Consumer protection liability when the bot makes the promise
A company deploys an agent that negotiates with customers. It offers a refund the company does not honour. The company's position is that the agent made an error, and consumer protection law is unlikely to find that interesting.
Is a business liable when its AI agent makes a promise it will not honour?
Very likely. Section 5 of the FTC Act reaches representations likely to mislead a reasonable consumer and material to their decision. Intent is generally not required, and nothing in the framework asks what produced the representation. “The model hallucinated” is an explanation, not a defence.
- Consumer protection liability attaches to the business for representations made in its name, regardless of what produced them.
- 'The model hallucinated' is an explanation, not a defence — the business chose to deploy a system that makes representations.
- Bounding what an agent can commit to, and giving the consumer a record of it, addresses both the liability and the dispute.
Part of AI oversight and regulation
The framework, briefly
Consumer protection regimes generally prohibit deceptive acts or practices. A representation is typically deceptive if it is likely to mislead a reasonable consumer and is material to their decision.
Two features matter here. Intent is generally not required. And the business bears responsibility for representations made in its name through its channels.
Why the usual defences do not fit
| Argument | Problem |
|---|---|
| The model produced unintended output | The business chose to deploy a system capable of making representations |
| The consumer should have known it was a bot | The representation was still made in the business's name |
| Terms of service disclaim agent statements | Disclaimers rarely cure a deceptive representation that induced the transaction |
| It was a one-off error | Systems produce errors at a rate; a rate is a practice |
The last row is the one businesses underestimate. An individual employee's misstatement is an incident. A deployed system's error rate across thousands of interactions looks structural, and structural is how these cases are framed.
Four representation types worth bounding
- Price and terms. Anything about cost, duration, renewal or cancellation.
- Remedies. Refunds, credits, replacements, waivers of fees.
- Capability. What the product or service does, particularly in regulated categories.
- Eligibility. Whether a consumer qualifies for something.
These are the classes where a representation is both material and easy for a generative system to produce plausibly. They are also enumerable, which makes them boundable.
Bounding at the boundary, not the prompt
Instructing a model not to promise refunds is a request. Preventing it from issuing one is a control.
# The agent proposes; the boundary decides
def issue_remedy(agent_request, conversation):
policy = remedy_policy_for(conversation.customer, conversation.issue)
if agent_request.type not in policy.permitted_types:
return refuse("outside permitted remedies")
if agent_request.value > policy.max_value:
return escalate_to_human(agent_request, conversation)
remedy = policy.render(agent_request) # canonical, from policy
receipt = sign_issuer(remedy) # business attests
deliver_to_consumer(remedy, receipt)
return remedy
# The consumer's record comes from the policy engine,
# not from the model's description of what it offered.The last comment is the substantive design point. If the consumer's record of what was promised is generated by the same model that made the promise, a dispute about what was said has no independent evidence on either side.
What the consumer should receive
A record that is useful to them, which incidentally is also the record that protects the business.
- What was agreed, in specific terms — amount, timing, conditions
- When, and through which channel
- That it was an automated interaction, where that is not obvious
- Something verifiable, so the record is not simply the business's assertion
A consumer holding this is less likely to complain to a regulator, because they have what they need. A business holding the counterpart is in a far better position if the consumer does.
The disclosure question
Whether a consumer must be told they are interacting with an automated system is an area of active regulatory attention, and requirements differ by jurisdiction and context.
The practical position is that disclosure is cheap and its absence is an easy criticism. A business that discloses clearly and bounds what the agent can commit to has addressed both the disclosure question and the substance, without waiting for the rule to settle.
This describes consumer protection principles at a general level and is not legal advice. Standards, enforcement posture and disclosure requirements vary by jurisdiction; take specifics to counsel.
Four representation classes worth bounding
| Class | Examples |
|---|---|
| Price and terms | Cost, duration, renewal, cancellation |
| Remedies | Refunds, credits, replacements, fee waivers |
| Capability | What the product does, especially in regulated categories |
| Eligibility | Whether a consumer qualifies for something |
These are enumerable, which is what makes them boundable. Instructing a model not to promise refunds is a request; preventing it from issuing one is a control.
Objections and honest limits
“Our terms of service disclaim agent statements.” Disclaimers rarely cure a deceptive representation that induced a transaction. The statement was made in the business's name through the business's channel.
“It was a one-off error.” Systems produce errors at a rate, and a rate is a practice. An individual employee's misstatement is an incident; a deployed system's error rate across thousands of interactions reads as structural, and structural is how these cases are framed.
Disclosure is an area of active regulatory attention and it is cheap. A business that discloses clearly and bounds what the agent can commit to has addressed both the disclosure question and the substance without waiting for the rule to settle.
Bounding what an agent can commit to
- Enumerate the representation classes. Price, remedies, capability, eligibility. Four lists, each short.
- Move the decision to a policy engine. The model proposes; the engine decides what is permitted and at what value.
- Render the consumer's record from the engine. Not from the model's description of what it offered.
- Give the consumer something verifiable. A record they can check is one they are less likely to complain about, and one that protects you if they do.
Terms used here
- Deception
- A representation likely to mislead a reasonable consumer, material to their decision. Intent is generally not an element.
- Material representation
- One likely to affect a consumer's choice. Price, remedies and eligibility almost always qualify.
- Policy engine
- A component that decides what the business is willing to offer, independent of what a model proposes.
Frequently asked questions
Can a business disclaim its agent's statements? Terms of service disclaimers rarely cure a deceptive representation that induced a transaction. The representation was made in the business's name.
Does intent matter? Generally not under deception standards. A representation likely to mislead a reasonable consumer, material to their decision, is the test.
How should agent commitments be bounded? At the system boundary rather than in the prompt. A policy engine decides what remedies are permitted; the model proposes and does not issue.
Why shouldn't the model write the consumer's record? Then a dispute about what was promised has no independent evidence. The record should come from the policy engine that issued the remedy.
Where this fits in Manav
Manav's role here is the consumer-side record: the remedy is rendered from the policy engine, countersigned by the business, and delivered to the consumer in a form either party can verify later without the other's cooperation.
Sources and further reading
- FTC Policy Statement on Deception
- State unfair and deceptive acts and practices statutes.
- Regulatory guidance on automated interaction disclosure.
- Published enforcement actions involving automated customer interactions.
- Federal Trade Commission Act, Section 5