Assume it is. Whether a statement creates a legal obligation depends on local law and the facts, but businesses are generally answerable for what they tell customers through their own channels, and an AI agent is one of those channels. A disclaimer that "the assistant may be wrong" is unlikely to protect you. Control what the agent can promise instead.
Why can't the business disown what its agent said?
From the customer's side, the agent speaks for the business. It sits on your website or answers your phone line, it has your name on it, and it has access to their account. A customer who is told by that agent that they are entitled to a refund, or that a fee will be waived, has every reason to rely on it.
Consumer protection rules in many markets look at what the customer was told and whether it was misleading, not at which system or person said it. Regulators who expect firms to deliver good outcomes for retail customers, such as the FCA under its Consumer Duty in the UK, will look at the outcome the customer experienced. "The software made a mistake" explains the cause; it does not change who is responsible.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What kinds of statements create the most risk?
| Statement | Why it is risky |
|---|---|
| "You'll get a full refund." | States an outcome the agent may not have authority to grant |
| "We'll waive that fee." | A concession the customer will rely on |
| "Your claim is covered." | An interpretation of a contract by something not authorised to interpret it |
| "This price is locked until Friday." | A commercial term the business may not have offered |
| "We'll call you back tomorrow." | A commitment that fails if no task was created |
| "You're entitled to compensation." | A legal conclusion |
Each of these is safe only when the agent is reporting something the business has already decided and recorded. If the refund was issued, saying so is fine. If it was only requested, saying "you'll get it" is a promise.
What may the agent say, and when must it hold?
The agent may state facts from systems of record, describe published terms and service standards, and confirm actions it has completed within its limits. It must hold when a customer asks for a commitment the agent cannot back with a recorded decision, and it must refuse to interpret contracts or rights beyond approved wording.
A useful test is whether the business would be comfortable reading the sentence aloud in a complaint hearing. "Your refund of 60 was issued today" passes. "Don't worry, you'll definitely get your money back" does not. A ready list of statements to check is in the support agent promise checklist.
An illustrative example
Take a regional airline's chat agent. A customer whose flight was cancelled asks whether they will be compensated.
A risky reply: "Yes, you're entitled to compensation for the cancellation, and we'll pay it within a week."
A safe reply: "Your flight was cancelled and I've rebooked you on the next available one. Compensation depends on the reason for the cancellation and the rules that apply to your route. I've opened a claim for you, and the claims team will confirm whether compensation is due and how much within their published timescale."
The second reply states facts, records an action and points to a decision someone else will make. Nothing in it can be relied on as a promise the airline did not intend.
Which controls reduce the risk?
- A list of permitted commitments. The agent may only commit to what systems have done or published standards guarantee.
- Approved wording for entitlements. Rights and coverage are described in reviewed language, never improvised.
- A reply check. Commitment language is detected and checked against the record before a reply goes out.
- Record of what was said. Every reply is filed, so if a customer relies on a statement, you can see exactly what it was.
- A fast route for mistakes. When the agent does promise something wrongly, a person decides quickly how to put it right for the customer.
Even simple account changes carry implied promises, such as confirming that letters will now go to a new address; see whether an AI agent can change a customer's address. The broader case for judging agents by restraint is made in judge a support agent by what it will not promise.
Where Pryme fits
In Pryme, the agent can only describe an action as done when the system's confirmation arrives as a sealed fact, and commitments it cannot back are refused at the moment of action. Each reply and verdict is filed in an append-only record. More is in the AI support agent that knows when to say no.
Questions people ask
Does telling customers they are talking to an AI change liability?
Disclosure is required in many cases and builds trust, but it does not turn the agent's statements into something the customer should ignore. Treat them as the business's statements.
Should we honour a mistaken promise?
That is a decision for a person, weighing fairness, cost and precedent. Many businesses honour small ones and fix the cause.
Can terms and conditions exclude responsibility for the agent?
Possibly in part, depending on the law, but relying on that is weak protection. Preventing the promise is far more effective.