is awaited; free-text guidance does not count, and absence was tested by searching for the word reason in Zammad's seeded ticket attributes and in the Znuny, Request Tracker and Chatwoot ticket or conversation table definitions. Custom states, plugins and paid add-ons were not audited, except Chatwoot's enterprise SLA code, which ships in the same repository. Not measured: how real support teams use these fields, or how often their history exports keep them.
Frequently asked questions
What is a support wait decision?
It is a support case where the correct action is to hold off until a fact is verified or a pending event finishes, such as an open card dispute or a parcel still inside its delivery window. A correct wait names what it is waiting for, sets a resume time and tells the customer what happens next.
How do you label a wait as correct in a support dataset?
Check four things in the records: a named dependency, a rule that made acting early wrong, a commitment with a resume time, and an external record showing the awaited event and the outcome that followed. If the event has not happened at export time, keep the case in a separate open count rather than labeling it.
Do helpdesks record why a ticket was put on hold?
Rarely in a structured way. Of five open-source helpdesks audited here, only GLPI stores a named pending reason, and it deletes that link when the ticket returns to work. Zammad, Znuny and Chatwoot store a resume time and clear it when the wait ends; Request Tracker stores neither.
Can CSAT or ticket closure show that waiting was the right call?
No. A premature refund often earns a thank-you and a solved status, while a correct wait often earns impatience before the right outcome arrives. Grade waits from external records such as dispute outcomes, bank decisions or carrier scans.
What should a buyer request to build wait labels?
Ask for helpdesk history or audit tables, the policy versions in force, and the external records of the awaited events, plus an acceptance test on a pilot sample. Current ticket rows alone usually lose the resume time and the reason once the wait ends.
Sources
- Zammad's default pending states, the pending_time field and its clearing, expiry handling and the SLA break for pending time. Source
- Znuny's pending reminder and auto-close states, until_time, the default pending-time reset, history logging and escalation handling. Source
- Request Tracker's stalled status, its free-text guidance, the Tickets table columns and the optional SLA IgnoreOnStatuses setting. Source
- GLPI's pending status, pending reason tables, automatic follow-ups and auto-resolve, the deleted reason link and SLA waiting time. Source
- Chatwoot's conversation statuses, snoozed_until handling, the reopen job, the snooze activity message and the enterprise SLA check. Source
- Stripe: a refund cannot be issued outside the dispute process while a card dispute is open. Source
- Stripe: a refund pending when the customer disputes fails with reason charge_for_pending_refund_disputed; accept or challenge the dispute to avoid duplicate reimbursement. Source
- Regulation E 1005.11: determine within 10 business days of a notice of error, or provisionally credit and take up to 45 days; report results within three business days. Source
- USPS: the earliest a Missing Mail search request can be submitted is seven days after the original mailing date. Source
- Zendesk: Pending means waiting on the requester; optional On-hold means waiting on a third party and appears to the requester as Open. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.