What it takes to let a dispute agent move money

Every dispute is a multi-system investigation against two clocks: the network's filing window, which kills the bank's right to recover, and the consumer-protection clock, which does not stop when the network window closes. Reg E runs on business days and forces provisional credit on debit cards; Reg Z runs on calendar days and billing cycles on credit cards. A desk that counts both correctly, every time, is already doing most of what the agent does - by hand, at about $37 a case in this model (roughly 35 minutes of fully loaded analyst time, an illustrative assumption).

The error that matters is not a slow case. It is a missed deadline, a credit posted to a claim that should have been denied, or a denial that should have been a credit. The control design is built around those three.

• A versioned policy, not a model setting

  The CRO sets the evidence threshold (default 0.95), an autonomous-credit ceiling per segment, the file floor and which categories may resolve alone. Every record carries the policy version it ran under.

• Code decides; the model narrates

  Scoring, the gate, filing rules, reason codes, clocks and GL postings are deterministic. A language model writes the analyst's summary from the engine's facts, and a grounding check drops any line that contradicts them.

• Six stop conditions

  Score below threshold, amount above ceiling, sanctions match, no recovery right (lapsed window or 3-D Secure shift), adverse action, missing evidence. Any one stops the agent before money moves.

• Denials are never automated

  A denial is adverse to the cardholder. The agent can recommend one - with the notice and the documents right it triggers - but only an analyst can execute it.

• An independent re-check

  A second implementation recomputes the score, the gate and the posting balance. If it disagrees with the engine, nothing moves.

• Overrides with reason codes

  An analyst can replace any recommendation, but must name the action and a reason code. Override rate by reason is a monitored metric, not an anecdote.

In the demo book this default policy resolves 27 of 38 disputes without an analyst, moves the blended cost to $18 a case (modeled) and executes 0 denials. The book is authored to exercise every outcome, so the ratio illustrates the mechanism; it is not a forecast for any bank's card book.

1. 1

   Shadow mode on closed disputes

   The agent decides every historical case; nothing moves. Compare its outcomes with what analysts did, fit the calibration on the bank's own labels, and replay every regulatory clock.

   Exit when: calibration is fitted and signed off by Model Risk; zero clock misses in replay; agreement on the would-be-autonomous set meets the CRO's target.

2. 2

   Recommendation mode, live

   Every case still goes to an analyst, now with a written recommendation, its postings and its deadline. The override rate becomes the leading indicator.

   Exit when: override rate on the would-be-autonomous set stays below the agreed ceiling for a full month; handle time falls.

3. 3

   Canary autonomy in one category

   Grant autonomy where evidence is strongest - duplicate processing, which shows 96.2% precision in this synthetic backtest - for a small share of volume under a low ceiling.

   Exit when: precision, reversal rate and QA sampling hold; zero Reg E or Reg Z breaches.

4. 4

   Policy-gated autonomy

   Widen categories and ceilings one policy version at a time, each backtested first on the Policy tab's method and monitored after.

   Ongoing: drift in score distribution, override and reversal rates, wrongly-credited exposure, clock breaches - reviewed by Model Risk.

Exit targets are the bank's to set; the gates are the proposal.

The evidence score is a model in the sense of SR 26-2, the model-risk guidance the Federal Reserve, OCC and FDIC issued in April 2026 to supersede SR 11-7. The filing rules are deterministic code, and the narrating language model is generative AI, which SR 26-2 leaves to the bank's own governance. The design produces the evidence a validation of the score asks for; each item below is generated by the product, not written after the fact:

• Conceptual soundness: every score factor, its weight and its rationale, visible per case
• Outcomes analysis: backtest and calibration by category (1,000 labelled disputes in this demo; the bank's own history in production)
• Change control: every policy version is stamped into every record it governed
• Ongoing monitoring: override rate by reason code, reversal rate, wrongly-credited exposure, clock breaches
• Narration governance: model id, prompt hash and grounding result recorded; no cardholder name sent; no decision authority
• Limitations stated: hand-set weights until calibrated on real outcomes; connectors and history in this build are synthetic

In the synthetic backtest the default policy resolves 44.3% of disputes inside policy at 87.8% precision; 54 credits went to claims later labelled invalid. A score of 0.99 does not mean 99%: that gap is exactly why the threshold is set on the calibrated rate, and why the packet reports it.

Vendor descriptions are summarised from each company's public materials, linked above. None of these companies is affiliated with or has reviewed this work.

Representment and pre-arbitration with their network time limits and GL moves; pre-dispute deflection; consumer letters drafted from the recommendation and checked like the narration; and merchant-evidence extraction feeding the score as structured, grounded fields. Each lands behind the same policy gate.