Companies mitigate compliance risk in agentic commerce by treating AI agents as regulated actors rather than software features: every agent action needs a verifiable authorization from a human or accountable system, hard limits enforced in code rather than in prompts, and a tamper-evident audit trail that binds each transaction back to its instruction. Existing obligations, from consumer protection to AML to data privacy, do not pause because an agent initiated the transaction. The company that deployed the agent still owns the outcome.
The core mistake I see is treating agent behavior as a prompt-engineering problem. A system prompt telling an agent to stay under a spending cap is a suggestion. A gateway that rejects any transaction over the cap is a control. Regulators and auditors recognize the second kind. Spend limits, merchant restrictions, category scoping, and rate limits belong in the transaction path, where the agent cannot talk its way past them.
The second pillar is authorization evidence. When an agent buys, cancels, or shares data, the company needs to answer a specific question later: who authorized this, and what exactly did they authorize? That means capturing the customer's instruction as a durable, scoped record at delegation time, not reconstructing it from chat logs after a dispute. Without that record, every contested agent transaction defaults to the company's liability, because it cannot prove the customer asked.
Third, the audit trail has to be tamper-evident. Standard application logs can be edited, which makes them weak evidence. Append-only or cryptographically chained logs of agent decisions and actions are cheap to run and materially change a company's position in a regulatory examination or a dispute.
Fourth, plan for disputes before volume arrives. Agentic transactions generate a new dispute category, the agent that misunderstood, which existing chargeback and complaints processes were not designed for. Companies need a defined path for "the agent exceeded its mandate" that is distinct from ordinary buyer's remorse, and records good enough to tell the two apart.
Finally, map the existing rulebook onto agent flows rather than waiting for an agent-specific one. As of September 2026, regulators including the UK's FCA have signaled that agent-initiated payments fall under existing authorization and consumer-duty frameworks, and no major jurisdiction has granted agentic commerce a lighter regime. The practical posture: assume current rules apply, document the mapping, and review vendor agent platforms with the same diligence applied to any outsourced regulated function.
The uncomfortable question is liability allocation between the agent operator, the platform, and the merchant when a compliant-looking transaction still goes wrong. That is unresolved, which is exactly why the evidence trail matters.
Go deeper: how the FCA is approaching agentic payments and the missing governance layer for agent commitments.