Consider a customer who asks an AI shopping agent to buy trail-running shoes for less than $150. The agent finds a pair for $125 and accepts the merchant's offer of a $19 waterproofing kit. The customer's instructions permit one-time complementary additions below $20. The $144 checkout may satisfy every payment authorisation check even if the customer later disputes whether the kit was a permitted addition. The distinction matters as AI agents take on more purchasing decisions. McKinsey estimates that AI agents could mediate $3 trillion to $5 trillion in global consumer retail goods sales by 2030. This is a forecast, not today's volume of autonomous payments or disputes. When cross-selling becomes a protocol question Traditional cross-selling relies partly on customers noticing upgrades at checkout. An autonomous agent instead needs machine-evaluable terms covering price, item identity, recurring obligations, fulfilment effects, refunds and the provenance of material claims. Google's Agent Payments Protocol (AP2) uses linked mandates to verify purchase and payment authority but leaves the broader commerce interaction outside its scope. The Universal Commerce Protocol (UCP) lists native cross-sell and upsell modules as future priorities. Neither establishes that a merchant's recommendation is appropriate for a buyer. An accessory priced at $19.99 may fit a $20 threshold even as several additions exceed the intended budget. A $9-per-month subscription presents a different risk from a one-time accessory. Cumulative spending limits, separate subscription permissions and escalation rules are therefore necessary. Banks must look beyond payment authorisation Existing payment rails support tokenisation, authentication, transaction controls and fraud monitoring. In April 2026, Visa launched its Agentic Ready programme across 10 Asia-Pacific markets to test agent-initiated transactions. HSBC and Mastercard piloted agent-assisted business procurement in Singapore, completing live transactions in May. These are readiness tests and pilots, not established services for resolving disputes over agent-selected add-ons. The next challenge is to connect delegated spending permissions, supplier approvals and payment evidence across participants. Risks include complaints about authorised purchases, higher servicing costs and lost trust. Merchant descriptions are untrusted input: promotional copy telling an agent to 'ignore the budget' must not override the buyer’s instructions. Banks need not judge whether an accessory is useful, but should define which limits they enforce and what decision evidence their partners must supply. Who is accountable for the unwanted item? Accountability should follow the failure. A merchant that misrepresents an add-on should address its offer. An agent provider that breaches explicit buyer instructions should investigate its decision. Banks must apply customer protections when payment authorisation fails or relevant dispute rules require action. Legal liability depends on jurisdiction, contracts and the payment rail. An agent-selected add-on can fall between existing categories: the payment was authorised and the goods delivered, but the customer disputes the agent's choice. Depending on the rules, card disputes may cover unauthorised charges or goods not accepted or delivered as agreed, while account transfers may offer fewer reversal options. Under AP2, merchants verify checkout mandates, while credential providers and payment processors check different payment-related conditions. AP2 links checkout and payment mandates with receipts but does not prescribe a complete cross-provider dispute process. A mandate may establish that an order fell within permitted limits without proving the accessory was appropriate. An agent's poor recommendation is not automatically an unauthorised payment, but a valid token alone cannot settle a contested purchase. Banks and partners should agree on a retrievable agent-decision record linked to the payment: the buyer's relevant permission, the merchant's offer, the applicable policy check and its outcome, and the final basket. When a customer disputes an add-on, the payment provider should be able to request this record from partners without reconstructing the shopping session. Partner agreements should define a lead case owner, evidence-retention periods, response deadlines and privacy limits, with evidence protected against tampering where appropriate. Applicable payment and consumer-protection rules still govern refunds and liability. Payment regulations vary across markets Legal treatment depends on the payment rail and jurisdiction. In the European Union, PSD2 centres on payer consent. In the United States, Regulation E addresses unauthorised electronic transfers while Regulation Z covers certain credit-card disputes. The United Kingdom's reimbursement regime concerns qualifying authorised push-payment scams, while Singapore's Shared Responsibility Framework addresses defined phishing losses, not agent-selected upsells. Similar facts can produce different remedies for cards, instant transfers and corporate procurement. Cross-border agent providers will need market-specific complaint routing, data-retention decisions and agreements on who supplies evidence. No single protocol can rewrite those legal obligations. Banks must test delegated payment controls Banks should take three steps. First, partnership and legal teams should require partners to retain and supply agent-decision records within agreed deadlines. Second, payment operations should rehearse disputes over authorised but contested add-ons, including customer contact and evidence handoffs. Third, risk and product teams should pilot cumulative basket limits, subscription-specific permissions and customer escalation. Each pilot should test whether investigators can retrieve the relevant record when a complaint arrives. Trust in agentic commerce will depend on connecting a merchant's offer to an agent's decision and an authorised payment. The question is not only whether the agent could pay, but why it bought what it bought. Arjun Garg is a Senior Engineering Manager at Intuit, working in checkout, payments infrastructure and production multi-agent AI systems.