Skip to content

AI uses for records

Bank fee dispute workflow data: one overdraft case that tests an agent's evidence handling

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

A bank fee dispute tests an agent's evidence handling when the case links the disputed fee to the transaction that triggered it, the consent and notice records the rule depends on, and the ledger entry that settles it. Grade the run on the final system state and the evidence cited, not on a polite reply, and include counterexamples where the customer's claim is wrong.

Agent run

Bank fee dispute workflow

This page covers bank fee dispute workflow data: one overdraft case that tests an agent's evidence handling.Illustrative

Key takeaways

  • A good fee dispute task has records that disagree, such as an opt-in flag with no matching consent capture or confirmation notice.
  • The deciding rule must be part of the packet, because whether a fee is allowed depends on the transaction type and the consent evidence.
  • Success is a ledger reversal or a justified denial linked to the case, checked against system state, not a status field or a closing message.
  • Counterexamples where the fee was valid stop an agent from scoring well by refunding everything.
  • Real packets need linked records from at least five systems, with account and personal identifiers removed before reuse.

How can a bank fee dispute test an agent's handling of evidence?

By giving the agent a dispute whose correct outcome depends on reading several linked records against a stated rule, then checking what the agent changed in the system. An overdraft fee dispute is ideal: the customer's claim is short, the records that settle it live in different systems, and at least one record is likely to look decisive while being incomplete.

The task has five parts: a starting state, a set of evidence items with stable IDs, the rule that applies, the decision the evidence supports and an outcome check against the final system state. An agent passes only if it reaches the supported decision, cites the evidence that supports it, and leaves the ledger, the case record and the customer correspondence in the expected state.

Why is a small fee dispute a strong test of evidence handling?

Because it is small enough to verify completely and messy enough to mislead. A single disputed fee touches the card authorization, the posting and fee engine, the consent record, the notices sent to the customer and the case management system. Each holds part of the truth, and the most visible field, often an opt-in status flag on the account, is not the evidence the rule actually asks for.

Agent benchmarks have moved toward judging outcomes rather than transcripts. The tau-bench paper, for example, scores a run by comparing the database state at the end of the conversation with an annotated goal state. A fee dispute fits that design: the right end state is a specific reversal, or a specific justified denial, and both can be checked mechanically.

Which rules decide the outcome of this dispute?

For a U.S. consumer account, mainly the overdraft opt-in rule in Regulation E. Section 1005.17 bars an institution from charging a fee for paying an ATM or one-time debit card overdraft unless the consumer affirmatively consented, through a consent obtained separately from other agreements, and the institution confirmed that consent in writing, including a statement that the consumer can revoke it. An institution whose procedures ensure fees are charged only on transactions paid after the confirmation was sent is treated as meeting the confirmation requirement.

The rule does not cover every overdraft. Checks and recurring debits fall outside 1005.17, and for them the question becomes whether the fee matches the account disclosure; Regulation DD section 1030.4 requires those disclosures to state each fee's amount and the conditions under which it is charged.

How the bank handles the dispute procedurally is a separate question. If the bank treats the notice as a Regulation E error, section 1005.11 sets the clock: a determination within 10 business days, or up to 45 days with provisional credit, and, where the bank finds no error, a written explanation that tells the consumer they may request the documents relied on. Errors there include a computational or bookkeeping error by the institution relating to an electronic transfer. Whether a fee complaint enters that track or the bank's complaint process is a policy decision the packet must state, not one the agent should guess.

The illustrative dispute packet

Everything below is invented for illustration: names, dates and identifiers are fictional and no money amounts are used. Starting state: a consumer checking account opened in March 2023; a secure message received on day 0 disputing an overdraft fee and stating the customer never agreed to overdraft coverage; the bank's illustrative policy routes fee disputes on electronic transactions into its Regulation E error track.

The illustrative dispute packet
Evidence IDRecord and source systemWhat it shows (illustrative)What the agent must do with it
E1Customer secure message, case systemDisputes one overdraft fee; says no opt-in was ever givenIdentify the disputed fee and the claim to test
E2Card authorization and posting record, card platformOne-time debit card purchase on day minus 6, paid into a negative balanceClassify the transaction type that triggered the fee
E3Fee posting, fee engineOverdraft fee posted day minus 5, linked to the E2 transaction IDConfirm the fee is tied to E2 and not to another item
E4Account profile, core bankingOpt-in status flag reads opted in, last changed day minus 9Treat as a status field, not proof of consent
E5Online banking consent logConsent capture event on day minus 9 with session ID and form versionCheck that consent was captured separately and when
E6Notice history, correspondence systemWritten confirmation of consent generated day minus 2, after E2 was paidCompare the confirmation date with the transaction date
E7Account disclosure version in forceLists the overdraft fee and the transaction types it applies toConfirm the fee was disclosed for this transaction type
E8Bank policy excerptFee disputes on electronic transactions follow the error trackApply the stated deadlines and letter requirements
E9Periodic statement for the cycleShows the fee within total overdraft fees for the period and year to dateKnow what the correction will change downstream

What does a correct run look like, step by step?

The supported decision in this packet is to reverse the fee. Consent was captured on day minus 9, but the written confirmation went out on day minus 2, after the purchase was paid on day minus 6. The customer's claim that she never opted in is wrong, yet the fee still fails, because the bank's illustrative procedures, following the 1005.17 confirmation provision, charge overdraft fees only on transactions paid after the confirmation was sent. An agent that stops at E4 denies the dispute; one that reads E1 alone refunds for the wrong reason. Only the reasoning through E5 and E6 is correct and explainable.

Steps

What does a correct run look like, step by step

Checklist · 8 items

  1. Open the case from E1 and record day 0 as the notice date under E8.
  2. Link E3 to E2 by transaction ID and classify E2 as a one-time debit card transaction covered by the opt-in rule.
  3. Read E4, then look for the consent capture behind it in E5 and the confirmation in E6.
  4. Compare dates: consent before the purchase, confirmation after it.
  5. Decide: reverse the E3 fee; note that consent now exists for future transactions.
  6. Post the reversal linked to the case and to E3, within the E8 deadline.
  7. Send a letter that explains the finding and cites the consent and confirmation dates.
  8. Close the case with the evidence IDs relied on and the reversal reference.

0 of 8 ticked. Tick each item as you go.

What counts as success, and what is it checked against?

Each criterion is checked against system records after the run, not against what the agent said it did. The statement check matters because Regulation DD section 1030.11 requires periodic statements to show total overdraft fees for the period and for the calendar year to date, and the official commentary allows a waived and credited fee to be reflected in the year-to-date figure; the reversal should flow through to that total as the bank's policy specifies.

What counts as success, and what is it checked against?
CriterionPass conditionChecked againstTypical failure
DecisionFee reversed for the E2 transaction onlyLedger after the runDenies because E4 says opted in
Evidence citedCase notes cite E2, E3, E5 and E6 by IDCase recordCites only the customer message
ScopeNo other fees or transactions changedLedger diffReverses every fee in the cycle
TimingReversal and letter within the E8 deadlineEvent timestampsCorrect decision posted after the deadline
Customer letterStates finding and dates; no promises outside policyCorrespondence recordGeneric apology with no reason
Downstream stateStatement total handled as policy specifiesStatement preview or adjustment recordLedger fixed, statement total left inconsistent

Counterexamples that stop an agent from refunding everything

A task set built only from cases where the customer is right teaches and rewards capitulation. Pair this packet with variants that change one record and flip the correct outcome, and score them the same way.

At a glance

Counterexamples that stop an agent from refunding everything

  • Valid opt-in: E6 shows the confirmation sent on day minus 8, before the purchase. The fee stands; the letter must cite the consent and confirmation dates, and, on the error track, tell the customer they may request the documents relied on.
  • Check overdraft: E2 is a paper check, not a debit card purchase. The opt-in rule does not apply; the question is whether E7 disclosed the fee for checks. If it did, the fee stands regardless of E4.
  • Missing evidence: E5 is absent and E4 is the only sign of consent. The agent should not invent a consent date; the supported action is to escalate or reverse under the bank's stated policy for unproven consent.
  • Wrong item: the customer disputes a fee that E3 links to a different transaction. The agent must correct the case scope before deciding.

How should results be scored across many runs and variants?

Per variant and per criterion, with repeated runs, because a single pass hides how often the agent gets the same case wrong. The tau-bench authors introduced a pass^k measure, the chance that all of k independent attempts at a task succeed, to show that agents which solve a task once often fail to solve it consistently. A dispute agent that reverses the right fee in most runs but denies it in some is not ready to touch customer accounts.

Report results by variant so the failure pattern is visible. An agent that passes the reversal case and fails the valid opt-in case is refunding by default; one that passes the opt-in case and fails the check overdraft case is applying the wrong rule to the transaction type. Break the score down by criterion as well: a correct decision posted after the deadline, or a correct ledger with an unexplained letter, is a different defect from a wrong decision.

Keep the expected end state for each variant in a separate answer file that the agent never sees, and version it with the packet. When the bank's policy or the disclosure changes, the expected outcomes may change too, and a stale answer file quietly turns correct runs into failures.

Which records would a real packet need, and what comes out first?

An approved real case needs the same links as the illustrative one: the dispute intake, the triggering transaction, the fee posting, the consent capture event, the notices actually sent, the disclosure version in force, the policy that routed the case, the correction entries and the final letter. Regulation E requires institutions to keep evidence of compliance for not less than two years, so recent dispute histories often have these records, though spread across systems.

Before any reuse, remove or replace account numbers, card numbers, names, addresses, session identifiers and free-text details that identify a customer, and keep the joins by replacing identifiers consistently. Label every packet as real or illustrative, record which system each evidence item came from, and keep the outcome as a system event, not a code typed into a case note.

Access permissions, privacy permission and the right to license are separate questions. That a team can export dispute records, and that the records have been de-identified, does not settle whether the bank or its vendor may license them; contracts and counsel decide that.

Bringing real dispute histories to a licensing review

Linked dispute histories are the kind of operational record AI developers ask for when they build and test agents, and SourceX helps institutions and their service providers license selected ones, covering rights review, contracting and preparation. On volume, licensees tend to favor histories of at least two years and tens of thousands of disputes or more in one system, while rarer work may qualify at a smaller size; every deal gets its own assessment.

Before any case history is exported, an institution can check its eligibility through a short assessment that asks for no data.

Frequently asked questions

Can a bank charge an overdraft fee on a debit card purchase without opt-in?

Under Regulation E section 1005.17, a fee for paying an ATM or one-time debit card overdraft requires the consumer's affirmative consent and a written confirmation of it. The bank may still pay such a transaction without consent, but cannot charge the fee for doing so. Checks and recurring debits are outside that rule.

Why not grade the agent on its reply to the customer?

A reply can sound right while the ledger is wrong, or the reverse. Checking the final system state, the evidence cited and the timing against an expected outcome measures whether the work was done. The letter is one criterion among several.

Why include cases where the customer is wrong?

Without them, an agent can score well by granting every dispute. Counterexamples where a valid consent record or a disclosed check fee supports the charge test whether the agent follows the evidence rather than the customer's framing.

How long does a bank have to resolve a Regulation E error?

Section 1005.11 sets 10 business days to determine whether an error occurred, extendable to 45 days if the account is provisionally credited, with longer periods in some cases. Whether a fee dispute enters that track depends on the claim and the bank's policy, which a task packet should state.

Can the packet use real customer disputes?

Yes, if the institution approves the use, the records are de-identified with joins preserved, and the right to use them for this purpose is confirmed. Label each packet as real or illustrative so reviewers know what the evidence represents.

Sources

  • Regulation E bars fees for paying ATM and one-time debit card overdrafts without affirmative, separately obtained consent confirmed in writing. Source
  • Regulation E error resolution sets 10 business days, up to 45 days with provisional credit, and a notice of the right to request documents relied on. Source
  • Institutions must keep evidence of Regulation E compliance for not less than two years. Source
  • Regulation DD account disclosures must state each fee and the conditions under which it is imposed. Source
  • Periodic statements must show total overdraft fees for the period and year to date, and a waived fee may be reflected. Source
  • tau-bench scores agents by comparing the final database state with an annotated goal state and measures consistency across trials with pass^k. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies