A logistics manager can type “ship from the closest warehouse with available stock” or “hold ground shipments if next-day costs less than $10 more,” and the AI executes it across more than 7,000 locations and 250 million inventory records.
Retailers have used routing software for years. What’s different here is how the decision gets made. Earlier tools followed fixed rules set by a human in advance. This system weighs profitability, inventory and shipping costs in real time and makes a live judgment call on where each order goes. That puts AI in control of whether a customer gets what they expected, not just moving the order through a path someone already mapped.
When that judgment call goes wrong, the customer calls their bank. The payment may have cleared without error. The problem is one step back, in a fulfillment decision the payments chain never saw.
This is the core risk now in front of merchants and acquirers. When an AI agent routes an order to the wrong warehouse, misreads available inventory or trades delivery reliability for a margin gain, the processor records a clean transaction. The merchant gets a dispute with no easy explanation, because the decision that caused it happened in a system with no audit trail the payments chain can read.
Chargebacks911 identified that gap in May. Card networks are rolling out frameworks for AI-initiated transactions, but the infrastructure to handle disputes and assign liability in those transactions is largely unbuilt. “The card networks have built dispute frameworks over decades around one idea: the cardholder either did or did not authorize this transaction,” said Monica Eaton, founder and CEO of Chargebacks911, according to a report from Finopotamus. “In an agentic world, that question doesn’t have a clean answer.”
Datos Insights, in research commissioned by Mastercard, projected that global chargeback volume will grow 24% from 2025 to 2028, reaching 324 million transactions annually, before agentic commerce is fully counted. Fulfillment AI that routes incorrectly after a valid payment shifts dispute exposure into a part of the transaction lifecycle that existing frameworks were never built to reach.
The Protections Being Built Stop at Checkout
Mastercard introduced Verifiable Intent in March. The framework links a consumer’s identity, their specific instructions and the transaction outcome into a single verifiable record any party can consult in a dispute. Google, Fiserv, IBM and Checkout.com are early partners. “If something goes wrong, everyone needs facts, not guesswork,” Pablo Fourez, chief digital officer at Mastercard, said in a statement at the time.
Visa has moved in the same direction. Through Visa Intelligent Commerce, the network issues payment tokens tied to specific agents and specific use cases. A token for a grocery shopping agent cannot authorize a travel booking. One with a $200 limit cannot complete a $500 transaction. Both networks are building toward a point where identity, authorization and payment credentials must be verified before any AI agent can complete a purchase.
Both efforts address what happens at checkout. Salesforce’s Agentic Order Management operates after checkout has already cleared. Warehouse selection, carrier assignment and profit optimization happen in fulfillment, where none of those protections currently reach.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
