How to Stop AI Agents From Making Things Up

Every buyer asks us the same question before they sign anything. “What happens when your AI just makes something up?” It’s the right question. AI hallucinations in customer service aren’t a theoretical risk — an invented price or a confirmed appointment that never existed is a real bill someone has to pay. We build and run AI voice agents for a living, so we’ll tell you exactly how we keep them from inventing facts.

The short version: a good agent answers only from your verified facts and says “let me check on that” for everything else. The long version is below, because the details are where it actually goes wrong.

What a hallucination actually is, in plain words

A large language model doesn’t look things up the way a database does. It predicts the next most likely word based on patterns it learned from a huge pile of text. Most of the time that’s useful. The problem is the model has no internal sense of “I don’t know.”

When it hits a gap, it fills the gap. Confidently. With grammar so clean it sounds exactly as sure as when it’s right. Ask an ungrounded model what your Saturday hours are and it won’t hesitate — it’ll guess “9 to 5” because that’s a common answer for businesses like yours, and it’ll say it like it read it off your door.

That’s a hallucination. A plausible answer that nobody verified.

Why it’s worse in customer service than anywhere else

If a model hallucinates in a brainstorm, you shrug and move on. On a customer call, the hallucination becomes a promise. The caller heard a price. The caller wrote down a time. The caller is now planning their day around something your business never agreed to.

Say you run an HVAC company and the agent quotes a flat $89 diagnostic fee it invented. Now you either eat the difference to keep the customer happy, or you argue with someone who has a recording of your own system promising $89. Both options cost you. One of them costs you the review.

This is exactly why “the call got answered” isn’t the bar. A wrong answer delivered smoothly is worse than a missed call, because a missed call doesn’t create a liability you don’t know about yet.

Grounding: the agent only knows what you told it

The fix isn’t a smarter model. It’s a tighter leash. We ground the agent in your verified facts — your real hours, your real prices, your service area, your policies, what you do and don’t handle. Those facts live in a source the agent has to pull from before it answers.

In plain words, it works like an open-book exam where the book is the only thing allowed. When a caller asks a question, the system retrieves the relevant facts from your verified set and hands them to the model with strict instructions: answer from these, and only these. Engineers call the retrieval part RAG. You can call it “the agent isn’t allowed to wing it.”

And when the answer isn’t in the book? The agent doesn’t guess. It says “I don’t have that in front of me — let me get someone who does” and routes the call. That sentence is the whole game. An agent that comfortably admits the limit beats one that bluffs every time.

No-go zones: things the agent never improvises

Grounding handles the gray areas. Some areas shouldn’t be gray at all. We hard-wire explicit no-go zones — categories where the agent is never allowed to generate an answer, period, no matter how confident it sounds.

• Bookings. The agent never confirms a time it hasn’t actually written into your scheduling system. If the calendar write fails, the caller hears “let me have someone confirm that,” not a fake confirmation.
• Pricing. Real numbers from your price list, or a handoff. No estimating, no “probably around.” A guessed quote is a quote you’re now on the hook for.
• Medical, legal, and financial advice. If you run a dental clinic or a law firm, the agent books the consult. It does not opine on your toothache or your case. Ever.
• Anything safety-related. A gas smell, a flood, a break-in — that’s a human, immediately.

The line we draw with clients: the agent is allowed to inform from verified facts and to schedule against a real system. It is not allowed to decide, diagnose, or invent. Everything past that line is a handoff.

Escalation isn’t a fallback — it’s the feature

People treat the human handoff like a confession of weakness, as if a good agent should handle everything alone. We think that’s backwards. Knowing when to escalate is the most valuable thing the agent does.

With Mercvox — our AI voice receptionist for trades businesses, live across North America — emergencies route straight to a human. That’s a design choice, not a limitation we’re apologizing for. The agent picks up in under two seconds, handles the routine 24/7, and the moment a call is outside what it can safely answer, it gets a person involved.

Good escalation has triggers. Low confidence on a retrieved answer. Certain keywords. A caller who’s clearly frustrated. A request that touches a no-go zone. When any of those fire, the agent stops trying to be clever and starts trying to be useful, which means getting out of the way.

Test for it before launch: red-team your own agent

You don’t find out whether grounding holds by hoping. You attack the thing on purpose before a customer does. We red-team every agent with the questions a real caller will eventually ask — the awkward ones, the edge cases, the traps.

• Ask about a service you don’t offer and see if it politely declines or confidently invents a price for it.
• Ask about hours on a holiday. Ask about a location you don’t serve.
• Try to get it to confirm a booking when the calendar is full or down.
• Push for medical or legal advice and check that it hands off instead of guessing.
• Ask the same thing five different ways, including the rude way, and make sure it doesn’t crack under rephrasing.

Every answer that should have been “let me check” but came back as a confident invention is a bug we fix before go-live. This is unglamorous and it’s most of the work. It’s also the difference between a demo and something you’d put on your main line.

Why “just use a better model” doesn’t fix it

Here’s the part that gets oversold, so we’ll be blunt. A more capable model hallucinates less often, not never. It still has no native sense of what’s true about your specific business, because your Saturday hours were never in its training data. They couldn’t have been.

Worse, a smarter model hallucinates more convincingly. The fluency goes up, so the wrong answers get harder to catch. If your whole reliability strategy is “we upgraded to the newest model,” you’ve made the failures rarer and prettier, not gone. The thing that actually moves the needle is the boring stuff: grounding, no-go zones, escalation triggers, and testing. The model is one ingredient. It isn’t the recipe.

What to ask a vendor before you trust them with your phone

If you’re evaluating anyone — including us — make them answer these.

• Where does the agent get its facts, and what happens when the answer isn’t there?
• What are the hard no-go zones, and how are they enforced — a polite instruction or an actual block?
• When and how does it escalate to a human? Show me the triggers.
• How did you test it before launch? Can I watch you try to break it?
• When it confirms a booking, is that confirmation tied to a real write in my scheduling software, or is it just saying words?

A vendor who handles those without flinching has thought about reliability. A vendor who waves the question away with “our model is really advanced” has told you which one matters more to them. We dig into this distinction more in our piece on custom AI agents versus off-the-shelf chatbots , and you can see how we think about this across industries on our services page.

FAQ

Can you make an AI agent that never hallucinates?

No, and anyone promising zero is selling you something. What you can do is build it so that when the agent doesn’t know, it says so and hands off instead of inventing. The goal isn’t a model that’s never wrong. It’s a system that never confidently states a fact it can’t back up.

Is hallucination the reason some people prefer a human answering service?

Sometimes, though a person can also misremember your hours or quote the wrong price. A well-grounded agent pulls from one verified source every time, which is more consistent than a tired human at 2am. We compare the trade-offs honestly in AI receptionist versus answering service .

How long does it take to set up grounding for my business?

Most of the time goes into gathering your verified facts — hours, prices, policies, service area — and red-teaming the result. Once those are clean, the agent has a tight source to work from. The facts are the slow part, not the technology.

If you want an agent you can actually trust on your main line, that starts with how it handles the things it doesn’t know. Book a free intro call and we’ll walk you through how we’d ground one for your business and show you exactly where it hands off to a human.