Order Fraud Review: Every Flagged Order Released or Cancelled on Its Evidence

A flagged order waits while the evidence sits in three places

Every order the fraud screen flags waits for a person while the customer waits for a shipping email. The review is the same each time: look at the charge's address check and card security code result, see whether earlier charges on the same customer were refunded or disputed, check whether the customer has verified, then release, hold or cancel under the rule order. Speed matters, and so does being right in both directions.

Flagged orders are tracked in Attio, charges and refunds live in Stripe under the customer rather than the order, and the review policy and a generated worksheet sit in Google Drive. Customers answer verification requests by email, the payments team posts chargeback news, and the fraud lead occasionally moves the auto-cancel score line. When that line moves, every order in the queue has to be looked at again.

Who works in the room

Two agents share the fraud room. Nova is the default, so a question about any order's decision goes to it unnamed. Nova reads Stripe through a lasting grant the payments lead approved for this room, and looks up charges by each order's Stripe customer id, never by the order number. Ada owns the worksheet check and the customer contact. When an approved decision needs a customer to verify, Nova mentions Ada with the order and what is missing, and Ada drafts the request.

The review in seven steps

The rule order, the score line and the recording fields are in Google Drive and in the room guidelines. Each step ends with a person.

A lower score line means a full re-check

A lower auto-cancel line looks as if it only affects orders just above it. Because the score check sits inside the full rule order, Nova re-derives every order in scope from the top rather than filtering for scores near the line, and lists each decision that changed along with the rule that changed it. The fraud lead approves the set in one pass.

Corrections and replies are handled the same way. A corrected card check on one charge can change the decision on that customer's other flagged orders, so the re-derivation follows the customer, not the message. Every one of these turns is kept in the run history, so when a customer or the payments team asks why an order was cancelled, the answer is a record rather than a recollection.

What the checks catch

• A Stripe charge matched to an order by order number instead of customer id, which pulls in another customer's evidence
• A worksheet flag set on an order the worksheet never covered
• A lowered score line applied only to orders near the line, leaving the rest decided under the old one
• A verification reply matched to the wrong order when one customer has several flagged
• Two replies on the chargeback thread, or one that leaves out a cancelled order

What stays with people

• Approving every release, hold and cancel before Attio is written
• Releasing or cancelling the order in the store
• Sending verification requests and thread replies, which Ada prepares as drafts in Gmail
• Setting the auto-cancel score line

Integrations in this workflow

Kylon is not a fraud screening tool, and it does not replace the one that flags your orders. It sits across the records and payment data the review team already uses, so the coordination between them has one home.

• Attio
• Gmail
• Google Drive
• Slack
• Stripe

• Attio tracks each flagged order with its score, decision, reason and worksheet flag
• Stripe provides each customer's charges with address and card check results, refunds and disputes
• Google Drive keeps the review policy, the rule order and the generated worksheet
• Gmail carries verification requests and the chargeback thread, and each message is a draft an analyst sends
• Slack is where the payments team posts processed chargebacks and analysts flag urgent orders

Where to start

Start with one day's flagged queue. Connect the Attio list of flagged orders and give the room a lasting grant on Stripe, then spell out one rule in the room guidelines: charges are found by customer id, every change to the score line re-derives the whole queue, and the fraud lead approves every decision before it is written.

“A flagged order gets the same full check whether it is first in the queue or fortieth.”

Set it up with your agent

Open a new Room, add your agent and paste this prompt. It connects the tools, builds the App with you and asks for your rules one step at a time, and nothing goes out until a person approves it.