Do AI Receptionists Get Things Wrong? Hallucinations, Liability, and How to Avoid Them

Yes, an AI receptionist can get things wrong, and this is the failure that should worry a business owner most. A robotic voice is annoying. A confident wrong answer is damaging, and as of 2024 it's a documented legal liability: an AI that invents a price you don't charge, quotes an opening hour you don't keep, or promises a service you don't offer can cost you a customer, a refund, or worse, before you even know the call happened.

The good news is this is preventable, and it comes down to how the agent is built, not whether AI "can be trusted" in some vague sense. Here's why AI agents make things up, why it's your problem and not the AI's, and what a well-built one does to stop it.

The case that put a number on the risk

In February 2024, a tribunal ordered Air Canada to honour a bereavement-fare refund policy that didn't exist, one its own website chatbot had invented and told a grieving customer. Air Canada argued it couldn't be held responsible for what its chatbot said, suggesting the bot was "responsible for its own actions." The tribunal called that a "remarkable submission" and ruled the airline was liable for information on its own website, whether it came from a static page or a chatbot (The Guardian).

The principle is simple and it travels: you own what your AI tells your customers.

In New Zealand, that lands under the Fair Trading Act 1986. The government's own Responsible AI guidance for businesses lists the Act explicitly, warning that misleading or deceptive conduct arising from AI outputs can breach it, with substantial penalties. A made-up price or promise from your AI receptionist isn't a quirky bug. It's a representation your business made.

Why AI agents make things up

The technical name is hallucination. A language model is built to produce a plausible-sounding response, and if it doesn't have the real answer, it can still generate one that sounds confident. On a phone call, where there's pressure to fill silence and keep moving, a poorly-configured agent reaches for a guess rather than admitting it doesn't know. The caller, hearing a confident voice, believes it.

The root cause is almost always two things: no guardrails, and bad source data. Gartner's 2026 research found 38% of AI project failures trace directly to poor data quality. An agent pointed at a stale price list or a half-finished FAQ will confidently repeat whatever it was given. Garbage in, liability out.

The fix: an agent that knows its limits

A well-built agent is set up to do the opposite of guess. When it hits a question it doesn't have a real answer to, it should say so plainly and take a message. "I'm not certain on that, I'll get Julian to confirm and call you back" is a far better outcome than a made-up figure, because it protects both the caller and your reputation.

Two things prevent hallucinations in practice:

• Working from your real, current details, not assumptions. The agent answers from what you've actually told it about your business, and nothing more.
• A clean fallback when it doesn't know. Instead of bluffing, it captures the question and hands it to you. You give the authoritative answer when you call back.

How AppAI handles it

AppAI is set up to capture and hand off rather than guess. If a caller asks something the agent isn't sure of, it takes the details and texts you a summary so you can ring back with the right answer, instead of putting words in your mouth. You stay the source of truth on anything that matters, which is exactly what keeps you on the right side of the Fair Trading Act.

It's also kept current. If there's something the agent should know, a new price, a change in what you offer, the wording can be updated the same day. So the fix for "it didn't know X" is simply to tell it X, and from then on it does. That also closes the data-quality gap that drives so many AI failures.

This is the same honest design that runs through the whole product: the agent answers, qualifies, and captures, then hands the decision to you. For the bigger picture of where AI agents fall short and how to judge one, see why AI phone answering services suck and how to pick one that doesn't, and for what the agent does well, see what an AI receptionist actually is.

How to test it before you trust it

Don't assume, check. Ring the agent and ask it something it shouldn't confidently know: an unusual request, an edge-case price, a service that's a bit outside the obvious. A good agent will admit the limit and offer to take a message. A bad one will make something up. That single test tells you more about whether you can trust it than any feature list, and it's far cheaper than finding out via a customer complaint.

You can run that test on AppAI right now: ring the live demo line on 03 565 9950 and try to catch it out. It's the real product answering in a natural Kiwi voice.

If it earns your trust, setup is done for you and usually live in about five business days, it's $350 + GST a month with no lock-in, and you keep your number. Founding members (the first 10 Canterbury businesses) get $350/month locked in for life, details on the pricing page.