Why AI Chatbots Quote the Wrong Price (And How to Fix It)
OS
Oskar
·6 min read
A chatbot that quotes the wrong price is not a bug. It's what happens when you ask a language model a factual question and it has no fact to retrieve. The model doesn't know your prices unless something hands it your prices, so it does the only thing it can do: it predicts a number that sounds plausible.
That distinction matters more than most vendors let on. A generic AI chatbot built on a general-purpose model has read millions of pages about pricing in general. It hasn't read your Preisliste from last month. When a visitor asks what a Heizungswartung costs, the model answers anyway, because refusing to answer isn't something base language models are naturally good at.
Frequently asked questions
Related Posts
This is a solved problem, but only if you build the chatbot the right way. Below is why the wrong-price failure happens, what it actually costs a small business, and the specific fixes that stop it.
Why AI chatbots make up prices in the first place
Large language models generate text one likely word at a time. They aren't databases and they don't look up facts unless you connect them to a source of facts, a process called retrieval-augmented generation, or RAG. Without that connection, a model asked about your prices does the same thing it would do if asked about a fictional restaurant's menu: it produces something that reads correctly, in the right format, with a currency symbol and a number, and none of it is grounded in anything you actually charge.
This is why a chatbot can sound completely confident while being completely wrong. There's no hedge in its tone, no "I'm not sure," unless you explicitly train it to say that. It will quote 89 euros for a service call as easily as it would quote 890, because both numbers are equally plausible to a system that is pattern-matching language, not consulting your books.
What a wrong quote actually costs you
The damage isn't hypothetical. In February 2024, a Canadian tribunal ruled that Air Canada had to honor a discount its support chatbot invented for a grieving customer, on the basis that the airline was responsible for information its own chatbot provided. The case, Moffatt v. Air Canada, 2024 BCCRT 149, is now the reference point lawyers cite when they explain that "the bot said so" is not a legal defense.
For a Handwerk business, the version of this problem is smaller in scale but just as real. A chatbot that tells a customer "Heizungswartung ab 89€" when your actual starting price is 149€ creates one of three outcomes: you eat the difference to avoid a dispute, you argue with a customer who has a screenshot, or you lose the lead once they find the real number and feel misled. None of those outcomes are cheap for a business that runs on referrals and local reputation.
How to stop an AI chatbot from guessing your prices
The fix isn't "better prompting." A good prompt can reduce how often a model guesses, but it can't give the model information it was never given. Fixing the wrong-price problem means changing what the chatbot has access to, and what it's allowed to say when it doesn't know something.
Ground the bot in your real price list, not a description of it. Upload the actual Preisliste, Leistungsverzeichnis, or service-page pricing tables as source documents, not a paragraph in a system prompt that summarizes them from memory.
Break pricing into small, specific chunks. A single 40-page PDF with one price buried on page 22 retrieves worse than a price list split by service line, because retrieval works by matching the customer's question to the closest chunk of text, and closer chunks mean more accurate matches.
Tell the model exactly what to do when it finds no match. The instruction should be a plain rule: if the exact service or price isn't in the retrieved documents, say so and offer a human follow-up, instead of estimating.
Set an escalation threshold. For anything above a set amount, say 300€, or anything involving a custom quote such as Altbausanierung or non-standard electrical work, route the conversation to a callback request instead of letting the bot commit to a number.
Re-sync your price data every time it changes. A chatbot trained once in January on prices that changed in March will confidently quote January's numbers. Retraining needs to be a recurring step, not a launch-day task.
Test it against your own price list before customers do. Ask the bot the ten pricing questions your team gets asked most often and check every answer against the actual sheet. This catches gaps before a customer does.
A concrete example: training a chatbot on a Handwerk price sheet
Say you run an SHK (Sanitär-Heizung-Klima) business with a standard price sheet: Heizungswartung, Rohrreinigung, Notdienst-Anfahrt, and a handful of flat-rate services, plus a note that custom installations are quoted on-site. Uploaded as source documents into a RAG chatbot like AmueAI, that sheet becomes searchable, so when a visitor asks about maintenance pricing, the bot retrieves the exact line, quotes the exact number, and can point to where that price comes from.
If someone asks about a bathroom renovation with non-standard plumbing, the bot has no matching chunk to retrieve, and a correctly configured one says so. It explains that custom jobs are quoted after a site visit and offers to book one, instead of inventing a number. That's the difference between a chatbot that protects your margins and one that quietly erodes them.
Rule-based bot vs. RAG chatbot, for price questions specifically
Not every chatbot on the market works the same way, and the difference matters most exactly where pricing is concerned.
Rule-based / decision-tree bot: Only ever says what a human explicitly scripted into a flow. Safe from hallucination, but breaks the moment a customer asks anything outside the tree, and every price change means editing the flow by hand.
RAG chatbot without escalation rules: Reads your documents but still guesses when a question falls outside them, unless it was explicitly told to say it doesn't know. This is the setup that produces confident wrong answers.
RAG chatbot with grounding and escalation (what AmueAI is built to do): Retrieves the exact price from your uploaded documents, and hands the conversation to a human the moment a question falls outside what's in the data.
Pull up your own price list right now and imagine a stranger asking your website chatbot about the top five line items on it. If you're not certain it would answer all five correctly, that isn't a future risk, it's the exact failure mode customers screenshot and post to a local Facebook group. Fixing it isn't complicated: give the bot your actual documents, tell it when to stay quiet, and check its answers the way you'd check a new employee's.
AI Agents
Rule-Based vs RAG Chatbot: How to Actually Decide
Most rule-based vs RAG chatbot guides skip cost and setup time entirely. Here's a concrete test using your last 100 support questions, what each option actually costs a small business, and when a RAG chatbot like AmueAI beats a decision tree.
