Will the AI Say the Wrong Thing? How Grounded Knowledge Bases Keep Sales Messaging Compliant

If you’re thinking about adding AI automation to your lead generation workflow, you’ve probably asked this question at least once:

“Will the AI say the wrong thing to my leads”

Our own account executives hear this often on sales calls.

One prospect told us they broke a competitor’s AI agent in about five minutes, even though it ran on a 40,000-character knowledge base. Another had used a tool that kept asking a lead “what time today?” after the lead had already said tomorrow. A lender asked us how we would make sure the agent never implied a guaranteed rate.

Meera runs lead conversations for B2C sales teams, and we’ve launched campaigns generating millions of text interactions across insurance, financial services, and education. Here is what that work has taught us about why AI sales agents go off script, and how grounded knowledge bases, scripted answers, and fallback rules keep them compliant.

Key takeaways

• AI sales agents go wrong in three ways: they invent answers, ignore what the lead said, or drift into regulated language.
• A grounded agent answers only from content your team has approved.
• A bigger knowledge base raises the risk of wrong answers, so keep yours small, current, and approved.
• Sensitive topics should trigger scripted replies, and questions without an approved answer should go to a person.

Three ways AI sales agents say the wrong thing

Each failure has a different cause, so it helps to pull them apart.

1. It invents an answer. Language models are built to produce fluent text. When a loosely built agent can’t find an answer in its approved content, it fills the gap with whatever sounds right. That is how a five-minute test breaks a 40,000-character knowledge base: ask something the documents don’t cover, and see what comes back.
   
2. It ignores what the lead said. The “what time today?” loop comes from conversation design. The agent missed the lead’s intent and kept following its script. More knowledge won’t fix that. The agent has to recognize intent, such as a request for a call tomorrow, and act on it.
3. It drifts into regulated language. This is the one compliance teams lose sleep over. An agent that implies a guaranteed rate, promises coverage, or confirms eligibility creates exposure for the business. The CFPB flagged this risk in its chatbot issue spotlight, warning that inaccurate information about a financial product can seriously harm consumers.

And the business owns the mistake. In the Air Canada ruling, a British Columbia tribunal held the airline liable for bad bereavement fare advice from its website chatbot. The tribunal rejected Air Canada’s argument that the chatbot was responsible for its own actions.

What a grounded knowledge base does

Think of an open-book exam where the book is the only allowed source. The technique is called retrieval-augmented generation, or RAG. When a lead asks a question, the system first searches your approved content for the relevant passage. The model then writes its reply from that passage, instead of from everything it absorbed during training.

Meera works this way. It answers leads using your approved content and language, drawn from your knowledge base and FAQs. Meera does not run as an open-ended chatbot. Conversations can be fully scripted, built on controlled decision paths, or grounded in your knowledge base, with limits on which intents the agent handles and when it escalates.

A good knowledge base usually includes:

• Your website content
• Current FAQs
• Product, pricing, or program documents your compliance team has signed off on
• Approved answers to common objections
  

For more on how two-way texting works under the hood, see our guide to AI SMS chatbots.

Why 10,000 hours of call recordings makes things worse

When teams want an agent to sound like their best reps, the instinct is to load everything: every script, every training deck, sometimes thousands of hours of call recordings. More context feels like it should mean better answers. In practice, more content makes wrong answers more likely.

Call recordings are full of things you would never want an agent to repeat. Reps quote rates that expired months ago and improvise promises to close a deal. Leads share personal details. When retrieval pulls from that pile, it can surface two passages that contradict each other, and the model blends them into one confident, wrong answer.

Research points the same way. One Stanford study found that model accuracy often drops sharply when the relevant information sits in the middle of a long input, even for models designed for long inputs. A bigger pile gives the model more places to lose the right answer.

A small, current, approved knowledge base works better. Before you add a document, ask one question: would our compliance team sign off on the agent saying this word for word? If the answer is no, leave it out. Old rate sheets, expired promotions, raw transcripts, and internal notes all fail that test. Recordings still have a use, though. Mine them for the questions leads ask most, then write approved answers to those questions.

The three guardrail layers Meera uses

Grounding handles most questions. Regulated teams need two more layers on top, plus controls on the platform itself.

1. Grounded answers for everyday questions. Questions about process, timelines, availability, and next steps get answered from your approved knowledge base. Meera’s team builds that knowledge base from your website and internal documents before launch.

2. Scripted answers for sensitive topics. Some topics are too risky for a generated reply, however well grounded. For those, your team writes the answer. When a lead raises that topic, Meera sends the approved wording instead of composing its own. Teams often script objection handling and competitor comparisons, and lenders will want rate and eligibility questions on that list. In insurance, Meera never presents itself as a licensed agent. When a question needs one, Meera routes the lead to a licensed agent on your team.

3. A handoff when there’s no approved answer. If Meera can’t find the answer in your approved content, it doesn’t guess. It tells the lead it will connect them with someone on your team. On voice calls, questions outside the agent’s scope transfer to a human agent automatically.

The controls underneath. Meera’s compliance controls cover the rules around every message: quiet hours based on the recipient’s time zone, DNC Registry scrubbing, consent tracking, opt-in and opt-out management, and conversation logs for audits. Meera also registers your brand and campaigns with carriers, and it is SOC 2 Type II certified.

Before anything goes live, a dedicated AI expert builds your conversation flows and trains your team. You can also test the agent yourself first. Text it the questions that worry you most and read exactly what it sends back.

FAQ

What is a grounded knowledge base?

A grounded knowledge base is a set of approved documents an AI agent must answer from. The agent retrieves the relevant passage first, then writes its reply from that passage.

Can an AI sales agent discuss interest rates?

An agent can share rate information your compliance team has approved. Most lenders are safer scripting rate questions or handing them to a loan specialist. Check the rules for your products with your legal team.

Who is liable when an AI agent says something wrong?

Generally, the business that deploys it. In the Air Canada case, the tribunal held the airline responsible for its chatbot’s advice. This isn’t legal advice, so confirm your obligations with counsel.

Should you load call recordings into an AI agent’s knowledge base?

Usually not. Recordings contain outdated information, off-script promises, and personal data. Use them to find the questions leads ask, then write approved answers.