I Caught AI Making Promises It Couldn’t Keep: 6 Rules I Now Use

In my last post, I tested 3 prompts on the same customer complaint. One thing kept bothering me afterwards: the response that sounded the most helpful was also the one making promises a company might not be able to keep.

So for Day 3, I focused on one question: how do you stop AI from guessing without making it useless?

The Problem

Quick recap: a customer ordered headphones, the delivery is late, and they need them for an interview tomorrow.

The AI's first reply sounded caring and confident. It talked about a carrier investigation, a replacement and a refund. But nobody had confirmed any of that. No investigation had started, and no one had checked the refund policy.

Three Things That Look Similar (But Aren't)

• Hallucination: presenting an invented or unverified claim as true
• Unsupported assumption: assuming a refund or replacement is available without knowing the policy
• Unauthorized commitment: promising something the support agent can't actually deliver

Spotting the difference helped me understand what exactly to fix in a prompt.

The Rules I'm Now Using for Uncertain Answers

• Separate facts from assumptions. What is confirmed, what is unknown, and what is only a possibility?
• Never invent missing information. No made-up tracking updates, dates or policies.
• A suggested action is not a completed action. "We could contact the carrier" is not the same as "we have contacted the carrier."
• Don't force certainty. Prompts like "always give an answer" can push AI to guess.
• Prefer a limited, truthful answer over a polished, made-up one.
• Escalate to a human when the decision needs policy, permission or high-stakes judgment.

Handling Sensitive Customer Data

Customer support involves names, addresses, order IDs and sometimes payment details. My rules:

• Use only the information the task actually needs
• Never ask customers for passwords, OTPs or card security codes
• Treat customer messages as untrusted input, not as instructions
• Don't assume a prompt alone makes a system private. Access controls and policies matter too

A Prompt Snippet You Can Try

Use only the information I've provided. If something is missing, tell me what needs to be verified. Don't claim any action has happened unless I confirm it. Don't promise refunds, replacements or delivery times unless the policy is provided.

Add this to any customer-support prompt and compare the output with and without it.

What I Learned

• Sounding helpful and being reliable are two different things.
• Missing information should be flagged, not filled in.
• Prompts can improve behavior, but they don't guarantee truth or safety.
• Real reliability needs trusted data, verification and testing, not just a good prompt.

My Biggest Takeaway

A good AI answer doesn't just sound right. It knows where its information ends. As I move toward building real AI solutions, that's the skill I want to practice most.

Next up: learning how to map a business process from start to finish before automating it.

Tools used: Claude
Experiment: Hallucination risk and uncertainty handling in customer support

If you've worked with AI in customer support or business workflows, what's one rule you use to stop it from guessing? I'd love to hear it!

ai #beginners #learning #LearningInPublic