Systems and records
Support promise tracking data: did the team deliver what was promised?
By SourceX Editorial · Updated
9 sections
Short answer
To verify a support promise, link the commitment in the conversation to the record in the system that had to act, such as a billing refund, a Jira fix or a logged callback, and check that it happened by the stated time. Then confirm the problem stayed solved: no reopen, no follow-up ticket and no repeat contact within a set window. A solved ticket status alone proves neither.
System export
Support promise tracking
Key takeaways
- A promise is only verifiable when it names an action, an owner and a time, and points to the system where that action is recorded.
- Ticket status measures the support queue, not delivery; the proof sits in billing, engineering, logistics or telephony records.
- Fulfillment needs a maturity window: Zendesk reopens and follow-up tickets created from closed tickets are the evidence that a fix held.
- Handoffs, corrections and promises made beyond an agent's authority need their own fields rather than being folded into one outcome.
- A schema with enumerated outcomes and required links can be validated mechanically before anyone trusts the labels.
- Store pointers to comments and records instead of copying free text, which keeps privacy review manageable.
What record shows a support promise was actually kept?
You verify it by finding the record of the promised action in the system that performs it, checking the timing against what was said, and then watching for evidence that the customer's problem came back. Three links, each to a different record: the commitment, the action and the after-effect.
Support teams usually stop at the first link. The ticket shows the agent wrote that a refund would post in five business days, then the ticket was solved. Whether the refund posted lives in the billing system; whether the customer came back lives in later tickets, calls or chats. Without those links the dataset records intentions, not outcomes.
This matters most when the promise was made by an AI agent or a new hire, because the team needs to know whether commitments are being made that operations cannot meet. It also makes support history far more useful as training data, since a model learns what was actually delivered rather than what was said.
What counts as a promise in a support conversation?
A promise is a statement that commits the company to a specific future action for this customer, with an owner and, explicitly or by policy, a time. Courtesy phrases do not count; a commitment that cannot be checked against a record is noted as unverifiable rather than guessed. The table maps common commitment types to the system that proves delivery; example wording is illustrative.
| Commitment type | Illustrative wording | System that acts | Proof of delivery |
|---|---|---|---|
| Callback | A specialist will call you tomorrow morning | Telephony or contact center platform | Outbound call log to the customer's number within the window |
| Refund or credit | The credit will appear on your next statement | Billing or payments system | Credit memo or refund transaction tied to the account |
| Escalation to engineering | I have raised this with our engineers | Jira or another issue tracker | Linked issue created, with the support case referenced |
| Fix in a release | This will be fixed in the next update | Issue tracker and release records | Linked issue resolved and shipped in a dated release |
| Replacement shipment | A replacement ships today | Order management or warehouse system | Shipment record with carrier scan by the stated date |
| Policy exception | We will waive the restocking charge this time | Order or billing system plus approval record | Adjustment posted and an approval by someone with authority |
| Status update | I will email you by Friday with an update | Help desk | Public reply on the case by the stated date |
The promise, action and follow-up evidence schema
The schema below records one row per promise, not per ticket, because a single case can hold several commitments with different owners and outcomes. Fields that point to evidence store references, such as comment or transaction IDs, rather than copied text.
At a glance
The promise, action and follow-up evidence schema
Each link is a small provenance statement in the sense of the W3C PROV ontology: an outcome label is derived from an action record, which was generated by an activity and attributed to an agent. Recording links that way makes it obvious which label rests on which evidence when a reviewer challenges it.
| Field | Type | Purpose | Validation check |
|---|---|---|---|
| case_id and promise_id | string | Identify the case and the individual commitment | Both required; promise_id unique |
| promise_ref and made_at | comment ID and timestamp | Point to the reply containing the commitment | Reference resolves to a public reply on the case |
| promise_type and made_by_role | enum and enum | Callback, refund, escalation, fix, shipment, exception or update; made by a human agent, AI agent, supervisor or other team | Both values in their enumerated lists; AI agent flagged for separate analysis |
| policy_version and authority_ref | string and record ID or null | Support policy in force when the promise was made, and the approval needed for exceptions beyond the agent's limit | Version effective on made_at; authority_ref required when promise_type is exception |
| due_at | timestamp | Deadline stated or implied by policy | Later than made_at; source of deadline recorded |
| owner_chain and corrections | ordered list and list of comment IDs | Every handoff of the commitment, and later replies that changed or withdrew it | Each handoff has a role and timestamp; each correction is after made_at |
| acting_system, action_ref and action_at | string, record ID and timestamp | Where delivery is recorded and when it actually happened | Required for any kept outcome; action_at compared with due_at |
| window_end | timestamp | End of the period watched for the problem returning | Outcome cannot be final before window_end |
| recurrence_refs | list of case IDs | Reopens, follow-up tickets and repeat contacts on the same issue | Each dated within the window |
| outcome | enum | Kept on time, kept late, partly kept, broken, superseded, not mature | Consistent with action_at, due_at and window_end |
Which validator checks catch a bad promise record?
The checks are deliberately mechanical so anyone can rerun them. JSON Schema's validation vocabulary handles the first two directly: required lists the fields that must be present and enum restricts a value to the listed options. The rest are cross-field rules that a short script applies after schema validation.
- Check 1: every required field is present, and promise_type, made_by_role and outcome hold allowed values.
- Check 2: due_at is later than made_at, and the source of the deadline is recorded.
- Check 3: any kept outcome has an action_ref and an action_at.
- Check 4: kept on time requires action_at on or before due_at; kept late requires it after.
- Check 5: an outcome other than not mature requires the current date to be past window_end.
- Check 6: superseded requires at least one correction that withdraws or replaces the promise.
- Check 7: an exception promise without an authority_ref is flagged as made beyond authority, whatever the outcome.
- Check 8: a kept outcome with a recurrence inside the window is downgraded for review rather than accepted.
Where does the evidence sit in Zendesk, Salesforce and Jira?
Each platform keeps a different slice, and knowing which one keeps which slice saves weeks of guessing. In Zendesk, ticket audits are a read-only history of every update to a ticket, with events such as field changes, new comments and notifications; for change capture at scale Zendesk's own docs point to the incremental ticket event export instead. The ticket metrics object includes a reopens count and solved time. Zendesk also states that metrics are not supported for imported tickets, so cases migrated from an older help desk need their recurrence evidence rebuilt from audits or source data.
In Salesforce Service Cloud, entitlement milestones are the closest built-in promise record. Milestone statuses distinguish compliant work from open violations and closed violations, where a milestone was missed but the step was still completed, which maps directly to kept late. A Salesforce knowledge article notes that milestones should be completed before the case closes, because the completion date cannot be changed once the milestone has exited the entitlement process, so closed cases with open milestones are a known data-quality gap.
In Jira, the link between a support case and engineering work is a work item link. Atlassian lists link types such as is caused by, causes, relates to and implements; record which type was used, because relates to says far less than implements about whether the fix addressed the customer's problem.
What proves the problem stayed solved after the ticket closed?
Proof is the absence of recurrence over a window long enough for the problem to come back, plus positive evidence where the action itself can be observed later, such as a credit that stayed posted. Pick the window by promise type and write it down before labelling; a shipment might need a couple of weeks, while a fix promised for the next release needs the release date plus a period of use. These lengths are judgment calls for each team, not standards.
In Zendesk, a solved ticket can still be reopened, and the reopens metric counts how many times that happened. Once a ticket is closed it cannot be reopened, so a new ticket is created as a follow-up that references the closed one through via_followup_source_id, and the parent lists its follow-ups. Count both. Then search other channels on the same account for the same issue, because a customer who gave up on email may call instead.
Until the window ends, the outcome is not mature. Counting those cases as kept is the most common way promise data overstates delivery.
A refund that posted and a fix that never shipped
Both records are illustrative and invented to show the structure; they are not customer data. The first passes every check. The second shows the pattern most support exports contain.
At a glance
A refund that posted and a fix that never shipped
- Valid: promise_type refund; made_by_role human agent; made_at day 0; due_at day 5 by policy version 7; action_ref a credit memo in billing; action_at day 3; owner_chain agent then billing team; recurrence_refs empty; window_end day 35; outcome kept on time. Checks 1 to 8 pass.
- Failing: promise_type fix; made_by_role AI agent; made_at day 0; due_at next release; action_ref missing; ticket solved on day 1; a follow-up ticket from the same customer on day 20 references the closed case; outcome recorded as kept on time.
- Why it fails: check 3 (kept with no action record), check 8 (recurrence inside the window) and check 5 if the release has not shipped. Corrected label: not mature until the linked Jira issue ships, then broken or kept late depending on the release date.
- What the failure teaches: the ticket was closed because the AI agent's reply ended the conversation, not because anything was delivered.
How should handoffs, corrections and authority change the label?
Handoffs do not change who the customer was promised something by, but they explain where delivery broke. Keep the full owner_chain; a broken promise whose last owner was a team with no access to the acting system is a process finding, not an agent error.
Corrections need care. If an agent later writes that the earlier date was wrong and gives a new one, the first promise is superseded and the new one is tracked on its own. If the agent silently misses the date and nobody corrects it, the promise is broken even if a later action eventually happens.
Authority is separate from outcome. A waived charge granted without the approval the policy version required can be kept on time and still flagged as beyond authority. Keeping that as its own field lets a team study over-promising without punishing the agents who delivered.
Licensing promise and follow-through records
Promise records with linked actions and mature outcomes are among the more distinctive kinds of support history, because they show what a company actually did rather than what it said. Privacy review still comes first: case references, account identifiers and reply text need a reviewer's sign-off before anything leaves the help desk.
A support organization that wants to offer this history to AI developers can do so through SourceX, which takes care of sourcing, rights review, contracting, preparation and delivery. Buyers tend to look for two or more years of consistent records and tens of thousands of cases in one system, though each deal is judged on its own. The process opens with a brief questionnaire about the company, before any ticket export is shared.
Frequently asked questions
What fields does a support promise record need?
At minimum: the case and promise IDs, a reference to the reply that made the promise, the promise type, who made it, the policy version, the deadline, the acting system and action record, the time the action happened, any corrections, the end of the watch window, recurrence links and an enumerated outcome.
Is a solved ticket proof that a promise was delivered?
No. A solved status shows the support queue moved on, often because the conversation ended. Delivery is proven in the system that acts, such as billing, the issue tracker or the warehouse, and durability is shown by the absence of reopens or follow-up tickets during a set window.
How long should you wait before labelling a promise as kept?
Set the window by promise type before labelling, long enough for the problem to recur if the fix did not work. Until the window ends, mark the outcome as not mature and leave it out of any analysis of kept or broken promises.
How do Zendesk follow-up tickets show a problem came back?
A closed Zendesk ticket cannot be reopened, so a new ticket is created as a follow-up that references the closed one, and the closed ticket lists its follow-ups. Counting follow-ups alongside reopens gives a fuller picture of recurrence than reopens alone.
Should promises made by AI support agents be tracked separately?
Track them in the same schema but flag the role, so the team can compare kept rates and authority flags between AI and human agents. AI agents can end conversations with commitments that no downstream team has accepted, which is exactly what the action link reveals.
Sources
- Zendesk closed tickets cannot be reopened; follow-up tickets reference the closed ticket through via_followup_source_id. Source
- Zendesk ticket metrics include a reopens count and the time a ticket was solved. Source
- Zendesk ticket audits are a read-only history of ticket updates; the docs recommend the incremental ticket event export for change capture. Source
- Zendesk states that metrics are not supported for imported tickets. Source
- Salesforce milestone statuses distinguish compliant, open violation and closed violation, where a violated milestone was still completed. Source
- Salesforce notes milestones should be completed before a case closes, since the completion date cannot be changed after the milestone exits the entitlement process. Source
- Jira work item link types include is caused by, causes, relates to and implements. Source
- The W3C PROV ontology relates entities, activities and agents through properties such as wasGeneratedBy, wasDerivedFrom and wasAttributedTo. Source
- JSON Schema validation defines required for mandatory properties and enum for a fixed list of allowed values. Source
See if your company qualifies
A short company assessment. No data uploads are needed.