Who is liable when your AI receptionist is wrong?

If your AI receptionist tells a caller the wrong price, the wrong hours or a promise nobody approved, the business that deployed it is the one the caller will blame, and a court or regulator is likely to agree. Your client carries most of the exposure. You carry some, because you built and sold the system. The AI vendor underneath usually carries the least. We are not lawyers and this is not legal advice, but the pattern is consistent enough that you should plan for it before your first client goes live.

Who is responsible when an AI receptionist gets it wrong?

The business whose name is on the phone line. Callers do not know or care which model, platform or agency sits behind the voice. They hear "Thanks for calling Northside Heating" and treat what follows as Northside Heating speaking.

The clearest public example is a chatbot case, not a voice one. In Moffatt v. Air Canada, a customer relied on the airline's website chatbot for bereavement fare rules, and the chatbot's answer contradicted another page on the same site. Air Canada argued the chatbot was a separate entity responsible for its own actions. The tribunal rejected that and held the airline responsible for information on its website whether it comes from a static page or a chatbot, and found the airline had not taken reasonable care to make sure the chatbot was accurate.

That is one small-claims-style decision in one jurisdiction. It does not bind a court in your state. But it tells you what the argument "the bot did it" is worth when a customer has a screenshot.

Regulators say the same thing. The FTC's line, quoted in a Morgan Lewis analysis of AI booking agents, is that there is "no AI exemption from the laws on the books." Materially wrong details about price or cancellation terms can be misleading even if no human typed them.

What does that mean for the agency that sold it?

It means your client is first in line, but you are second. If the AI says something wrong because you loaded the wrong price band, skipped a service the client does not offer, or never tested the after-hours script, the client will look at you. Whether they have a legal claim depends on your contract and your state. Whether they have a business reason to fire you does not depend on anything.

Be honest with yourself about where errors come from. Wrong answers tend to trace back to a handful of causes.

• The business gave the agent a price, a service list or a policy that was out of date or ambiguous.
• Nobody wrote down what the agent must never promise, such as a firm quote, a diagnosis or a guaranteed arrival window.
• The calendar or CRM connection failed and the agent confirmed a booking that did not exist.
• The model improvised something plausible to fill a gap in its instructions.

The first two are intake problems, and you control intake. That is why we built the onboarding path around collecting hours, services, price bands, calendar and lead routing once, in the client's own words, before the build starts. Our intake checklist post walks through what to ask for.

What about the AI vendor or the platform?

Assume they will not cover you. Platform terms commonly disclaim accuracy, put the review burden on the customer and cap liability at what you paid. The Morgan Lewis piece describes consumer terms in travel this way, noting that they often disclaim accuracy and cap liability at the amount paid. Those clauses may limit what you can recover. They do not decide who fields the angry phone call.

We are not going to tell you what any particular contract says, including ours. Read the terms of every platform in your stack, including Lumina Sales, and have a lawyer review the agreement you give your own clients. That is the document that matters most, and it is the one you control.

What should your client agreement say?

You want a short list of clauses, written by someone licensed in your state. The topics worth raising with them are these.

Scope of authority. State what the agent may do, such as answer questions, take messages and request appointments, and what it may not do, such as quote final prices, give professional advice or cancel jobs.

Client-supplied information. Say that the agent answers from information the client provided and approved, and that the client is responsible for keeping it current. This one clause does a lot of work, because most errors are stale data.

Review and sign-off. Have the client listen to test calls and approve the script and price bands in writing before launch. A signed approval is evidence that the client saw what the agent would say.

Change requests. Say how the client updates prices or hours and how fast you will apply the change. A client who changed their pricing in September and told nobody is not an AI failure.

Limits on your own liability, plus whether you carry insurance that covers this kind of work. Ask an insurance broker about technology errors and omissions coverage before you need it.

How do you cut the odds of a wrong answer in the first place?

Contracts allocate blame after the fact. Prevention is cheaper.

Test with the ugly questions, as we cover in testing an AI receptionist before it goes live. Ask the agent about a service the business does not offer, a price you know is not in the script, and a request that needs a human. Write down what it says.

Give the agent a refusal list. Tell it plainly to say "I can't quote that, but I can have someone call you back" for firm prices, medical or legal advice, and anything involving money changing hands. A short, specific refusal list beats a long, vague tone guide.

Separate estimates from commitments. Morgan Lewis recommends keeping estimates clearly apart from confirmed information, and says the interface and the disclaimer should tell the same story. In practice, the agent says "ranges usually run from X to Y, and the owner confirms the final number" instead of "that will cost X."

Keep records. Call recordings, transcripts and booking logs are how you find out what the agent actually said, and they protect everyone if there is a dispute. Check your state's recording consent rules and how long your client wants records kept.

Give callers a human exit. If the caller asks for a person, or the agent cannot resolve something, it should take a message or hand off. Our post on disclosure rules for voice agents covers the related question of telling callers it is an AI.

Own the maintenance. Prices change. Seasons change. Somebody has to review the script on a schedule, and our post on who maintains the system after it goes live is worth a read, and it is a natural paid retainer for you.

Where does this advice fall short?

Three honest limits.

First, no process gets the error rate to zero. Language models can produce a fluent wrong answer even with good instructions. The goal is a low rate on low-stakes questions and a handoff on high-stakes ones, not perfection.

Second, some trades need extra care. A law firm intake line, a medical practice or anything near financial advice carries regulatory risk that a generic script does not cover. If you sell into those, get a specialist to review what the agent may say before you go live. The FTC has also pursued companies for overstating what their AI can do. In one 2024 action, DoNotPay agreed to pay $193,000 over claims about an untested AI lawyer, so do not oversell your own system either.

Third, we cannot give you legal advice and neither can any blog post. Laws differ by state and country, and the case law on AI errors is young.

What does a prospect see in our demo about this?

Our demos run against a sandbox, and the page states that the AI is a demo and that bookings made in it are simulated. That disclosure exists for the same reason this post does. A prospect should never think a demo booking is real, and you should never have to explain why it was not. You can try it yourself in the live demo and see exactly what a prospect sees.

FAQ

Can I make the client fully responsible for AI mistakes?

You can write the contract to put most responsibility on the client for information they supplied and approved, and a lawyer can help you draft it. A clause does not stop an unhappy client from leaving or a customer from complaining, and it may not hold up if you misrepresented what the system does. Treat it as a floor, not a shield.

Do I need a disclaimer on the AI receptionist's greeting?

Possibly. Some jurisdictions and industries have disclosure rules, and a clear statement that the caller is speaking with an AI assistant helps set expectations. Whether a disclaimer protects you depends on the facts. The Morgan Lewis article warns that the experience and the disclaimer need to be consistent with each other.

What happens if the AI books an appointment that does not exist?

That is the worst case for client trust, and it usually comes from a failed calendar connection that the agent reported as success. Test the failure path during setup, make the agent confirm only after the calendar returns a success, and review booking logs against the real calendar in the first weeks.

Should I carry insurance?

Ask a broker. Technology errors and omissions policies exist for software and consulting work, but whether one covers AI output errors varies by insurer, so ask about that exact scenario and get the answer in writing.

Does Lumina Sales cover me if the AI gets something wrong?

Read the terms you agree to when you sign up and ask us directly if something is unclear. We would rather you ask before a client launches than after. The platform gives you the demos, portal, fulfillment and onboarding path, and the agreement with your client is yours to write.

Next step

Before your next client signs, write a one-page list of what the agent may promise and what it must hand off, and put it in front of the client for a signature. Then run it through test calls. If you are still deciding on a platform, the pricing section lists exactly what the $49 plan includes, and you can start your workspace here.