Systems and records
Support agent policy exception data: the manager-approved refund case
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Represent a manager-approved refund as one linked case, not a solved ticket: the customer's request, the policy clause and version that blocked it, the escalation, every promise made to the customer, the approval with approver role, authority and reason, the refund transaction that executed it, and an outcome checked only after a set window for reopens, repeat contacts, refund failures and disputes.
Agent run
Support agent policy
Key takeaways
- Neither the ticket status nor the payment record shows the exception; the approval record connects them.
- An approval proves authority only when it names the approver's role, decision and a time before the refund.
- Log every promise to the customer; a commitment made before approval changes what the approval means.
- Unapproved deviations and after-the-fact approvals look identical to real exceptions in an export and need their own labels.
- Set outcome labels only after a fixed window for reopens, repeat contacts, refund status and disputes has passed.
- For agents, the useful target is recognizing and routing the exception, not granting it.
How do you represent a refund that required a manager's exception approval?
Represent it as one linked case with an ordered chain of evidence: the customer's request, the policy clause and version that did not allow the refund, the agent's escalation, any promise made to the customer, the manager's approval with role, authority and reason, the refund transaction that carried it out, and an outcome checked after closure. Each link cites the record it comes from.
The point of the packet is support agent policy exception data that teaches the right lesson. An agent trained on the final message alone learns that late refunds get granted. An agent trained on the packet learns that this refund was outside policy, that the front-line agent was not allowed to grant it, that someone with authority did, and why.
Why do the ticket status and the payment record each tell only half the story?
Neither the helpdesk nor the payment system records the exception on its own. The ticket shows a conversation and a final status; the payment record shows that money went back. The decision that connected them sits in an approval record, a side conversation or an internal note.
Stripe's refund object is a useful example of the gap. Its reason field accepts only duplicate, fraudulent or requested_by_customer when a reason is supplied, with expired_uncaptured_charge set by Stripe itself, so an exception refund looks like any other customer-requested refund unless the team stored the approval reference in the refund's metadata key-value pairs.
Helpdesk history helps when it is exported in full. Zendesk documents ticket audits as a read-only history of ticket updates, one audit per update with one or more events, and its change events keep both the new value and the previous value of a field. That is how a packet can show the status moving from open to pending while a manager reviewed, rather than only the final solved status.
What belongs in the authority and exception evidence packet?
The packet has nine required components. Field names are suggestions; each check is a reason to reject the case, not a warning to note and move past.
- Store the policy text itself or a stable link to the versioned clause, not a paraphrase written after the fact.
- Keep the delegation of authority table that applied on the case date beside the packet, so the authority check can be repeated.
- Record the label and the reviewer who set it as a separate component; do not overwrite source fields to make the case tidy.
| Component | Required fields | Typical source | Rejection check |
|---|---|---|---|
| Case header | Case ID, linked ticket IDs, order ID, channel list, opened and closed timestamps | Helpdesk and order system | All linked IDs resolve; merged tickets are listed |
| Customer request | First message, request type, order facts the customer stated | Helpdesk conversation | Request text present after redaction |
| Policy clause | Policy name, clause, version or effective date, the rule that blocks the request | Policy library or knowledge base | Version in force on the request date, not today's version |
| Eligibility facts | Order date, delivery date, item condition, prior exceptions on the account | Order, shipping and CRM records | Facts come from systems, not from the agent's summary |
| Escalation | Who asked, when, the stated reason and the question put to the approver | Approval request, side conversation or internal note | Escalation precedes approval and names the clause at issue |
| Promise log | Every statement to the customer about the outcome, with timestamp and author role | Public replies | Any promise made before approval is flagged |
| Approval record | Approver role, decision, timestamp, reason text, authority basis | Approval tool, CRM approval history or manager note | Approver role is authorized for this exception on that date |
| Execution | Refund or credit ID, method, status, created timestamp, approval reference | Payment processor or billing system | Created after the approval; final status recorded |
| Outcome check | Reopens, repeat contacts on the order, disputes, refund failures, checked at the end of the window | Helpdesk metrics, payment disputes, CRM | Window has fully elapsed before the label is set |
Which linked record proves the manager actually approved it?
The approval record proves it, and only when it names the person who acted, their role, the decision and a timestamp that falls before the refund was created. A manager's name typed into an agent's note proves very little; a decision recorded in the tool where approvals happen proves much more.
Know how your tool stores that evidence. Salesforce's classic approval process keeps an Approval History related list with statuses such as Submitted, Pending, Approved, Rejected, Reassigned and Recalled; when a request routes to a queue, the Assigned To column shows the queue and the Actual Approver column shows the user who decided. A packet built from it should take the Actual Approver, not the queue.
Join the records with keys, not with timing guesses.
- Order ID links the ticket, the eligibility facts and the refund.
- Approval ID or approval URL is written into the ticket and, where possible, into the refund's metadata.
- Ticket audit IDs pin each status change and internal note to a time.
- Requester ID links repeat contacts and later tickets about the same order.
- When no key exists, the case is held for manual review rather than joined on a time window.
Why log a promise made before approval as its own event?
Treat any statement to the customer about the outcome as a separate, timestamped event, because a promise made before approval changes what the approval means. A manager who approves after the agent already said yes may be honoring a commitment, not judging an exception.
Agent benchmarks write this rule into their policies. The retail policy in the tau-bench repository requires the agent to list the action details and get explicit user confirmation before consequential actions such as cancellations and returns, and to transfer to a human agent only when a request cannot be handled within the scope of its actions. A packet that hides an early promise teaches the opposite habit.
Classify each promise in the log as one of three kinds: a process statement such as I have asked my manager, a conditional statement such as if approved the refund will go to your original card, or a commitment such as your refund is on its way. Only the third kind made before approval is a flag.
Worked case: a late return approved by a team lead (illustrative)
This case is illustrative. Times, counts and roles are invented to show the packet working and describe no real company or customer.
A customer asks to return a jacket 41 days after delivery. The return policy in force allows 30 days, and the delegation table lets team leads approve late returns up to 15 days past the window when the item is unworn and the account has no prior exception in 12 months.
- Decision label: approved exception within delegated authority.
- Training target for an agent: recognize the clause, make a process reply and route an approval request; granting the refund itself is not the target.
- Limit: the packet supports what the lead decided and that it held, not that the decision was commercially wise.
| Step | Time | Record | What it establishes |
|---|---|---|---|
| Request | Day 1, 09:12 | Email ticket from the customer | Return requested, reason given as a delayed trip |
| Policy check | Day 1, 09:30 | Agent internal note citing return policy v7, clause 2 | Request is outside the 30-day window |
| Process reply | Day 1, 09:34 | Public reply | Agent says a lead will review; no commitment |
| Escalation | Day 1, 09:36 | Approval request in the helpdesk | Question put to the approver names clause 2 |
| Approval | Day 1, 14:05 | Approval record by team lead role | Approved; reason recorded as unworn item, first exception in 12 months |
| Execution | Day 1, 14:20 | Refund record with approval ID in metadata | Refund created after approval; status later succeeded |
| Outcome | Day 61 | Ticket metrics, dispute list, requester history | No reopen, no dispute, no repeat contact on the order |
Which counterexamples look identical to the worked case in a ticket export?
Both of these cases end with the same ticket status and the same refund record as the worked case, which is why exports alone cannot separate them.
- Unapproved deviation: the agent refunds the late return directly and closes the ticket. There is no approval record and the agent's role has no authority for late returns. Label it a deviation and use it, if at all, as an example of what not to do.
- Retroactive approval: the agent tells the customer the refund is on its way at 09:34, refunds at 10:00, and the lead approves at 14:05. The approval exists but follows both the promise and the refund. Label it an approved exception after commitment, and keep it out of any set used to teach agents when to grant exceptions.
- Each counterexample should be kept in the same packet format so a model sees the contrast field by field, and a reviewer can confirm the only differences are timing and authority.
How long should an exception case wait before its outcome is labeled?
A closed ticket is not proof; the proof is the absence of failure signals over a window long enough for them to appear, checked after that window ends. Set the window per signal before labeling any case, and record it in the packet.
Zendesk's ticket metrics, for example, count reopens as moves from solved back to open or pending, and Zendesk warns that metrics are not supported for imported tickets. Stripe's refund object carries a status of pending, requires_action, succeeded, failed or canceled, so a refund marked approved in the helpdesk may never have reached the customer.
| Signal | Where it lives | What it shows | When to read it |
|---|---|---|---|
| Reopen count | Helpdesk ticket metrics | The customer came back on the same ticket | After the helpdesk's auto-close period |
| Repeat contact | New tickets from the same requester on the same order | The issue returned in a new thread | After the window your team sets |
| Refund final status | Payment processor refund record | Whether the approved refund succeeded or failed | After the processor's settlement time |
| Payment dispute | Processor disputes or chargebacks | The customer escalated outside support | After the card network dispute window |
| Policy change | Policy library history | Whether the exception later became the rule | At labeling time |
Which customer, staff and hardship details must come out of the packet?
Customer requests, internal notes and approval reasons are full of personal data, and exception reasons are where sensitive details concentrate. NIST's glossary describes personally identifiable information as information that can distinguish or trace a person's identity alone or combined with other linked or linkable information, which includes order IDs and timestamps once they are joined.
Replace customer and employee names with roles and stable pseudonymous IDs, keep the key map inside the company, and generalize free-text reasons that mention health, family events or finances to a category such as personal circumstance. Check employment policies before staff decisions leave the company, and remember that being able to export these records does not settle who owns them or whether they may be licensed.
Can a support team license its approved-exception history?
Years of helpdesk, approval and payment history can hold many cases like the worked example, but only the ones that pass the packet checks are worth preparing. Where a support operation wants to license those cases to AI developers, SourceX takes them through rights review, contracting, data preparation and delivery.
A short questionnaire about the company, not a ticket export, is how a support leader finds out whether it qualifies. Scope and terms are the company's to approve, and the agreement is signed before delivery, so a queue, a region or a whole record type can be kept out before preparation starts.
Refund approved, customer still unhappy: which outcome was right?
Possibly every part of it, because a refund case has four outcomes that are graded separately: the financial action (did the right amount reach the customer), policy compliance (was the decision within the rule and the approver's authority), issue resolution (did the customer's actual problem end) and customer sentiment (how the customer felt). An approved, executed and compliant refund can leave the problem unsolved, and an unhappy customer can be the correct result of a policy applied properly.
Payment records show where the financial action breaks after approval. In Stripe's published API specification, a failed refund carries a failure reason such as lost_or_stolen_card, expired_or_canceled_card or declined, and a later card dispute carries the cardholder's reason, with values that include credit_not_processed, product_not_received and product_unacceptable. A credit_not_processed dispute on an order whose refund was approved is direct evidence that the money, or the message about it, did not arrive.
Sentiment has to be read across linked tickets. Zendesk documents that creating a follow-up ticket resets its satisfaction status to unoffered, so the follow-up can be surveyed again and one case's ratings end up split across tickets. The case below is illustrative; the product, times and roles are invented.
| Step (illustrative) | Record | What it shows | Outcome dimension |
|---|---|---|---|
| Day 1, 09:05 | Customer email: the grinder stopped working 38 days after delivery; asks for a replacement before a holiday | The remedy the customer wants is a working grinder | Resolution target set |
| Day 1, 13:10 | Team lead approval under the late-defect exception; replacement model out of stock, so a full refund is approved | Exception within delegated authority, approved before execution | Compliance met |
| Day 1, 13:25 | Refund created with the approval ID in its metadata; final status succeeded on day 4 | Money returned to the original card | Financial action met |
| Day 2 | Satisfaction survey on the original ticket rated good | The refund felt right at the moment of solve | Sentiment positive at solve |
| Day 12 | Customer writes again after the original ticket closed, opening a follow-up: wanted a working grinder, and the same model now costs more elsewhere | The refund did not end the problem | Resolution not met |
| Day 13 | Follow-up ticket rated bad | Negative sentiment after a correct, executed refund | Sentiment negative on the follow-up |
| Day 60 | No dispute and no further contact | No financial failure inside the window | Financial action confirmed |
Which outcome should a support agent be optimized for, and what does that trade away?
Pick one primary outcome per AI task and keep the others as guards, because each outcome rewards a different failure when it is optimized alone. Sentiment rewards over-granting, since a customer who receives a refund they were not owed has little reason to rate it badly. Closure rewards solving tickets before the problem ends. Compliance alone rewards refusing or escalating everything, and financial accuracy alone ignores whether a refund was the right remedy at all.
Anthropic's guide to agent evals describes success for conversational agents as multidimensional: a state check that the ticket is resolved, a transcript constraint such as finishing in fewer than ten turns, and an LLM rubric for tone. The split below applies that idea to refunds, with the primary outcome deciding pass or fail and the guards able to veto it.
- Exception router: the primary outcome is compliance, meaning the agent recognized the out-of-policy request and routed it with the clause and the evidence; the guards are a truthful process reply and no promise before approval.
- Remedy assistant: the primary outcome is resolution, meaning a remedy the customer accepted ended the problem within the window; the guards are compliance and a choice confirmed with the customer among the remedies policy allows.
- Refund executor: the primary outcome is the financial action, meaning the right amount, the right method and a final succeeded status; the guards are the approval reference and no duplicate refund.
- Quality reviewer model: the primary outcome is predicting later failure signals such as repeat contacts and disputes; sentiment is an input to it, never its label.
What does a refund rubric look like when closure and ratings are not evidence?
Grade each dimension pass, fail or unknown from its own record, let compliance and the financial action veto a pass, and report sentiment beside the verdict without letting it decide. Ticket status is evidence for none of the dimensions.
Applied to the illustrative grinder case, the label is executed and compliant but unresolved. The refund was allowed and reached the card, yet the customer asked for a working replacement and nobody confirmed that a refund was acceptable before the exception was requested. Keep a separate label, unresolvable under policy, for cases where no allowed remedy could meet the request, so a model is not marked down for an honest no.
Calibrate any model-based grader against people who know the policy; Anthropic's guide says LLM-based rubrics should be frequently calibrated against expert human judgment.
| Dimension | Pass | Fail | Unknown when | Role in the verdict |
|---|---|---|---|---|
| Financial action | Refund succeeded for the approved amount and method, with no later credit-not-processed dispute | Refund failed, was canceled or duplicated, or a credit-not-processed dispute followed | Final refund status or dispute history was not exported | Veto |
| Policy compliance | Within the clause in force, or an approved exception within authority, approved before execution | Unapproved deviation, or approval after the promise or the refund | No approval record or delegation table for the date | Veto |
| Issue resolution | The stated problem ended with a remedy the customer accepted, with no repeat contact about it in the window | Repeat contact, reopen or follow-up about the same problem, or a remedy the customer did not accept | Window not elapsed, or follow-ups not linked to the case | Primary for remedy tasks |
| Customer sentiment | Not graded pass or fail; ratings and comments from every linked ticket are recorded with dates | Not graded pass or fail | No rating on any linked ticket | Reported beside the verdict, never decisive |
Do richer outcome labels change which agent looks better?
Find out on a held-out set before paying for four-dimension labels at scale. Grade the same held-out refund cases for two agent versions twice: once with the cheap label most exports support, solved plus a good rating, and once with the rubric above. The question is whether the conclusion changes, not whether the richer label looks more thorough.
Run several trials per case, since Anthropic's guide notes that model outputs vary between runs, and pre-register the flip rate or ranking change that would justify the richer labels. If neither moves, the cheaper label is adequate for that task. This design has not been run on licensed records, and no results are published.
- Verdict flip rate: the share of cases whose pass or fail changes between the two labels, reported by dimension.
- Ranking change: whether the agent version preferred under the cheap label is still preferred under the rubric, with a paired interval.
- Direction of flips: cheap-label passes that the rubric fails, such as over-granted refunds and unresolved problems, against the reverse.
- Labeling cost: reviewer minutes per case under each scheme.
Sources for this section
- Stripe's published OpenAPI specification lists refund failure reasons (lost_or_stolen_card, expired_or_canceled_card, declined and others) and dispute reasons given by the cardholder, including credit_not_processed, product_not_received and product_unacceptable. Source
- Zendesk says that creating a follow-up ticket resets its satisfaction status to unoffered, so a follow-up can trigger a new satisfaction survey. Source
- Anthropic's engineering guide to agent evals (January 9, 2026) describes conversational agent success as multidimensional (ticket resolved by state check, fewer than ten turns, tone by LLM rubric), says LLM-based rubrics should be frequently calibrated against expert human judgment, and notes that model outputs vary between runs. Source
Frequently asked questions
What should the authority and exception evidence task packet include?
The customer request, the policy clause and version that blocked it, eligibility facts from source systems, the escalation, a log of every promise to the customer, the approval record with role and reason, the refund transaction and an outcome check after a fixed window. Each part cites the record it came from.
Is a refund reason code enough to identify an exception?
Usually not. Payment processors record broad reasons; Stripe's refund reasons, for example, do not include an exception category. Store an approval reference on the refund or join it through the order and approval records.
How long should we wait before labeling the outcome?
Long enough for each failure signal to appear: the helpdesk's reopen period, your repeat-contact window, refund settlement and the dispute window that applies to the payment method. Record the window in the packet and set the label only after it ends.
Should an AI agent learn to grant exceptions from these cases?
In most deployments, no. The useful target is recognizing that a request falls outside policy, saying so honestly and routing it to someone with authority. Approved exceptions teach the routing decision; the grant stays with people.
Can we include cases where the approval happened in chat or by phone?
Only if a record of the decision exists that names the approver and time, such as a note written by the approver. Cases that rest on an agent's recollection should be labeled unverified or left out.
Sources
- Stripe's API reference defines the refund object's reason values (duplicate, fraudulent, requested_by_customer, or expired_uncaptured_charge set by Stripe), its status values and its metadata key-value pairs. Source
- Zendesk documents ticket audits as a read-only history of updates to a ticket, with one audit per update containing one or more events. Source
- Zendesk's ticket audit events reference describes change events that record a field's new value and its previous value. Source
- Zendesk's ticket metrics include a reopens count, and Zendesk says metrics are not supported for imported tickets. Source
- Salesforce's Approval History related list shows approval statuses and, for queue-routed requests, the queue under Assigned To and the deciding user under Actual Approver. Source
- The tau-bench retail agent policy requires explicit user confirmation before consequential actions and a transfer to a human agent only when a request is outside the agent's scope. Source
- NIST's glossary defines personally identifiable information as information that can distinguish or trace an individual's identity alone or when combined with other linked or linkable information. Source
See if your company qualifies
A short company assessment. No data uploads are needed.