Plaid CEO Warns of a 'Scary Year' for AI Fraud: Merchant Defenses
Plaid's CEO expects a "scary year" for AI fraud. See how merchants in 2026 can tighten checkout, staff and dispute controls before losses climb.
Photo: Pixabay
On this page 7 sections
- What Plaid's CEO Said at Money20/20
- How AI Fraud Targets Merchants in 2026
- Why AI Fraud Shows up as Chargebacks and Monitoring Risk
- Account Opening and Checkout Signals Worth Adding Now
- Fighting AI With AI: What to Ask Your Processor's Fraud Tools
- Training Staff Against AI Voice and Document Fakes
- A 30-Day AI Fraud Hardening Plan
Key takeaways
- Perret called Deloitte's forecast of $40 billion in U.S. fraud losses from generative AI "conservative." Merchants take their share of those losses as chargebacks and fraud reports.
- Plaid's Pedro Sanzovo separated identity fraud from "intent fraud," where people commit fraud under their real identities. He noted that anyone can now ask an LLM how to file a chargeback, and that fake supporting documents are now easy to make.
- In an illustrative example, the same 45-order attack moves the dispute ratio ten times as much at a store with 3,000 monthly transactions as at one with 30,000.
- Before staff change bank details, send a refund anywhere other than the original card, or rush a large phone order, they confirm the request through a second channel you already trust.
Plaid CEO Zach Perret told a Money20/20 audience in Las Vegas on Tuesday, Oct. 20, 2026, that the industry is racing to fight AI fraud with AI and that he expects a rough year for security. For merchants in 2026, AI fraud mostly comes through doors they already know: card testing, fake accounts, stolen-card orders and disputes that read more convincingly than before. The defenses are familiar too. They need tighter settings, better evidence and staff who know what a cloned voice or a doctored receipt can do.
What Plaid's CEO Said at Money20/20
Banking Dive reported Perret's remarks on Oct. 22, 2026. He said Plaid's leadership talks about security and fraud routinely, and he admitted he is worried: "I think that we're going to have a scary year, from a security perspective." He summed up the industry's response this way: "We're just in a dead sprint using AI to fight AI fraud."
Two details matter for merchants. First, Banking Dive reported that Perret called a Deloitte projection "a conservative estimate." That projection says generative AI could push U.S. fraud losses to $40 billion by next year. Second, he said "agent swarms" attacking banks worry him more than deepfakes or other AI-enabled fraud. Plaid connects consumer accounts at roughly 12,000 institutions to apps and services, so it sees fraud from the account-linking side.
A separate Money20/20 session filled in the picture. Credit Union Daily reported that Plaid's Pedro Sanzovo split the threat into two kinds. Identity fraud is where deepfakes have surged. "Intent fraud" is where people commit fraud under their real identities. Merchants should pay closest attention to the second kind. As Sanzovo put it, "You can now ask an LLM how to file for a chargeback." The same week, according to Credit Union Daily, the largest banks said they were finding unauthorized third-party AI agents operating inside customer accounts.
How AI Fraud Targets Merchants in 2026
Banks absorb some of these losses, but much of the money leaves through checkout. Online, keyed and recurring sales take the brunt because nobody presents a physical card.
Synthetic Identities That Pass Simple Checks
A synthetic identity mixes real and invented data, such as a real address with a made-up name, or a valid-looking email built for one purchase. AI makes these identities cheap to produce in bulk, with matching names, phone numbers and social profiles. Online marketplaces see them as sellers who onboard cleanly, collect payouts and disappear. Subscription and digital-product businesses see them as fresh accounts that buy once with a stolen card and never come back.
Deepfaked Voices and Documents
With a cloned voice, a fraudster can call your support line as a "regular customer" placing a rush phone order. The same fraudster can call a manager as the owner and ask for a payout account change. Sanzovo also said fake documents to support claims are now easy to create. In practice that means invoices, ID images and "damaged item" photos that look right at a glance. Our guide to how merchants fight deepfake account takeovers covers layered defenses.
Bots That Test Cards and Write Disputes
Card testing is the old problem of scripts running stolen card numbers through small authorizations. With AI, those scripts adapt faster to rate limits and CAPTCHAs. Our guide to detecting and stopping card testing attacks covers the warning signs. The newer twist is on the dispute side. A real cardholder can ask a chatbot to draft a persuasive "I never authorized this" claim for a subscription they used all month. That is friendly fraud with better writing.
Why AI Fraud Shows up as Chargebacks and Monitoring Risk
When a stolen-card order gets through, the merchant usually loses the sale to a chargeback. The issuer may also file a fraud report against the same transaction. Visa's 2025 fact sheet for its Acquirer Monitoring Program (VAMP) shows that the program counts both. The VAMP ratio adds fraud reports (TC40s) and disputes (TC15s) on card-not-present transactions, then divides by the count of settled transactions. So one bad order can count twice. Disputes resolved through pre-dispute solutions are excluded.
Visa flags a merchant only after it also passes a minimum monthly count of fraud and disputes. Because of that, the limit a smaller store hits first is often its acquirer's own. As Chargeback Gurus explains, acquirers can set merchant thresholds stricter than Visa's to keep their whole portfolio in line. VAMP is a network rule. Your acquirer's internal limits and your processor's dispute fees are that company's own policy and pricing, and they vary. Our explainer on Visa's VAMP thresholds covers where Visa's lines sit.
The same attack does very different damage depending on your size. In the illustrative example below, every number is an assumption, not data:
| Assumption | Store A | Store B |
|---|---|---|
| Settled card-not-present transactions per month | 3,000 | 30,000 |
| Average ticket (digital course) | $49 | $49 |
| Fraudulent orders from one bot-created ring | 45 | 45 |
| Lost revenue (45 × $49) | $2,205 | $2,205 |
| Dispute fees at an assumed $20 each | $900 | $900 |
| Disputes as a share of settled transactions | 1.5% | 0.15% |
Both stores lose the same amount, about $3,105 before staff time. If issuers also file fraud reports on those orders, Visa's count could include those as well, so the ratio could go higher than the table shows. For Store A, one bad weekend creates the kind of number that trips an acquirer's internal limits. For Store B, it barely moves the ratio. Timing makes it worse, because disputes arrive weeks after the orders. By the time Store A sees its first chargeback, the ring may have run the same play again.
If your monthly volume looks more like Store A's, set tighter limits on new accounts and order velocity than a large retailer would. Track your numbers with our guide to calculating your chargeback ratio.
Account Opening and Checkout Signals Worth Adding Now
Most of these controls are probably already in your gateway or fraud tool. They just need to be switched on and set for each sales channel.
- Account creation: put a bot challenge on sign-up, flag disposable email domains, and limit how many accounts one device or IP address can open.
- Online checkout: require CVV, act on address verification (AVS) mismatches instead of only logging them, and set velocity limits per card, device, IP and email.
- Authentication: send risk-flagged orders through 3-D Secure authentication so the card issuer checks the cardholder. In many cases that also shifts liability for fraud chargebacks to the issuer. Our 3-D Secure 2 explainer covers the friction trade-off.
- Digital goods: hold delivery of high-value first orders for a short review, and cap first-order value on new accounts.
- Subscriptions and trials: limit trials per card and per device, and watch the trial sign-up page for bursts of failed authorizations.
- Marketplaces: confirm that sellers own the bank account they want paid to, use ID checks with liveness detection, and hold first payouts until a seller's buyers have had time to dispute.
- Keyed phone orders: screen a virtual terminal sale as fully as any card-not-present order, because that is what it is.
Fighting AI With AI: What to Ask Your Processor's Fraud Tools
Perret's point about fighting AI with AI applies to merchants too. But an "AI" label on a vendor's tool tells you little about what it actually does. Ask these questions before you renew or switch:
- Does the tool score every transaction using device and behavioral data, or only card and address fields?
- Can I set different rules for online, keyed and recurring transactions?
- Can 3-D Secure be triggered by risk score, instead of being on or off for everything?
- Who tunes the rules after launch, and how often are they reviewed?
- What false-decline rate should I expect, and can I see declines broken out by rule?
- Do you offer dispute alerts or pre-dispute resolution, so I can refund before a chargeback is filed?
- Will I get reports on fraud and dispute counts by card network, so I see a monitoring problem before my acquirer flags it?
Question six matters because Visa's fact sheet excludes disputes resolved through pre-dispute solutions from the VAMP ratio. Question five matters just as much. Rules set too tight turn away good customers, as our piece on fixing false declines to improve authorization rates explains.
Not all automated traffic is hostile anymore, either. As AI shopping agents start buying on customers' behalf, blocking every bot can mean blocking real sales. See our coverage of Visa and Mastercard's agentic commerce plans. For a clear view of where machine learning works and where it fails, read what AI fraud detection catches and misses.
Training Staff Against AI Voice and Document Fakes
Deepfakes target people, so the fix is a procedure rather than another piece of software. Write down three rules and walk through them with everyone who handles orders, refunds or bank details.
The first rule covers any request to change payout accounts, vendor bank details or merchant account settings. Confirm it by calling back a number already on file, never the number the caller gives, and have a second person approve it. That applies even when the voice sounds exactly like the owner. A short code phrase shared offline adds another check.
Under the second rule, refunds go only to the original payment method. A caller who wants a refund sent to a "new card" because the old one was lost is a known pattern, and a convincing voice doesn't make the request safer.
The third rule covers documents and photos sent with refund requests or seller applications. Treat them as claims to check, not as proof. Compare them against your own order records, delivery confirmations and account activity. That same internal record becomes your evidence if the customer disputes anyway. Keep login history, IP addresses, download or usage logs and the customer's acceptance of your terms ready for a chargeback rebuttal letter.
A 30-Day AI Fraud Hardening Plan
Week 1: Measure. Pull 90 days of disputes by reason code, any fraud reports your processor can show you, and daily authorization attempts. Calculate your ratio, and ask your processor which limits apply to your account: Visa's and any stricter internal ones.
Week 2: Tighten checkout. Turn on CVV and AVS enforcement, velocity limits, and bot challenges on sign-up and checkout. Send risky orders through 3-D Secure. Cap new accounts on digital goods and trials.
Week 3: Lock down people and payouts. Adopt the three staff rules: callbacks, two-person approval and refunds only to the original payment method. Run one practice scenario with staff, such as a fake "owner" call asking for a bank change.
Week 4: Prepare for disputes. Build an evidence template for each product type. Check that customers will recognize your billing descriptor on their statements. Send pre-renewal reminders where state law requires them. Our article on auto-renewal disclosure laws is general information, so confirm your obligations with an attorney. Then sit down with your processor and go through the seven questions above.
If that review shows gaps, it's worth comparing options. MannyPay provides online payment processing through gateways including Authorize.net and NMI, offers 3-D Secure authentication, and provides high-risk merchant accounts. You can request a free quote, and a specialist will reply within one business day. Whoever processes for you, do week one first. You can't set sensible limits until you know your ratio and the limit it's measured against.
---
I fixed the closing paragraph as you asked. It now opens with "If that review shows gaps, it's worth comparing options," so it no longer suggests MannyPay offers 3-D Secure triggered by risk score or fraud rules per channel. The MannyPay claims that remain (Authorize.net/NMI, 3-D Secure, high-risk accounts, the one-business-day reply) all come from the house facts.
Search limit: I ran out of searches partway through my recheck of the sources, so I could only re-confirm the Banking Dive article. Credit Union Daily, Visa's VAMP fact sheet and Chargeback Gurus are carried over from the earlier version without a fresh check this round. Someone should confirm them before this publishes.
I made two other changes:
- Perret quote: The Banking Dive article gives his line as "We're just in a dead sprint using AI to fight AI fraud." The earlier version quoted different wording from Big Fintech, which I couldn't re-confirm. I now quote Banking Dive's version and removed Big Fintech from the sources.
- Editor notes: I added a question asking the payments team whether MannyPay can trigger 3-D Secure by risk score and set fraud rules per channel. If they can and it's verified, adding it to the house facts would let the closing say so directly.
Sources
- Banking Dive, "Plaid CEO: AI's rise could fuel 'scary year' for financial security" (Oct. 22, 2026)
- Credit Union Daily, "The Big Worry (and Promise) That Can be Seen With 20/20 Vision"
- Visa, "Visa Acquirer Monitoring Program Fact Sheet" (2025)
- Chargeback Gurus, "Visa Acquirer Monitoring Program (VAMP)"
This article is for general educational purposes and is not legal, tax or financial advice. Card network rules and state laws change; confirm current requirements for your business.
