So, your AI made something up, and you didn’t catch it before it hit the field.

We hear this all the time from sales departments.

Artificial Intelligence is becoming more widespread among sales teams, which are already fast-paced and high-stakes environments. Of course, this means you run the risk of misinformation. 

In this case, the misinformation is a response hallucinated by an AI. You know, those answers that sound confident but aren’t real? In a customer-facting setting these outputs are a major trust breaker. 

That’s right. One false claim can set you back months of building trust.

In this post, we’re going to show you what causes AI hallucinations, how to catch them, and prevent bad responses from happening in the first place. 

It comes down to better habits, better prompts, and better tooling. 

Let’s dive in.

What is an AI Hallucination?

Just so it’s clear: an AI hallucination is the response you get from an AI system, or LLM, that sounds authentic but is not grounded in truth. 

In sales, timing is tight and trust is everything. To compete at a high pace today, your sales reps need to rely on automation and AI-powered workflows. That means information is processed at unprecedented speed – whether that be a response to a customer objection or a simple “how do I?” question about internal processes.

Sadly, the answers that AI gives you can sound real, when in fact they couldn’t be further from the truth.

Hallucinations are not always obvious, and they can sound right because AI states these answers with complete confidence. 

Here’s what they might look like: 

• “Our competitor doesn’t support this feature” (when they actually do)
  Your AI might accuse your competitors of lacking capabilities or overstate differences. You never want to look uneducated with a customer but now you run the risk of being over-educated on the wrong information.
  
• Product claims that aren’t aligned with your messaging.
  This would be the inverted use case, where your AI lies about your own product. This is a fast way to lose credibility with your buyer.
  
• Misattributed quote to a real person or company.
  This happens often and can really hurt your reputation if used prominently.

Why do these responses happen?

Simple: Large language models (LLMs) aren’t fact-checkers. When generating a response, they don’t scour the internet for reputable links to check their work. They’re not grounded in truth or verifying their responses against authoritative sources. Instead, they predict the best possible next words based on a combination of training data sources such as the World Wide Web.

On its own, an AI doesn’t “care” about accuracy – even if it sounds like it does.

Without proper grounding and instructions on what is an authoritative source of truth, LLMs will fill gaps in their knowledge with plausible-sounding guesses. And they’ll present the information as if it’s a verifiable fact. 

If you’re in a hurry to get a response over to your client, you may be convinced enough to copy and paste.

Big mistake.

5 Ways to Spot Hallucinations Quickly

Now, how do you catch a hallucination before it goes out? The key to prevention is to always be a bit skeptical and to recognize patterns in what drives hallucinatory responses. 

Here are the five warning signs to watch for when it comes to AI making stuff up: 

1. “Too Good to Be True” Claims

You’ve heard the expression many times: if it sounds too good to be true, it probably is. This is nowhere more valid than with LLMs. If the response sounds overly polished or flawless, take a beat. 

You might see claims such as, “Our product guarantees 100% uptime and infinite scalability.” 

Any extreme or perfect response like this one is almost never going to be true. The hard part is looking for the more subtle, confident statements that aren’t obviously false.

Recognize: AI tends to over-embellish to sound helpful. Ask it to explain how it arrived at an answer and to reveal its sources to boost credibility.

2. Missing or Vague Sources

Hallucinations won’t have a clear source to back up the claim. After all, if there was a great source link, it likely wouldn’t be a fake response.

Watch out for “source drift.” This is when the AI will reference a document or link, but the claim doesn’t actually appear in the source text. 

Recognize: If the AI won’t cite the specific source text from a reputable link, take caution.

3. Unfamiliar or Inaccurate Phrasing

You know your company, you know its lingo, and you can pick up on language that doesn’t align. Personal expertise is the most common way teams discover hallucinations. If the LLM uses a term that doesn’t sound familiar to you, like “hyper-automated ingestion pipeline,” and such verbiage doesn’t appear anywhere on your website or internal docs, it’s a red flag. 

Recognize: AI may borrow language from another vendor in your space. Be specific about which entity you’re referring to (company names, people, etc) to reduce risk.

4. Mismatch Between Input and Output

Sometimes, you’ll upload a document for your LLM to use, but you start noticing that it’s going outside the bounds of the doc itself. Without clear rules and blocks, the AI will look to its broader model knowledge to answer your queries. This is true even if you provided specific documentation. 

Recognize: If the AI goes off-topic from the source doc, it’s more likely to be guessing. 

5. Question Repetition

This one is perhaps the easiest to catch: if the AI repeats back to you word for word what you asked it, beware. For example: 

Q: “Does the vendor support a follow-the-sun support model with 99.9% uptime?”

A: “Yes, the vendor supports a follow-the-sun support model with 99.9% uptime.”

Recognize: If AI echoes you, it’s not answering you, it’s most likely hallucinating. 

Pro Tip: Train your sales reps to always double-check responses against your company knowledge base, CRM, or content repository. If they can’t find a match, they should escalate before using the AI answer. 

How to Fix and Prevent LLM Hallucinations

The good news is, you don’t have to just accept AI-generated lies. With the right skills, your team can reduce them and even eliminate them (at least from making their way to your customers) entirely. 

Here’s what to do: 

Ground Your AI with Source Documents You Trust (RAG)

This process is referred to as Retrieval-Augmented Generation (RAG). It is a fancy way of saying, “Don’t let the AI guess.” 

RAG enables you to feed documents that provide a source of truth to your model. You ground the AI with your internal company information, such as sales collateral, PDFs, customer data, and product specs.

There are numerous sources you can use.

That way, you can narrow the scope of its knowledge and output answers based on your approved sources of truth. 

You can build your own RAG system, or you can use an AI knowledge base specifically designed to unify, centralize, and pull answers from your company documents.

Always Ask AI to “Show Your Work“

You’ve been hearing that you need to show your work since the second grade.

Well, now it’s the robot’s turn.

Your entire team should get into the habit of asking your LLM to “cite the source.” You can also ask it to “show the document this answer came from.” 

If the AI can’t give you the source that backs what it’s saying, it’s not a trustworthy answer. 

Most LLMs will not automatically show where the information was derived from. You’ll want to add this rule to your prompts or use a tool like 1up which automatically demonstrates how sources influence your answer.

Tighten Your Prompts

Vague questions will breed vague answers. The more specific and concise your query, the more likely you are to receive an accurate, specific, and clear answer. 

Frame your questions with prompts like: 

“Based only on this document…”

or

“Don’t include information not found in the provided URL…”

This way, you’ll be limiting the model’s imagination and thereby boosting its accuracy. 

Allow your AI the Option to “Do Nothing”

A Large Language Model is trained on generating text. It WANTS to provide an answer. It doesn’t know it can say “I don’t know” without being penalized.

You’ll want to give your AI the option to say it cannot respond, or simply do nothing.

Use Platform Feedback Tools

Most AI tools will let you upvote and downvote the answers they deliver. Use this to your advantage. 

Feedback loops like this will teach the system which responses are the most useful. And which ones are just plain wrong? 

The improve answer feature with 1up allows for real-time corrections, so the AI won’t make the same mistake next time.

Training Your Team to Be AI-Literate

Okay, so you know you can train AI with upvotes and downvotes. You can also ground it and be specific. 

But you also need to train your team to be AI-literate. That is, they need to know how to work with AI to get the best results. 

Here’s how to train them to do just that: 

Create a Review Checklist

Encourage your team to be skeptical about AI outputs. When a sales rep is reviewing AI-generated content, train them to ask: 

• Does this align with our messaging? 

• Do I have clear, reputable sources to back up these claims? 

• Should I run this by a teammate for review before sharing with a customer?

When they have to check each of these answers off every time they review, they’ll get better and better at spotting, and preventing, AI hallucinations. 

Run Side-by-Side Audits 

Show your team real examples of AI hallucinations, so they can begin to develop instincts. Place those hallucinations next to grounded responses and have them point out the differences. 

The more they are exposed to AI falsehoods, the better they’ll get at picking up on those subtleties. 

Encourage the Use of Prompt Guardrails

It’s always good to standardize any process you have. In this case, you can standardize how your reps interact with AI by creating pre-vetted prompt templates. 

For example: 

• Safe: “Summarize only from this case study.”
• Risky: “What makes our product better than Competitor X?” 

Sharing a list of prompt templates (or mandating them, if you can) will significantly increase predictability in your teams’ AI outputs. When you design your prompts with guardrails in mind, you plan for safety.

Designate AI Champions

Train a few team members deeply so they can review outputs and educate others. And make sure to choose team members who will be engaged and enthusiastic. 

For example: 

With an AI tool your team is already using that hallucinates, bring it two or three of your strongest sales reps and have them run through the training. Work on scenarios, have them use the LLM, and show them how to identify and prevent AI hallucinations. Help them become masters at the game. 

Then, have those team members work with the rest of the team to show them how to do the same. 

When you get peer support and cheerleading, you’re more likely to get buy-in and actual learning. 

AI is Here to Stay, So Use it Wisely

In the end, AI should accelerate your team, not give you anxiety. It’s a tool and an assistant, but you still have to help it do its job. You can do this by learning how to train this tool. 

Hallucinations are fixable and preventable when you know how to spot them and plan for them. 

With the right habits, prompts, a review process, strong prompt hygiene, your team can gain trust in AI outputs.

FAQs on AI Hallucinations