When the Agent Gets It Wrong: Liability, Consent and Recourse in AI-Initiated Commerce

Agentic commerce — autonomous AI agents researching, negotiating and transacting on behalf of consumers and businesses — has crossed from novelty to measurable share of e-commerce volume in 2026. Visa Intelligent Commerce, Mastercard Agent Pay and the OpenAI Operator and Anthropic Claude integrations now collectively process meaningful daily transaction volumes. The merchant-side and issuer-side fraud frameworks for these flows are still catching up.

The harder problem, increasingly, is not preventing fraudulent agents — that is solvable with the authentication and mandate-binding architecture the schemes have specified. The harder problem is what happens when a legitimate agent, operating within its mandate, makes a purchase the consumer regrets, or that turns out to have been mis-prompted, or that exposes the consumer to a fee or commitment they did not consciously approve. Liability and consumer-protection frameworks built for human-initiated commerce do not answer these questions cleanly.

The strategic point is that the regulatory frame around agentic commerce is going to be set in the next 12-18 months. Institutions and merchants engaging with that frame in good faith now will find themselves with workable controls. Those that do not will find themselves in a chargeback environment they are not equipped to defend in.

Regulatory and Market Context

The European Commission's AI Liability Directive, in its post-trilogue form, contains explicit provisions for adverse outcomes from autonomous-agent transactions, with the rebuttable presumption of fault sitting on the deploying party unless mandate, audit-trail and consent evidence is producible. The UK Law Commission's October 2025 paper on AI agents and contract law floats analogous proposals.

In the US, the CFPB's January 2026 advisory on autonomous-agent purchases and Regulation Z framed agent-initiated card transactions as squarely within the existing dispute-and-chargeback regime, with one important addition — the consumer's right of recourse is not extinguished by the existence of the agent mandate, only narrowed where the mandate is appropriately scoped.

What the Data Is Showing

Early agentic-commerce dispute data is small but instructive. Disputes on agent-initiated transactions are running at roughly 2.4x the rate of comparable human-initiated card-not-present transactions, but the composition is different — fewer fraud disputes, more 'did not authorise' and 'not as described' disputes. Issuers are finding that the absence of a clean human-decision moment makes traditional dispute adjudication harder, not impossible.

Merchants accepting agent-initiated transactions are responding by demanding richer mandate metadata at authorisation — agent identity, mandate scope, consent timestamp, originating consumer device. Where this is provided cleanly, dispute losses on agent-initiated transactions are tracking close to human-initiated rates. Where it is not, they are running materially higher.

Implications for Financial Institutions

For issuers, the right model is to treat the mandate metadata as the equivalent of a strong-customer-authentication artefact — store it, surface it on dispute, and rely on it for liability allocation. Issuers that cannot produce mandate metadata on disputed agent-initiated transactions will lose those disputes, and the upstream effects on consumer trust in the issuer are larger than the dispute economics alone suggest.

For acquirers and merchants, the design problem is consent UX — the consumer's pre-authorisation of the agent's mandate has to be specific enough to bind, recent enough to be probative, and clear enough that a reasonable consumer would understand it. Generic 'I authorise this agent to make purchases on my behalf' language will not survive scrutiny under the new frameworks.

Conclusion

The agentic-commerce liability frame is being written now, in the form of issuer dispute decisions, scheme rules and emerging regulatory advisories. The institutions that engage with it as a control-design problem rather than a wait-and-see legal problem are setting the precedents others will follow. The ones that wait will be reading those precedents in their next chargeback report.

Suggested Next Steps

• Define your institution's position on agent-initiated transactions across issuing, acquiring and merchant-services books and document it formally.

• Add mandate-metadata capture and storage to your authorisation and dispute pipelines if not already in place.

• Audit your consent-UX patterns for any agent-initiated commerce flows you support and benchmark against the EC AI Liability Directive evidentiary expectations.

• Engage scheme contacts on the rule changes coming through 2026 for agent-initiated card transactions and prepare a scheme-rules-readiness response.

Sources: European Commission AI Liability Directive (post-trilogue text, 2026); UK Law Commission Paper on AI Agents and Contract Law (Oct 2025); CFPB Advisory on Autonomous-Agent Purchases under Regulation Z (Jan 2026); Visa Intelligent Commerce specification; Mastercard Agent Pay specification.

TrustSphere Risk Index — Vendor Spotlight: Forter

The TrustSphere Risk Index is a quarterly assessment of 221 financial-crime vendors across 8 categories and 11 capability dimensions including data coverage, real-time performance, network analytics, model-risk transparency and integration depth. The March 2026 index update is now available to TrustSphere clients.

In the Risk Orchestration / CNP category, Forter scored 62% in the March 2026 index — strong scores on identity-graph depth, mandate-aware decisioning and merchant-side dispute support. Forter is one of the few orchestration vendors actively building product capability around agent-initiated commerce as a distinct flow rather than treating it as a special case of CNP.

For institutions evaluating providers in this space, Forter is one of several credible options — vendor fit depends heavily on existing architecture, deployment model and downstream tooling. Contact TrustSphere for a comprehensive vendor suitability assessment tailored to your institution.

TrustSphere helps financial institutions design and deploy intelligent fraud and financial crime detection solutions. Visit www.trustsphere.ai