Key points

• AI agents that resolve refunds, change orders and answer policy questions have moved from pilots to production, and the vendor contract is where liability for their errors is actually allocated.
• A company is generally responsible to its customers for what its customer-facing agent says and does, regardless of whether a vendor built it. Tribunals have already rejected the argument that a chatbot is a separate entity.
• Standard software-as-a-service terms were written for tools that return results to a human. They fit poorly when the system takes actions, commits the company to outcomes or handles disputes.
• The terms that matter most are authority limits, accuracy and testing commitments, incident notice, data use, and a liability cap that reflects the risks an agent can create.
• Retailers and consumer brands also need to map the agent against consumer protection rules on refunds, pricing statements and complaint handling.

From chatbot to agent

Customer service automation used to mean a scripted bot that routed questions. The current generation is different. An agent can read an order history, interpret a policy, issue a refund, change a delivery address or cancel a subscription, and it does so by reasoning over a prompt and a set of tools rather than following a fixed decision tree. That flexibility is the commercial appeal. It is also what changes the legal analysis, because the system can take actions that nobody specifically approved in advance.

For retailers and other consumer-facing businesses, the volume matters too. An error that would be a single bad call in a human contact center can repeat across thousands of conversations before anyone notices.

Who answers to the customer

The starting point is that the company deploying the agent answers to its customers. In a widely reported 2024 decision, a Canadian tribunal held an airline responsible for incorrect refund information given by its website chatbot and rejected the argument that the chatbot was a separate legal entity responsible for its own statements. The reasoning is straightforward, and we expect other tribunals and regulators to follow it. A customer is entitled to rely on what a company's channel tells them, and the company cannot point to its supplier.

Consumer protection regulators take a similar view. In the United States, the FTC has said that existing rules on unfair and deceptive practices apply to AI-driven interactions, and state attorneys general have signaled the same. In the UK, the CMA's enhanced direct enforcement powers over consumer law apply to a business's conduct however it is delivered. The practical consequence is that the allocation of risk with the vendor has to be negotiated, because it will not come from the customer relationship.

Where standard vendor terms fall short

Most software contracts assume that a human uses a tool and takes responsibility for the output. Disclaimers that outputs may be inaccurate and must be reviewed make sense for a drafting assistant. They make little sense for an agent that talks to customers without review. A vendor who disclaims all responsibility for accuracy while marketing the product as autonomous is asking the customer to take a risk the customer cannot control.

Liability caps are the second pressure point. A cap set at twelve months of fees may be adequate for a failure to deliver a service. It is unlikely to cover the cost of regulatory remediation, mass refunds or a class claim that follows a systematic error. Buyers should ask which categories of loss are excluded from the cap and whether the vendor carries insurance that matches the exposure.

Terms to negotiate

Scope of authority. Define exactly what the agent may do, including refund limits, the products and policies it may speak to and the circumstances in which it must hand off to a human. Authority limits should be enforced technically, not only stated in documentation.

Testing and performance commitments. Require the vendor to test against the buyer's policies before launch and after material model or prompt changes. Where possible, set measurable thresholds for accuracy and escalation, with a right to suspend the agent if they are not met.

Change notice. Agents built on third-party models can change behavior when the underlying model is updated. The contract should require advance notice of material changes and a way to roll back.

Logging and audit. Retain full conversation records and the actions taken, in a form that supports complaint handling, regulatory requests and litigation. Clarify who holds the logs and for how long.

Incident response. Define what counts as an incident, such as repeated misstatements of policy or unauthorized commitments, and require prompt notice, cooperation and remediation at the vendor's cost where the vendor is at fault.

Data use. Customer conversations contain personal data and sometimes payment or health details. Limit the vendor's use of that data for model training, require appropriate security and confirm where processing takes place.

Indemnities and caps. Seek indemnities for third-party intellectual property claims and for regulatory penalties caused by vendor breach, and negotiate higher caps or super-caps for data, confidentiality and these agent-specific risks.

Consumer law checks inside the company

Contract terms address risk transfer, not compliance. Before launch, teams should check that the agent's statements about pricing, refunds, delivery and cancellation are consistent with the company's own terms and with consumer rules, including cancellation and subscription requirements. Customers should be told when they are dealing with an automated system, and there should be an accessible route to a person. Transparency obligations for chatbots are also coming into force in the European Union, as discussed in our note on EU AI Act transparency duties.

Complaint data should be monitored. Early signals, such as spikes in refund disputes or escalations about a single policy, are often the first sign that an agent is misreading instructions.

A practical sequence

We suggest that retailers treat agent deployment as a staged process. Begin with a narrow set of actions and low-value limits, test against real policies, and widen authority as results justify it. Keep the contract flexible enough to support that progression, with rights to change scope and to terminate for repeated failures. Our related note on agentic commerce covers the reverse situation, in which agents act for the consumer.

Corvane Ryle's Artificial Intelligence practice and Technology Transactions practice negotiate these arrangements for retailers and technology vendors, and Kenji Watanabe-Price can discuss how current market terms compare.