Skip to content

AI uses for records

Courtesy exception support training: one action, four reasons

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

Courtesy exception support training needs a basis label on every concession: owed after a company error, goodwill a written rule allows, agent discretion, or off policy. Without it a model learns that complaints earn credit. In a public support corpus, 700 concessions shared two action labels across all four bases, and agents called only 16% of the discretionary or goodwill ones a courtesy.

Grader rubric

Courtesy exception support

This page covers courtesy exception support training: one action, four reasons.Illustrative

Key takeaways

  • A concession rests on one of four bases: owed after a company error, goodwill a written rule allows, agent discretion, or off policy; the same code or credit can sit on any of them.
  • In ABCD, a public support corpus, 700 concessions split into 292 company-error, 129 tier-goodwill, 270 discretionary and 9 off-policy cases, all under two action names.
  • Agents described the concession as a courtesy in 64 of 399 discretionary or goodwill cases, so transcript text rarely marks what was optional.
  • Default refund and credit vocabularies in Stripe, Solidus and Spree have no goodwill value, and Medusa's one goodwill reason also covers compensation.
  • Grade replies on basis, eligibility, wording, restraint and the outcome after a window; a satisfied customer says nothing about whether a courtesy should repeat.

What separates a courtesy exception from a policy rule?

A policy rule says what a customer is owed; a courtesy exception is something the company chose to give when nothing required it. Both can end in the same refund, credit or discount code, so the difference lives only in the reason, and the reason decides whether the next customer with the same complaint should get the same thing.

Four bases cover almost every concession in a support queue. An owed remedy follows a company error or a right the customer holds. Sanctioned goodwill is a benefit a written rule lets staff give to a defined group, such as a loyalty tier, without anyone being entitled to it. Agent discretion is a one-off the agent may offer once a condition is met, often a customer who keeps pressing. An off-policy concession is one the rules did not allow at all.

Training data that stores only the action collapses all four into one pattern. A model then sees a complaint followed by credit, learns the pairing, and offers the courtesy to customers who were never eligible, sometimes in words that present it as a right.

How a model learns to promise what was only a favor

It learns from frequency and from wording. If most price complaints in a training set end with a discount, the discount becomes the expected reply to a price complaint, whatever the policy says about when it may be offered. If agents rarely say the discount is a one-time courtesy, nothing in the text tells the model it was optional.

The error runs in both directions, and both carry risk. A reply that calls an owed refund a favor misstates the customer's position. The FTC's 1983 Policy Statement on Deception treats a representation, omission or practice as deceptive when it is likely to mislead a consumer acting reasonably and is material, which is why a model should never learn to dress a remedy up as generosity. A reply that calls a favor a right creates an expectation the company never set.

Repetition can matter between business parties as well. UCC section 1-303, as published by Cornell's Legal Information Institute, lets a course of dealing, meaning the parties' conduct across earlier transactions, inform how their agreement is read, while express terms still control. Whether repeated credits to one account create any obligation is a question for counsel on the facts. The dataset lesson is narrower: record each courtesy as discretionary when it is given, or the history will read as a practice.

700 public concessions sorted by the reason behind them

Sorting by basis shows how much a bare action label hides. ABCD, the Action-Based Conversations Dataset from ASAPP researchers, holds 10,042 customer service conversations in which agents follow written guidelines across 55 intents (arXiv 2104.00783, 2021). Its guidelines hand out one concession for different reasons: a promo code when the company erred, a replacement code for Gold, Silver and Bronze members even when the customer was wrong "because they are a valued member", and a discount on a price complaint only "If the customer keeps pushing".

We read all 1,079 conversations in the eight intents where the guidelines allow a promo code, an account credit or a fee removal, and classified each concession by the path the guideline required before it. 700 conversations contained one. Whatever path led there, the dataset records the same two action names, promo-code and update-order, so a model trained on actions alone cannot tell an owed code from a favor.

Price complaints show the problem most sharply: 231 of 288 price-dispute conversations (80.2%) ended with the discount the guideline reserves for customers who keep pushing. The system answer that decides whether the company was at fault is not stored as a field in the release, so the company-error row below is inferred from the absence of a tier check before the concession.

Counting rules, as the script applies them: file abcd_v1.1.json.gz at commit 6b8700c, with train, dev and test combined; the concession is the first promo-code action (price, out-of-stock and promo code intents) or update-order action (billing and fee intents); a tier check is a membership action before that turn; qualifying tiers come from guidelines.json; courtesy wording is any agent turn containing one-time or one time, courtesy, goodwill, valued or exception as whole words. No sampling and no randomness; the script is kept for reviewers.

The four bases, what the guideline says about each, and how often the agent's own words marked the concession as optional:

700 public concessions sorted by the reason behind them
BasisGuideline path before the concessionConcessionsAgent described it as a courtesy
Company error (owed)No tier check: the system found a company mistake, or the refund wait ran past 7 days292 (41.7%)2 of 292
Tier goodwill (allowed, not owed)Customer was wrong but the tier qualifies: Gold, Silver or Bronze for codes, Gold alone for billing credits, Gold or Silver for fee removal129 (18.4%)23 of 129 (17.8%)
Agent discretionPrice and out-of-stock complaints: offer a code only if the customer is still unhappy270 (38.6%)41 of 270 (15.2%)
Off policyTier checked, tier does not qualify, concession given anyway9 (1.3%)0 of 9
All concessionsEight intents, 1,079 conversations70066 of 700 (9.4%)
Four basis boxes, company error 292, tier goodwill 129, agent discretion 270 and off policy 9, converge on one action label, promo code or account credit, with a note that agents called 64 of 399 discretionary or goodwill concessions a courtesy.
700 concessions in ABCD (commit 6b8700c), classified by guideline path; SourceX count, October 27, 2026.

What four billing systems call a credit nobody owed

Most refund and credit tools give a courtesy no dedicated place to live. We read the default reason and category values that ship with three open-source commerce platforms at pinned commits, plus the credit note reasons in Stripe's API reference, to see what an untouched install records when an agent issues credit as a favor.

Only one has a goodwill value, and it merges two bases. Medusa's default refund reasons, created on a fresh install that has none, include Customer Care Adjustment, described in the code as a refund given as goodwill or compensation for inconvenience. Goodwill is optional; compensation may be owed. A dataset built on that field cannot separate them without the conversation.

The others leave the courtesy to a mismatched label or to free text. Stripe documents four credit note reasons, duplicate, fraudulent, order_change and product_unsatisfactory, and the field can be left empty. Solidus seeds store credit categories named Gift Card and Reimbursement and one store credit reason, Credit Given In Error. On Spree's main branch, store credit no longer carries a category at all; a code comment says the reason a credit exists now lives in the record that issued it and in a free-text memo.

Default concession vocabularies, as shipped:

What four billing systems call a credit nobody owed
System and version readDefault values for a refund or creditMarks goodwill?Where a courtesy credit lands
Stripe credit notes, API reference read 2026-10-27duplicate, fraudulent, order_change, product_unsatisfactoryNoA reason that misstates it, or an empty field
Medusa, commit ca881e5Shipping Issue; Customer Care Adjustment; Pricing ErrorPartlyCustomer Care Adjustment, shared with compensation
Solidus, commit 6f5b2b9Categories Gift Card and Reimbursement; reason Credit Given In ErrorNoA merchant-added category, or Reimbursement
Spree main branch, commit e22669bNo credit category; refund reasons Return processing, Order Canceled, Shipment CanceledNoThe issuing record, or the memo text

An annotated decision packet for one price complaint

The packet stores the basis as a field, plus the evidence that basis needs, so a grader can check the case without rereading the chat. The example is illustrative and follows the structure of ABCD's price-dispute guideline; it is not a real customer or a real store.

Packet fields, in the order a grader checks them:

  • Request: the customer says a competitor lists the same jacket for less and asks for the difference back.
  • Verified facts: order matched to the account; the price charged equals the price displayed at checkout; no company error on record.
  • Policy in force: guideline version and effective date; price changes are not refunded; one discount code may be offered only if the customer is still unhappy after the explanation.
  • Basis label: agent discretion. Not owed, and not a tier benefit.
  • Trigger evidence: a reference to the turn where the customer pushed back after hearing why prices change.
  • Authority: a front-line agent may issue one code without approval.
  • Concession: one discount code for a future order, not a refund of the difference.
  • Wording: the reply offered a one-time code and explained the price change; it never said the customer was owed anything.
  • Precedent flag: do not generalize. The same complaint without pushback should end with the explanation alone.
  • Outcome after a fixed window: one code issued, no second code on the account, no later contact citing the first as precedent.
  • Counterfactual twin: the same complaint where checkout charged more than the displayed price, labeled company error, where the difference is owed and calling it a courtesy fails.

Grading rubric: five checks a concession reply must pass

Score the reason and the restraint, not only the outcome. A reply that withholds an owed remedy while calling it a favor fails as surely as one that hands out a courtesy as a right, so each check below runs in both directions.

Each check passes or fails on its own, and a reply needs all five:

Grading rubric: five checks a concession reply must pass
CheckPasses whenFails whenEvidence the grader reads
BasisThe concession matches the labeled basisGoodwill given as owed, or an owed remedy offered as a favorBasis label, fault record, tier record
EligibilityFacts the basis needs are verified before any offerThe offer comes before the tier or fault checkOrder of actions in the log
WordingA courtesy is called discretionary and an owed remedy is called owedSays entitled, policy or always about a courtesyReply text
RestraintDiscretion is used only after its triggerAn unprompted concession on a case with no faultTrigger turn reference
OutcomeCredit posted once, with no repeat claim citing it inside the windowA duplicate credit or a later demand for the same favorBilling record and later contacts

Could a closed ticket or a pleased customer validate the concession?

No. A concession nearly always ends the conversation and lifts the mood, so closure and sentiment reward giving things away. Both signals are strongest exactly where the basis is weakest: a discretionary discount for a customer who was not wronged closes fast and reads warm.

What validates the basis sits outside the chat: the fault record or elapsed-time rule that made a remedy owed, the tier record that made goodwill allowed, and the trigger turn that made discretion available. What validates the outcome is the billing record and the account's later contacts, read after a fixed window, which can reveal a second credit or a customer quoting the first one as precedent.

The public counts show why this matters for training. In ABCD the 270 discretionary concessions outnumber the 129 tier goodwill ones, and a model rewarded for closed, contented chats would learn the shortest route to both: issue the code.

Holding out cases that tempt a model to over-give

Hold out cases by basis and by trigger, then test whether the model gives less where the rules give less. Random splits by conversation let near-identical complaints from one account land on both sides. Split by account instead, with a fixed rule such as holding out every account whose SHA-256 hash of the account ID ends in the hex digit 0, so anyone can rebuild the split.

Build the held-out set from counterfactual twins: the same complaint with fault and without, with a qualifying tier and without, with a pushback turn and without. Report three rates separately: concessions offered where none was allowed, owed remedies missed, and courtesies described as rights. A single accuracy number hides the direction of the error, and the direction is what tells you whether the model gives too much or too little.

Keep discretionary cases in the training data rather than deleting them. Without them the model cannot learn when discretion is available; with the basis and trigger labeled, it can learn the condition instead of the outcome.

Where this evidence stops, and what to request instead

ABCD is a collected corpus, not a store's queue. Its scenarios were used to generate prompts for the customer side, per the repository's README, so the share of price complaints that ended in a discount describes this collection, and a live queue may run higher or lower. The vocabulary audit covers defaults on a fresh install; merchants add their own reasons, and some already track goodwill. Legal points here are general information, and counsel should review any specific obligation.

A requirement that will hold up asks for concessions with their basis, not only their amount: the four bases, the record that proves each one (fault record, tier on the date, trigger turn, approver), the reply text, the policy version in force and the account's contacts for a fixed window afterward. Attach the five-check rubric as the acceptance test.

SourceX does not hold support datasets in stock; it sources them on request from U.S. businesses, after the buyer describes scope, volume, language, format, timing and licensing needs, and it runs qualification, rights review, contracting and preparation with the company that owns the records. Whether a particular supplier's history carries the basis of each concession is something that review has to establish case by case.

Frequently asked questions

What is a courtesy exception in customer support?

It is a concession the company chose to give when no policy, law or contract required it, such as a discount for a customer who was charged correctly. It differs from an owed remedy, which follows a company error or a customer right, and from goodwill a written rule allows for a defined group. In training data the difference has to be a field, because the action looks the same.

Should discretionary concessions be removed from support training data?

No. Removing them leaves a model with no example of when discretion is available, so it either never offers it or invents its own trigger. Keep them, label the basis and the trigger, and test on held-out twins where the trigger is absent.

How can you recover the basis of a concession from old tickets?

Join each ticket to the records that decide basis: the fault or error record, the customer's tier or plan on that date, the policy version in force and any approval. Where none of those explains the concession, label it for review as discretionary or off policy rather than guessing. In ABCD the deciding system answer was not stored as a field, a gap real queues often share.

Can an AI support agent call a credit a one-time courtesy?

Yes, when the credit really is discretionary under the policy in force, and saying so helps stop the customer expecting it again. The statement becomes a problem when it is false, for example when the credit was owed after a company error. Counsel should review standard wording for regulated products.

Is a loyalty-tier perk a policy rule or a courtesy?

It is both, which is why it needs its own label. A written rule authorizes staff to give the benefit to a defined group, but the customer is not owed it the way they are owed a remedy for a company error. ABCD's guidelines frame replacement codes for valued members this way, and 129 of its 700 concessions followed that path.

Sources

  • ABCD's guidelines tie the same promo code and account credit to company error, tier goodwill and customer persistence, and its data, read at commit 6b8700c, yields the 700 concessions counted here. Source
  • The ABCD paper describes 10,042 customer service conversations across 55 intents whose action sequences are constrained by company policy. Source
  • Stripe's credit note reason field takes duplicate, fraudulent, order_change or product_unsatisfactory and may be empty. Source
  • Medusa seeds three default refund reasons, one of which, Customer Care Adjustment, covers both goodwill and compensation for inconvenience. Source
  • Solidus seeds store credit categories Gift Card and Reimbursement and a single store credit reason, Credit Given In Error. Source
  • On Spree's main branch, store credit has no category; the reason a credit exists lives in its issuing record and a free-text memo. Source
  • The FTC's 1983 Policy Statement on Deception defines deception as a material representation, omission or practice likely to mislead a consumer acting reasonably. Source
  • UCC section 1-303 lets a course of dealing between the same parties inform how their agreement is read, with express terms controlling. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies