What Happens if an AI Receptionist Gives My Customers Wrong Information?

Here's an illustration. A lawn care company in Spring gets a Saturday call from a homeowner asking if they trim oak trees. The company doesn't. But the AI's instructions never said so, and the knowledge base lists "landscaping services." The AI, trying to be helpful, says yes and books an estimate. On Monday, the owner reads the summary, winces, and makes an awkward callback.

Nobody lost a fortune. But it's exactly the kind of error owners worry about, and it's almost entirely preventable.

The Short Answer

AI receptionists make mistakes mostly when they're asked something their knowledge base doesn't cover and they haven't been told to say "I don't know." You prevent most errors by keeping the AI to approved information, listing what it must never promise, and routing unknowns to a person. You catch the rest by reviewing transcripts and fixing the cause.

What the Research Says About the Stakes

The reason to use an AI receptionist at all is that calls go unanswered. Small business call studies find roughly 40% to over 60% of inbound calls are missed, depending on the study and industry. A missed call is a certain loss. An answered call with a small error is usually a recoverable one.

That said, first impressions matter. Harvard Business Review reported in 2014 that acquiring a new customer costs 5 to 25 times more than keeping an existing one. You don't want a careless answer souring a relationship you worked hard to win.

There's also a paper trail to think about. Transcripts and recordings are how you catch errors, so handle them properly. Texas is a one-party consent state for call recording, but California and some other states require all-party consent. I recommend a recording disclosure on every call regardless. Healthcare offices also need a vendor that signs a Business Associate Agreement under HIPAA.

What a Wrong Answer Can Cost

Most AI errors are cheap: a wrong hour, a booking that needs to be moved. The expensive ones involve promises.

Run the math for a pricing error in your business:

• Approved quote the AI should give: ____ (say "starting at $89 for a service call")
• What a poorly limited AI might imply: ____ (say "a repair is about $89")
• Real repair cost: ____ (say $450)

Now your tech is at the door walking back a $361 gap with an annoyed customer. Do that twice a month and you'll have unhappy customers and a review problem. That's why limits on pricing language matter more than almost any other setting.

Why AI Receptionists Get Things Wrong

When I test AI receptionists, including our own line at FoundOnline, the errors almost never look like a robot going haywire. They look like a polite, confident answer to a question the knowledge base didn't cover.

The common causes:

• Gaps. The question wasn't covered, and the AI filled the space.
• Stale information. Last year's prices or a service you've dropped.
• Vague wording. "Landscaping services" covers a lot of ground.
• No permission to say "I don't know." The AI tries to help when it should hand off.

The surprising part is how often the fix is one sentence: if the answer isn't in the knowledge base, say someone will follow up, and don't guess.

Where Errors Cost the Most

Highest stakes: healthcare, legal, and financial services. A wrong answer can create liability. Keep the AI to scheduling and general information, with no advice of any kind.

Medium stakes: pricing in home services. Roofing, HVAC replacements, foundation work. A wrong number sets an expectation your crew has to undo.

Lower stakes: hours, service area, routine booking. Errors here are easy to spot and fix.

The higher the stakes, the more the AI should collect details and hand off rather than answer.

The Biggest Mistake: Perfect or Disaster Thinking

Owners tend to fall into two camps. Some assume the AI will be perfect, so they never read a transcript, and small errors repeat for months. Others assume it'll be a disaster, so they never try it and keep missing a big share of their calls.

What works is treating the AI like a new hire. Give it clear information, limit what it can promise, check its work closely for the first month, and correct it at the source. A new human receptionist would make mistakes too. You'd train them.

The Myth: "AI Hallucinates, So It Can't Be Trusted"

Vendor content implies the AI is always accurate. It isn't, and neither is any person on a phone.

Skeptic content says AI makes things up, so it can't be trusted with customers. That ignores the guardrails available today.

A grounded AI with a narrow scope and an "I don't know" path tends to make fewer costly errors than a rushed employee juggling three lines at lunchtime. And unlike that employee, every one of its mistakes shows up in a transcript you can read.

How I Set Up Quality Assurance

My rule is that prescription without diagnosis is malpractice. That applies to errors too: find the cause before changing anything.

1. Ground it. A short, current knowledge base, and instructions to answer only from it.
2. Set limits. A written list of what the AI must never say or promise.
3. Set the fallback. Transfer to the owner or take a detailed message when it doesn't know.
4. Test before launch. Call it and ask hard, odd, and off-topic questions.
5. Turn on transcripts and summaries, with a recording disclosure.
6. Review every call for the first two weeks, then a weekly sample.
7. Fix errors at the source in the knowledge base or instructions, then call and confirm the fix.
8. Make it right. If a caller got wrong information, call them back.

The Insider Tactic: Give the AI Permission Not to Know

Most errors come from an AI trying to be helpful without the facts. So write explicit "don't answer" rules:

• Don't quote prices beyond the approved list.
• Don't promise arrival times.
• Don't diagnose problems.
• Don't confirm services that aren't listed.

Then give it a graceful exit, something like: "That's a great question for the owner. I'll have him call you back shortly." Pair that with a text summary to you. Callers accept "let me get you the right person" far better than a wrong answer they act on.

One Thing to Do This Week

Write your "never say" list. Five to ten things your phone should never promise:

• Exact prices for jobs that need an inspection
• Same-day service you can't guarantee
• Services you don't offer
• Medical, legal, or financial advice
• Discounts that don't exist

Give it to whoever answers your phone today. If you add AI later, it goes straight into the instructions.

What Changed With AI?

Older phone bots rarely gave wrong information because they rarely gave any information. They routed you through menus.

Modern voice AI can hold a real conversation, which is what makes it useful and also what lets it be confidently wrong if it isn't grounded. The upside is that today's platforms let you tie the AI to a specific knowledge base, log full transcripts, and review every call.

You could never replay what your receptionist told a caller last Tuesday. Now you can, and each fix sticks.

How do I know if my AI receptionist said something wrong?

Review call transcripts and summaries. Read every call for the first two weeks, then sample weekly.

Can I stop the AI from quoting prices?

Yes. Limit it to approved pricing language, like a service call fee, and have it book an inspection for anything else.

What should the AI do when it doesn't know an answer?

Tell the caller someone will follow up, take their details, and send you a summary or transfer the call.

Does fixing an error once fix it for every future call?

When you correct the knowledge base or instructions, the change applies to every call from then on. Make a test call to confirm.

The Bottom Line

An AI receptionist will occasionally get something wrong, just like any person on the phone. The difference is that you can limit what it says, see every answer it gives, and fix mistakes once at the source.

If you'd like to test how an AI receptionist handles tricky questions, try the live demo at foundonline.pro and see what it does with a question it can't answer.