Systems and records
Business workflow idle time labels: mark delays without guessing causes
By SourceX Editorial · Updated
10 sections
Short answer
Label waiting periods in two layers. First, the wait state, read directly from events: queued, assigned but not started, suspended, parked in a waiting status, between handoffs, outside working hours or unobserved. Second, the cause, recorded only when a named record supports it and graded by evidence strength. Without such a record the cause stays unknown, and labeled waits must reconcile to each case's elapsed time.
License terms
Business workflow idle time
Key takeaways
- A gap between events is a fact; the reason for it is a claim that needs its own record.
- Status names such as On-hold or Waiting for customer reflect configuration intent, not what happened in a given case.
- Grade every cause from E3 recorded to E0 unknown, and train only on grades a record supports.
- Fix clock offsets, identifier joins and missing start events before measuring any wait.
- Labeled intervals should sum to each case's elapsed time and match the system's own wait metrics where they exist.
How do you label workflow waiting periods without inventing their causes?
Label every waiting period in two layers. The first layer is the wait state, which you can read directly from the event log: queued, assigned but not started, suspended, parked in a waiting status, between teams, or outside working hours. The second layer is the cause, which you record only when a named record supports it, together with an evidence strength grade. When no record supports a cause, the cause is unknown, and it stays unknown.
This split keeps idle time labels honest. A gap of three days between two activities is a fact. Saying the supplier was slow, the approver was on holiday or the team was overloaded is a claim, and each of those claims needs a different kind of evidence. Training data that blurs the two teaches a model to explain delays fluently and wrongly.
The rest of this page gives the annotation guide: the wait states, the evidence grades, the causes with their required evidence, the clock and identity checks that come first, a worked example and the reconciliation that closes the loop.
Why a status name is not a cause
Business systems name statuses after the reason someone expected a ticket to sit there. The name is a hint about intent at configuration time, not a record of what happened in a given case.
Zendesk is a useful illustration. Its requester wait time counts time in New, Open and On-hold and pauses while a ticket is Pending, and its agent work time leaves out On-hold because, in Zendesk's own explanation, tickets on hold are being worked on by a third-party vendor. In a particular help desk, agents may use On-hold for a part on backorder, an internal engineering question or simply to stop a timer. The metric definition carries an assumed cause; the case may not.
Jira Service Management works the same way from the other direction. Each SLA can have an optional condition for pausing the clock, with waiting for a customer to respond as the documented use case, and administrators choose which statuses trigger it. Two projects in the same company can pause on different statuses for different reasons. An annotator who copies the status name into a cause field is labeling the configuration, not the case.
The wait states you can read from the log
Wait states come from the structure of events, so two annotators working from the same log should always agree on them. They rely on lifecycle transitions: the XES standard's lifecycle extension defines schedule, assign, start, suspend, resume and complete among its standard transitions, and many ticketing, ERP and workflow tools record some subset of them under other names.
| Wait state | Starts at | Ends at | Needs these events |
|---|---|---|---|
| Queued | Item becomes ready or scheduled | Item is assigned to a person or team | Ready or schedule time and assign time |
| Assigned, not started | Assignment | First work action by the assignee | Assign time and a start or first-action time |
| Suspended | Work is paused by the assignee | Work resumes | Suspend and resume events, or status changes mapped to them |
| Parked in a waiting status | Entry into a status the system treats as waiting | Exit from that status | Status history with timestamps and the status catalog |
| Between handoffs | Completion of one activity | Start of the next activity, often in another system | End and start times joined by a case identifier |
| Outside working time | Close of the working calendar | Reopening of the calendar | Calendar, holidays and the time zone of the resource |
| Unobserved gap | Last event that explains the state | Next event that explains the state | Only the bounding events; nothing between them |
Grading the evidence behind a cause
Every cause label carries a grade. The grade describes how directly the record supports the cause, not how plausible the cause sounds. Graders should be able to point to the record ids behind any grade above zero.
| Grade | Meaning | Typical support | How a model may use it |
|---|---|---|---|
| E3 Recorded | A record states the reason in the case itself | Reason code on the status change, a dated request for information with its reply | Train and evaluate on the cause |
| E2 Corroborated | Two independent records agree and nothing contradicts them | Assignee busy on another case in the log plus a team calendar showing coverage gaps | Train with the grade as a feature; evaluate separately |
| E1 Inferred | One pattern in one source suggests a cause | Many items released at the same hour, suggesting a batch | Use for analysis only, not as a target label |
| E0 Unknown | No record speaks to the cause | Bounding events only | Keep the wait, leave the cause unknown |
Which causes are allowed, and what each one requires
Use a short, closed list of causes. A useful starting point comes from research on waiting times in business processes: Lashkevich, Milani, Chapela-Campa, Suvorau and Dumas classify waiting time between activities into batch processing, resource contention, prioritization, resource unavailability and extraneous factors, and their open-source implementation produces that breakdown from a log with start and end timestamps. Add an external dependency cause for waits on a customer, supplier or outside approver, because business records often document those explicitly.
The list matters less than the evidence rule attached to each entry. The table gives the minimum record for each cause and a common false attribution.
| Cause | Minimum evidence for E2 or E3 | Common false attribution |
|---|---|---|
| Batching | Several items released together at a recorded batch run or scheduled job | Calling any morning spike a batch when it is just start of shift |
| Resource contention | Assignee recorded as working another case during the whole wait | Assuming a busy team from total volume, without per-person events |
| Prioritization | Assignee started a later-arriving item first, with a priority field or rule | Reading every out-of-order start as deliberate priority |
| Resource unavailability | Calendar, shift or absence record covering the wait | Treating a long gap that spans a weekend as wholly caused by the weekend |
| External dependency | Dated outbound request and the inbound reply that ended the wait | Copying a status name such as Waiting for customer as proof |
| Unknown | None needed; this is the default | Replacing unknown with the most common cause to make data look complete |
Clock offsets, merged IDs and missing transitions: checks before measuring any wait
Every wait is a subtraction between two timestamps, so a wait is only as good as the clocks and the joins behind it. Run three checks on each source before any labeling starts.
Clocks first. Record whether each source stores UTC, local time with an offset or local time with no offset. RFC 3339 timestamps carry a stated offset from UTC, so 16:39 at minus eight hours and 00:39 the next day in UTC are the same instant; a source without offsets needs its zone documented before subtraction. Daylight saving changes and servers in different zones are where invented waits of exactly one or several hours come from.
Identifiers second. A handoff wait exists only if the case identifier links both sides. Merged customer records, reused ticket numbers after a migration and order lines split into new ids all create waits that never happened or hide ones that did. Sample linked pairs and confirm them by hand.
Transitions third. If a system records only status changes and never a start, assigned-not-started and active work collapse into one interval. Mark those intervals with the coarser state rather than estimating a start time. The waiting time analysis tool cited above, for instance, expects both a start and an end timestamp for each activity, which many business exports do not provide.
A worked example: a warranty claim that waits three times (illustrative)
Starting state: a warranty claim in a service desk passes from intake to a parts specialist, waits for the customer to send photos, waits for a supplier's decision and returns for closing. The full case spans nine calendar days. All values are illustrative.
Wait one, about six working hours between intake and the specialist's first action. The log shows assignment at intake, and the specialist's own events show work on four other claims during that window. State: assigned, not started. Cause: resource contention at E2, supported by the specialist's concurrent case events and the assignment time.
Wait two, two calendar days in the status Waiting for customer. The case has an outbound message requesting photos and an inbound reply with attachments that ends the status. State: parked in a waiting status. Cause: external dependency, customer, at E3, supported by the dated request and reply.
Wait three, four calendar days in the status On hold. There is no outbound message to the supplier, no supplier reference and no note. A colleague recalls the supplier being slow that month. State: parked in a waiting status. Cause: unknown at E0. The recollection is not a record and does not raise the grade; if the team later finds the supplier portal entry, the label can move to external dependency, supplier, at E3 with that record linked.
The exception worth keeping: the remaining gap before closing spans a weekend. Only the hours inside the closed calendar are labeled resource unavailability, from the calendar table. The working hours on either side stay in their own state, because a weekend explains the weekend, not the Friday afternoon before it.
Reconciling labeled waits back to the source totals
Labels should add up. For each case, the sum of active intervals, wait states and unobserved gaps must equal the case's elapsed time from first to last event, within a documented tolerance for timestamp rounding. A case that does not reconcile has a clock or join problem, and it should be fixed or excluded before its causes are counted.
Across the dataset, compare totals with the system's own metrics where they exist. If the help desk reports requester wait time for a ticket, the reconstructed New, Open and On-hold intervals for that ticket should match it; a systematic difference usually means a status mapping is wrong or business hours are being applied on one side only.
Report the share of wait time at each evidence grade alongside any cause breakdown. A breakdown in which most time is E0 is still useful, because it tells the workflow owner where records are missing, but it should never be presented as a measured explanation of delay.
When to leave causes unlabeled, and when to ask the workflow owner
Stop at wait states, with every cause unknown, when the system never recorded starts, when statuses are used inconsistently across teams or when the only explanation for most waits is memory. That dataset is still worth having: correctly bounded waits with honest unknowns are better training material than confident, fabricated reasons.
Bring in the workflow owner when a status catalog has no written definitions, when calendar data is missing for a team with shifts, or when external parties are involved and their reference numbers live in a separate portal. The owner can often locate the record that turns an E0 into an E3; an annotator should not guess on their behalf.
Narrow scope rather than lowering the bar. Labeling one process family well, with a reconciled clock and a written status catalog, gives a cleaner signal than labeling every workflow in the company at E1.
Why waits with honest unknowns add value to licensed workflow data
Agent builders want workflow histories that show real time, including the time nothing happened. Waits with bounded states, graded causes and reconciled totals let them model when to follow up, when to escalate and when waiting is simply correct, without learning a fictional story about every delay.
Companies with years of such history can license chosen workflow records through SourceX, which manages rights review, contracts, preparation and delivery. A short assessment with no uploads shows first whether the company fits, and no record ships until the company has approved the scope and signed.
Frequently asked questions
What is the difference between idle time and waiting time in a workflow?
Idle time usually describes a resource with nothing to do, while waiting time describes a case that is not being worked on. A case can wait while its assignee is fully busy elsewhere, which is why labels should describe the case's state and leave the resource's situation to the cause layer.
Should waiting time be measured in business hours or calendar hours?
Keep both. Calendar time is what the customer experiences, and business hours show what the team could have acted on. Store the calendar used, and label hours outside it as unavailability only from the calendar record.
Can a model infer the cause of a delay from text in comments?
Sometimes, and when a comment states the reason it can support an E3 label. A model's guess from indirect wording is at most E1 and should be kept out of target labels until a person or a second record confirms it.
How many causes can one waiting period have?
Overlapping causes happen, for example a wait that spans a weekend and a supplier reply. Split the interval at the boundaries the records support and give each piece its own cause and grade, rather than attaching two causes to one span.
What if the system only records status changes and no start times?
Then active work and waiting inside a status cannot be separated. Label the coarser interval as the status, note that starts were not recorded and avoid tools or metrics that assume start events exist.
Sources
- The open-source waiting time analysis tool classifies waiting time between activities into batching, resource contention, prioritization, resource unavailability and extraneous factors, from a log with start and end timestamps. Source
- Lashkevich, Milani, Chapela-Campa, Suvorau and Dumas present a data-driven method that classifies waiting times in business processes by cause (CAiSE 2023). Source
- The XES lifecycle extension defines a standard transactional model with transitions including schedule, assign, start, suspend, resume and complete. Source
- Zendesk's requester wait time is the combined time a ticket spends in New, Open and On-hold, and on-hold time is reported separately. Source
- Zendesk SLA requester wait time pauses in Pending, and agent work time leaves out On-hold because those tickets are being worked on by a third-party vendor. Source
- Jira Service Management SLAs can include an optional condition to pause counting time, for example while waiting for a customer to respond. Source
- RFC 3339 timestamps state an offset from UTC, so 1996-12-19T16:39:57-08:00 is the same instant as 1996-12-20T00:39:57Z. Source
See if your company qualifies
A short company assessment. No data uploads are needed.