How to Stop an AI Phone Agent From Making Things Up

AI phone agents make things up when a caller asks something their facts don't cover and nothing tells them to stop. Prevent it with four controls: give the agent complete facts for common questions, instruct it to say the team will follow up on anything not covered, list topics it must never answer (prices not listed, availability, advice), and review transcripts weekly to find and close gaps.

Why it happens

Language models generate likely-sounding text. Ask one "Do you offer weekend appointments?" with no facts about weekends, and "Yes, we offer limited Saturday appointments" sounds plausible. The model isn't lying on purpose; it's filling a gap.

The fix isn't a smarter model. It's removing gaps and making "I'll have the team follow up" the expected answer when a gap remains.

Control 1: Close the obvious gaps

Most invented answers fall into predictable categories:

Category	Example invented answer	Fact to add
Hours exceptions	"We're open on Memorial Day"	Holiday closures
Services	"We do commercial work too"	Services not offered
Prices	"That's usually around $200"	Price list or "quoted after visit"
Policies	"Cancellations are free"	Cancellation policy
Payment	"We offer financing"	Payment methods; financing yes/no
Area	"We can come to Springfield"	Service area including exclusions
Timing	"Someone can be there today"	Rule: never promise times

Write "we don't" facts as carefully as "we do" facts. See AI agent business facts sheet.

Control 2: A firm defer rule

Put this near the top of the instructions:

"Only state facts from the business facts. If a question isn't covered, say you'll have the team follow up, and take a message. Never guess."

Pair it with a natural line the agent can use: "That's a good question. I don't want to give you wrong information, so I'll have the team follow up. What's the best number?"

Control 3: No-go topics

Some topics should never be answered by the AI even if it seems to know:

• Specific availability or arrival times
• Discounts, refunds and exceptions
• Medical, legal, tax or financial advice
• Anything about other customers
• Comparisons with competitors
• Promises about outcomes ("this will fix it")

List them under "Never" in the instructions. See writing AI phone agent instructions.

Control 4: Test for gaps

Before launch, ask questions you know aren't covered:

• "Do you offer financing?"
• "Are you open on [holiday]?"
• "Can someone come today?"
• "Do you do [service you don't offer]?"
• "How much for [unlisted job]?"
• "Is [competitor] cheaper?"
• "What's your warranty?"

Pass means the agent defers. Fail means you need a fact or a rule. See test an AI receptionist before launch.

Control 5: Review transcripts

After launch, read conversations weekly and search for:

• Numbers and prices: are they all in your facts?
• Words like "usually," "typically," "around": often a sign of a guess
• Promises: "will," "guaranteed," "today"
• Answers to questions you never covered

Each finding becomes a new fact or rule. See review AI agent conversations.

Subtle forms of making things up

Not every invention is a wrong fact. Watch for:

• Stretching a fact: applying "free estimates for replacements" to repairs
• Combining facts: "We're open Saturday, so we can come Saturday" (office hours vs service hours)
• Inventing process: "You'll get a confirmation email in five minutes" when no such email exists
• Over-reassurance: "Don't worry, that's an easy fix"

Fix with more precise facts and a rule: "Don't describe processes or timelines that aren't in the business facts."

Why it matters legally and commercially

What the agent says, your business says. A made-up discount, warranty or timeline can create disputes, bad reviews and, in some contexts, legal exposure. For telemarketing calls, federal rules prohibit misrepresenting material terms. In professional services, ethics guidance such as ABA Formal Opinion 512 stresses verifying AI output and supervising AI tools. The common thread: AI output needs the same oversight you'd give a new hire.

A quick-reference checklist

• Facts cover the top 20 questions, including "we don't"
• Defer rule at the top of instructions
• No-go topics listed
• Gap questions tested before launch
• Weekly transcript review with a search for prices, promises and hedges
• A named owner who updates facts

Example: fixing a real gap

A pet grooming business noticed its AI agent telling callers "we can usually fit you in the same week." Nothing in the facts said that. Reviewing transcripts, the owner found callers asking "how soon can I get in?" and the agent filling the silence.

The fix had three parts:

1. Fact added: "Appointments usually book one to two weeks out. The team confirms exact times."
2. Rule added: "Never say how soon an appointment is available. Collect preferred days and the team will confirm."
3. Test added: "How soon can I get my dog in?" became a permanent test scenario.

The next week's transcripts showed the agent collecting preferred days and saying the team would confirm. The gap was closed in fifteen minutes because someone read the calls.

Signs your agent is guessing

Read a sample of transcripts and look for:

• Specific numbers you don't recognize
• Confident answers to questions you never anticipated
• Callers saying "that's not what I was told" on follow-up calls
• Staff hearing about promises they didn't make

Any of these means a gap exists. Treat it as a missing fact, not a broken product, and close it the same day. Keep a running list of every gap you've closed; over a few months it becomes the most complete description of what your customers actually ask.

Making "I don't know" sound good

Owners sometimes resist defer rules because "I'll have the team follow up" sounds weak. Phrasing fixes most of that. Give the agent a few natural options:

• "Good question. I want to get that right, so I'll have someone from the team confirm it. What's the best number?"
• "That depends on a few details the team will want to check. Can I take your name and number?"
• "I don't have that information, but I can make sure the right person calls you back today."

Callers generally accept a confident handoff far more readily than a confident wrong answer they discover later. A deferral also creates a reason for your team to call, which is often where a sale or a booking happens.

Callata's approach

Callata designs its agents around a facts-only rule. The agent editor's business facts field states that the agent never states facts that aren't there, and every agent's built-in rules say to only state facts from that section, never invent prices, availability, policies or promises, and say the team will follow up when something isn't covered. Appointment requests are saved for your team to confirm, so the agent doesn't claim a slot it can't see.

Each AI call is summarized and listed under Recent AI conversations, so the weekly review is quick. AI minutes are $0.25 per minute; Callata Office is $99 per month for five users, with additional users at $20 each. Build a facts-only agent in Callata.