← AI in online sellingC1 · LESSON 6 OF 7

C1, Lesson 6: Checkout, Fraud and Abandoned Carts: The Cost of Saying No

Tags: Sourced / General / Our view / Illustrative. Reading time about 13 minutes.

1. The story: the con artist who taught PayPal to watch

PayPal's early years nearly ended on fraud. A user known inside the company as "Igor" ran a scheme: make a fake sale, pay it from a second account with a stolen credit card, and withdraw the cash before the card owner reported the theft. That left PayPal holding the loss when the payment was reversed. In response PayPal built transaction-monitoring software, unveiled in 2000, and named it Igor. Co-founder and then chief technology officer Max Levchin later said it "was, basically, the reason we survived." [Sourced: American Banker, Aug 28, 2006] Levchin also said simple rules (time of day, payment amount) were easy for criminals to beat. [Sourced: same article] In December 2001 MIT Technology Review reported that PayPal claimed to have brought its fraud rate below 0.5 percent. [Sourced: MIT Technology Review, Dec 1, 2001; a company claim as reported] The lesson is that online checkout is a bet against strangers. Merchants carry the loss, and the tool that protects you also decides who is allowed to buy. [Our view]

2. The one idea

At checkout, AI does one job: it predicts whether this order is risky and assigns a score. The decision, approve, review or decline, is yours. Every decline of a real customer is a lost sale, so the goal is not to catch the most fraud; it is to minimize total cost: fraud losses plus lost good orders plus review time. [General]

3. The kitchen-table version

At your stall a stranger offers a $100 note for a $12 bottle. You squint at it. If you refuse every big note, you lose honest customers. If you accept every note, some are fake. A good stall keeper looks at several things together: how the person acts, whether the note feels right, whether they are a regular. That is a fraud score. [General]

4. Checkout and fraud: how the parts fit

  1. Cart: items, shipping, tax and total on screen.
  2. Checkout: contact, address, payment details.
  3. Authorization: the card network and the customer's bank say yes or no to the charge.
  4. Fraud screen: your tools (or your payment provider's) score the order from signals such as device, address match, order pattern, past history.
  5. Decision: approve, send to manual review, or decline.
  6. Afterwards: a customer can dispute the charge. This is a chargeback: the customer's bank reverses the payment and asks the merchant for evidence. [Sourced: Visa chargeback page, "What is a chargeback?"] Payment providers advertise machine-learning screening that scores each payment on hundreds or thousands of signals in well under a second (Stripe describes more than 1,000 characteristics and a decision in under 100 milliseconds). Treat these as vendor descriptions of their own product. [Sourced: Stripe blog, "How we built it: Stripe Radar"; vendor claim]

5. Why carts are abandoned

Baymard Institute's long-run average documented cart abandonment is about 70%, from many studies. [Sourced: Baymard, see Lesson 1] Not all of it is avoidable: much is browsing and price comparison. Baymard's own survey of 1,012 US adults on February 7, 2024 found that 48% had abandoned a cart because extra costs (shipping, taxes, fees) were too high, the top reason. [Sourced: eMarketer summarizing a Baymard survey, Feb 2024] A separate statistics listing gives other reasons: slow delivery, being forced to create an account, not trusting the site with card details, long checkout, unsatisfactory returns policy. [Sourced: Statista listing; secondary, survey details not confirmed] The practical point: most abandonment fixes are not AI at all. Show the total early, allow guest checkout, keep forms short. AI helps at the margin: recovery emails and offers, error detection, and fraud screening that does not decline the good customers. [Our view]

6. Worked example A: the cost of saying no

All numbers [Illustrative], computed by script. A store takes 10,000 orders a month at $36, 40% margin. 1% are fraud (100 orders). A fraud loss costs the order value plus a $15 chargeback fee: $51. Three settings of the same fraud model:

SettingFraud caughtGood orders wrongly declinedFraud lossLost margin on good ordersTotal cost
No screening0%0%$5,100$0$5,100
Strict80%3%$1,020$4,277$5,297
Balanced65%1%$1,785$1,426$3,211
Lenient50%0.3%$2,550$428$2,978
Read it: the strict setting catches the most fraud and costs more than doing nothing, because it declines 297 good orders. In this illustration the lenient setting wins. With a higher fraud rate or more expensive goods the answer flips. That is why you measure both numbers: fraud losses and false declines. [Our view]
Most teams track fraud losses (they get chargeback notices) and never see false declines (the customer just leaves). Ask your provider for the decline rate on orders later confirmed good.

7. Worked example B: abandoned cart recovery

From Lesson 1: 660 carts, 218 orders, 442 abandoned. Suppose you have emails for 40% (177) and a recovery email converts 8% of those: 14.1 orders, 14.1 x $36 x 0.40 = $204 margin a month. Not much, which is typical of small stores. The bigger prize is fixing the cause (hidden shipping cost) so fewer carts are abandoned at all. [Illustrative]

8. What it costs and what it breaks

  • False declines: lost sales and annoyed customers, usually invisible.
  • Manual review: someone must look at borderline orders; time and delay.
  • Over-trust in a score: criminals adapt; PayPal's lesson was that simple fixed rules get beaten (per Levchin). Models need updating.
  • Fairness: signals like address or device can disadvantage groups of honest customers; ask how your provider checks.
  • Privacy: fraud tools collect device and behavior data; disclose it in your privacy policy.
  • Chargebacks: even when you win the evidence fight, you spend time. Keep order, delivery and communication records.

9. Monday checklist

  1. Pull last month's fraud losses and chargeback fees. Pull the decline count.
  2. Ask your payment provider how many declined orders would have been good.
  3. Check what customers see about shipping cost before checkout.
  4. Offer guest checkout. Count form fields; remove any you do not need.
  5. Decide who reviews borderline orders and within what time.

10. Try it yourself (no code, 30 minutes)

Use the table in Section 6 with your own numbers: order count, average order, margin, fraud rate. Compute total cost for "no screening" and one stricter setting where you assume 2% of good orders are declined. Which wins?

11. Self-check

  1. What was "Igor" at PayPal? The nickname of a fraudster and of the transaction-monitoring system PayPal built in 2000 to stop him.
  2. In our table, why can the strict setting cost more than no screening? It declines many good orders, losing more margin than the fraud it stops.
  3. A fraud loss costs the order ($36) plus a $15 fee. What is it? $51.
  4. What was the top reason for cart abandonment in Baymard's February 2024 survey? Extra costs (shipping, tax, fees), 48%.
  5. Name two non-AI fixes for abandonment. Show the total early and offer guest checkout (or shorten the form).

12. Where this is going

When an AI shopping agent checks out for a person (Lesson 7), your fraud tool sees a new kind of buyer: not a human at a browser. Rules written for human behavior may block legitimate agent orders. Stores will need ways to tell "an authorized agent acting for a real customer" from "a bot." [Our view]

Sources

Review flags

  • RESOLVED (Oct 7, 2026): the Igor story. American Banker (Aug 28, 2006) and MIT Technology Review (Dec 1, 2001) were re-read. American Banker has the scheme (fake sale, stolen card, cash withdrawn before the reversal), "unveiled in 2000," the "was, basically, the reason we survived" quote, and the point that simple rules (time of day, amount) are easy to beat. MIT TR has "less than .5 percent" as what "PayPal claims." Both report PayPal's own account of itself, and the lesson says so. The book "The PayPal Wars" was not read and is not cited.
  • RESOLVED: the 48% figure. eMarketer's page (read) states it as a Baymard survey of 1,012 US adults on Feb 7, 2024; it is a secondary summary and the lesson labels it so. Statista's different list is used only for reasons, not numbers.
  • Stripe's description (1,000+ characteristics) is a vendor claim from Stripe's own blog; labeled.
  • Chargeback fees and rules vary by provider and card network; the $15 fee is illustrative.
  • All worked-example numbers (1% fraud, rates in the table, 8% recovery) are illustrative.

Claim ledger (Lesson 6)

#ClaimSourceVerifiedStatusRe-check
1Igor scheme: fake sale, stolen card, withdraw cash; software unveiled 2000American Banker 20062026-10-06Readnone
2Levchin: "basically, the reason we survived"; simple rules easy to beatAmerican Banker 20062026-10-06Readnone
3PayPal fraud rate claimed under 0.5%MIT Tech Review Dec 20012026-10-06Read excerpt; company claimnone
4Chargeback definitionVisa2026-10-06Readnone
5Stripe Radar 1,000+ characteristics, under 100 msStripe blog2026-10-06Vendor claim, excerpt2027-01-06
648% abandoned for extra costs (Feb 7, 2024, 1,012 US adults)eMarketer on Baymard2026-10-06Secondary summary2027-01-06
7Other reasons listStatista2026-10-06Secondary, excerptnone
8Fraud and recovery arithmeticOur computation2026-10-06Illustrative, script-checkednone