Try it: For judges: 3 cases, 90 seconds · The desk · Open a real sandbox chargeback · What was measured · AG Studio dashboard · Code on GitHub

What it does, in one example. A buyer's AI shopping agent bought $180 of dinner plates, and the cardholder told their bank they didn't authorise it (PP-R-IPX-10195230). Exhibit drafted a six-sentence answer. Every sentence cited a PayPal or store record, and a checker confirmed every figure in it. A tool that only cites its sources would have filed that draft. Exhibit didn't file it yet. A second model, playing PayPal's reviewer, asked what evidence the card networks actually recognise. Two of the sentences leaned on the agent's signed mandate, which the networks don't yet treat as proof, so Exhibit withdrew them. It filed on the stronger evidence instead: two earlier, never-disputed orders from the same device and IP address. When the reviewer pressed on those, Exhibit's answer was kept, and two of three hearings found for the seller.

For judges, in 30 seconds

  • Who it's for: small online sellers who take PayPal and lose chargebacks they could win, because the proof is scattered and writing it up takes an hour.
  • What it does: opens a real PayPal sandbox chargeback in one click, gathers the records from PayPal, writes the response with GPT-5.6 (every sentence citing a record), checks every figure in code, has a second model cross-examine the response, and then files it with PayPal, or refunds when the seller would lose.
  • Measured on GPT-5.6, 30 Oct, committed in the repo:
    • 16 of 18 held-out runs where the deciding fact is only in an email or listing (9 cases × 2 customers, written and labelled by a separate AI agent that never saw the prompt), against 2 for a rules baseline that reads every structured record (per 100 of those disputes, the rules keep at most $1,511).
    • 55 of 56 on the labelled main set, all 56 re-run on the final version (56 on 25 Oct, 53 on 29 Oct).
    • 1,028 of 1,049 true sentences with one fact changed were struck before filing.
    • At 28 measured hearings, 7 overruled answers held up: 3 against the reviewer's own objections, 4 against items the requirements card pressed. 16 objections were sustained. On 12 labelled fights, the first two hearings agreed on 10.
  • What it saves, honestly: on the held-out free-text cases, always fighting keeps more money per 100 disputes ($5,478 against Exhibit's $5,178). Exhibit's value is in the work and the risk: 56 fewer responses to write per 100, and 0 fights on cases the seller should refund, against 44 for always fighting. If a response takes the seller 30 minutes (an assumption, valued at $25 an hour), Exhibit comes out ahead after fees and time: $2,628 against $2,228.
  • Time, from the seller's lane's own records (13 cases): a median case was argued 10 s after PayPal's webhook, and the median fight was filed 17 min after it (8 filed; 5 refunded). 10 of the 13 needed no human step.

How it differs from other dispute desks. Many tools draft a response and attach evidence. Exhibit also tests the response the way PayPal will read it: on the PDF PayPal stores for the case, against the card network's requirements for that reason code, keeping an answer only when code can match it to a record. And it asks PayPal for a ruling only when two of those hearings agree.

How it works

  • One click, a real case. Exhibit takes a real card payment to a demo shop, adds tracking, and has the card issuer file a chargeback through PayPal's sandbox. Two of the seven stories are purchases by a buyer's AI agent. At /import a seller can paste their own order, tracking and emails instead; nothing is filed.
  • Argued before you arrive. PayPal's signed webhook wakes Exhibit. It gathers the order, payment, tracking, carrier scans, emails and listing as lettered exhibits, and GPT-5.6 writes the response one sentence at a time, each citing its exhibits.
  • Every sentence checked in code. Every amount, date, time, id, ZIP, carrier, signer and card number must be in the records the sentence cites. Quotes must match, and a refund can't be cited as the amount paid. A sentence with nothing to check is marked unverified, and the seller's own notes about unanswered emails are kept out.
  • Cross-examined before filing. A second model reads only the PDF PayPal will hold and objects where the network's requirements card is weak or missing (a "requirements card" lists what Visa or Mastercard asks for on that reason code). Exhibit answers from the records or concedes. The checker keeps an answer (OVERRULED) or doesn't (SUSTAINED). You can take the reviewer's chair yourself.
  • Filed, or refunded. The seller's choice goes through PayPal's Disputes API with a Bates-stamped PDF, or as an accepted claim. On a card chargeback a concession says the money stays with the cardholder and no further refund is due.
  • An autopilot that works through PayPal's Agent Toolkit. On the seller's own lane, the autopilot reads each case through the toolkit, holds a hearing, and files or refunds under written house rules, with a ten-minute undo window. It puts tracking up with create_shipment_tracking and refunds with accept_dispute_claim. Then it asks for the ruling.
  • On a desk. The AG Grid desk shows every dispute and who decided each outcome; the AG Studio dashboard has custom widgets, cross-filtering and a recorded analyst run. A seller's own agent can run the desk through an MCP server that sits next to PayPal's.

Measured: 55 of 56 right calls, all 56 re-run on the final version.

How we built it

  • PayPal (sandbox): Orders v2, Payments v2, Disputes v1 (provide-evidence with a PDF, accept-claim, adjudicate), webhooks verified with verify-webhook-signature, the chargeback simulator, and PayPal's Agent Toolkit (five reads, two writes). Section 6 has the detail.
  • AI: GPT-5.6 on Azure OpenAI, through one forced, strict function call. A rules engine argues through the same checks when no model is available.
  • Deterministic code decides: the checker, the requirements cards, the hearing's rulings, the two-of-three ruling panel, the mandate check and the house rules, all pure functions with unit tests in CI.
  • App: Next.js 15 and React 19 on Vercel, AG Grid, AG Studio 3.0, MCP over Streamable HTTP, pdf-lib, and Google Cloud Storage reached keylessly through Vercel's OIDC token. The design borrows from litigation paperwork: exhibit stickers, Bates numbers and a red ink stamp.

Measured: 1,028 of 1,049 altered true sentences struck before filing.

Challenges we ran into

  • Opening disputes by script. The buyer-side call needs consent we couldn't script, and card payments failed on a non-US sandbox account. A US account, one-call card checkouts and the chargeback simulator fixed it.
  • A reviewer we had to keep honest. The first reviewer agreed too easily. Driving it from the requirements card made it press every weak point, and then it struck a true answer, which we fixed and tested. One hearing's verdict also wasn't stable from run to run, so a ruling now needs two hearings that agree. We tuned on the same cases, and /proof says so.
  • The headline moved, and we published every move: 56 of 56 on 25 Oct, 53 of 56 on 29 Oct when the requirements card arrived, then 55 of 56 when all 56 were re-run on the final version.
  • The sandbox doesn't rule on every case. It hasn't ruled on any unauthorised-transaction chargeback Exhibit asked about (PP-R-JWB-10194389, PP-R-UPI-10195190, PP-R-BFL-10195216, PP-R-IPX-10195230), and two not-received requests (PP-R-UVT-10195188, PP-R-YAC-10195241) are still open. Each case, /proof and the home page say so.
  • A stipulated history is the fair test of an agent purchase. To show the earlier-orders evidence, a returning customer's two earlier orders are re-created as sandbox payments when the case opens. The reviewer reads the store's history as given, the way it would read a real seller's order book. The note on how it was re-created sits on the case page, with each payment's creation time.

Measured: 16 of 18 held-out free-text runs, against 2 for the rules.

Accomplishments that we're proud of

  • Full loops through PayPal's lifecycle.
    • PP-R-OHE-10195198 (charged twice) had no human in it. It was argued 10 s after PayPal's webhook and filed by the autopilot. After cross-examination on PayPal's copy, PayPal ruled for the seller at the autopilot's request.
    • On PP-R-HWG-10195220 (not as described), the first hearing went for the buyer, so the autopilot left the case to the seller. The seller filed. On PayPal's copy the reviewer raised its own objection, and Exhibit's answer from the listing and the buyer's own email was kept. PayPal ruled for the seller at Exhibit's request.
  • AI-agent purchases argued on the card networks' terms.
    • IPX is the example at the top of this story.
    • On PP-R-YAC-10195241 the agent bought within its mandate, and the cardholder says the parcel never came, a reason the sandbox does rule on. With no human step, the autopilot filed it 14 min after PayPal's webhook. Both hearings on PayPal's copy found for the seller, with the tracking number checked against the record. The ruling was asked at 03:40 UTC on 30 Oct; PayPal hasn't ruled.
  • A loss we asked for. On PP-R-SMJ-10195232 (not received, $27, no tracking), Exhibit advised a refund, and the seller filed a concession anyway. Both hearings went for the buyer, so Exhibit asked PayPal to rule for the buyer, and it did. The wording of that concession was corrected afterwards so it can't invite a second refund, and the case says so.
  • An AI buyer agent that refused a listing's injected instruction to overspend.

Measured: 10 of 13 seller's-lane cases needed no human step.

What we learned

  • PayPal's dispute lifecycle in detail: which links appear in which state, that the sandbox closes many cases on its own, and that unauthorised-transaction cases wait after a ruling request.
  • Structured output plus a deterministic checker beats asking a model to "be accurate". And measure the checker too: our eval found its mistakes.
  • A signed AP2 mandate decides who is responsible, but it isn't yet authorisation evidence the card networks recognise.
  • Your own database is a cache. The platform's record is the truth.

What's next for Exhibit

  • Connect a real store (Shopify, WooCommerce) directly, beyond /import.
  • Inquiry-stage tools for PayPal-account disputes: messages and partial-refund offers (the calls are written; they need cases opened as the buyer).
  • A held-out set labelled by working chargeback analysts.

Built by Rohan Borade and Aatmaja.

Built With

  • ag-grid
  • ag-studio
  • azure
  • azure-openai
  • ed25519
  • github-actions
  • google-cloud
  • gpt-5.6
  • model-context-protocol
  • next.js
  • openai
  • paypal
  • paypal-agent-toolkit
  • paypal-disputes-api
  • paypal-orders-api
  • paypal-payments-api
  • paypal-sandbox
  • paypal-tracking-api
  • paypal-webhooks
  • pdf-lib
  • react
  • typescript
  • vercel
Share this project:

Updates

Submission history