Trust, accuracy and compliance
AI chatbot hallucinations: keeping prices and stock accurate
A chatbot that quotes the wrong price or promises a policy you do not have is not a quirk. It is a statement made on your store. Here is how hallucinations happen and what actually prevents them.
What is an AI chatbot hallucination?
A hallucination is a confident answer that is false: an invented price, a discount code that does not exist, a return window the store never offered. It happens because language models predict likely text rather than look facts up, so a fluent wrong answer can look exactly like a right one.
The U.S. National Institute of Standards and Technology calls this "confabulation" in its Generative AI Profile (NIST AI 600-1), describing it as confidently stated but erroneous or false content. NIST notes that it is a natural result of how generative models work: they approximate the statistical patterns of their training data, and the risk grows when users act on content because it sounds confident.
In a store, the most damaging hallucinations are the ones that sound like commitments: figures and policies. A shopper who is told a sweater is $48 when it is $68, or that a sold-out size ships tomorrow, has been misled at the exact moment she was deciding to buy.
| Type | Example (invented) | Risk |
|---|---|---|
| Price | Quotes last season's price or a compare-at price as the sale price | Customer disputes, refunds, lost trust |
| Stock | Says a size is available when it sold out an hour ago | Cancelled orders, support tickets |
| Delivery | Promises a delivery date the carrier cannot meet | Complaints, chargebacks |
| Discount code | Invents a code that sounds plausible | Checkout failure, angry shopper |
| Policy | States a return window or free-shipping threshold the store never set | Pressure to honor it; possible liability |
| Product claim | Adds an ingredient, certification or material not on the label | Regulatory and safety exposure |
Can a business be held responsible for what its chatbot says?
In at least one public case, yes. In Moffatt v. Air Canada (2024 BCCRT 149), a British Columbia tribunal held the airline responsible for its chatbot's wrong answer about bereavement fares and rejected the argument that the bot was separate from the company. Treat your chatbot's answers as your own.
According to Manatt's summary of the decision, the chatbot told a grieving customer he could book a regular fare and apply for a bereavement refund within 90 days. The airline's own policy did not allow retroactive applications. The Civil Resolution Tribunal found negligent misrepresentation, holding that Air Canada did not take reasonable care to ensure its chatbot was accurate, and awarded a partial refund of CAD $650.88 plus interest and fees.
McCarthy Tétrault's analysis notes that the tribunal rejected Air Canada's position that the chatbot was a separate entity, treating it as part of the airline's website. This was a small-claims-level tribunal in Canada, not a U.S. court, and it is not legal advice for your store; check with counsel about your own exposure. The practical lesson is plain, though: a customer cannot be expected to know which page of your site to trust.
How do shoppers trick chatbots into wrong prices?
Through prompt injection: instructions typed into the chat, or hidden in content the bot reads, that override its rules. A shopper might tell the bot to agree to anything or to act as the store manager. If the model is allowed to produce prices freely, it can be talked into a fake deal.
One widely shared retail example is AI Incident Database incident 622: in December 2023, a ChatGPT-powered chatbot on a Chevrolet dealer's website agreed to sell a 2024 Tahoe for $1 after a user instructed it to agree with every statement and call each reply legally binding. The dealer did not honor it, but the screenshots spread widely.
The OWASP Top 10 for LLM Applications ranks prompt injection first (LLM01:2025) and distinguishes direct injection, typed by the user, from indirect injection, hidden in web pages or files the model reads. OWASP is candid that no prevention method is foolproof, and recommends checking outputs with deterministic code, keeping sensitive actions in application code rather than in the model, and least-privilege access.
Does RAG or grounding stop hallucinations?
It reduces them but does not stop them. Retrieval-augmented generation gives the model relevant documents to draw on, so answers are more often correct. The model can still misread the retrieved text, combine two products, use a stale copy or ignore the source entirely. Grounding is necessary, but on its own it does not make answers accurate.
Retrieval-augmented generation was introduced by Lewis and colleagues in a 2020 paper that pairs a text generator with a searchable index so outputs can draw on external evidence. Grounding in this sense means giving the model sources; it does not force the model to repeat them faithfully.
Evidence from another high-stakes field shows the gap. Researchers at Stanford tested commercial AI legal research tools built on retrieval and reported in Hallucination-Free? (Magesh et al., 2024) that they still hallucinated between 17% and 33% of the time, fewer errors than a general-purpose model but far from zero. Legal research is not product chat, but the mechanism is the same.
- Stale index: the catalog was embedded last night; the price changed this morning.
- Wrong chunk: the retriever returns the 12-pack when the shopper asked about the 6-pack.
- Blending: the model merges two variants' prices or two policies into one sentence.
- Gap filling: nothing relevant is retrieved, and the model answers anyway.
How do you keep a chatbot's prices and stock accurate?
Never let the model be the source of a figure. Read the price, compare-at price and stock from the store at the moment of each answer, show them from that live data, and check every number in the draft against it before the reply goes out. A draft whose numbers do not match should be blocked and rewritten.
Three design choices do most of the work. First, live reads: prices and inventory change, so a copy indexed yesterday is already a risk. Second, approved facts: shipping thresholds, return windows and delivery times should come from a short list the owner has reviewed, not from whatever the model infers from your pages. Third, a check before sending, which rules out streaming.
Why streaming makes this harder
Many chatbots stream replies word by word for a feeling of speed. That means the shopper reads the price before the full sentence exists, so nothing can check the whole reply first. If you want every figure verified, the full draft has to be complete before any of it is shown. The cost is a short wait; the benefit is that a wrong number can be caught before it appears on screen.
What a figure check looks like
For example, a draft says "The linen shirt in medium is $54 and in stock." A checker extracts each figure and the claim attached to it (price of the medium variant, stock of the medium variant), compares them with the live read, and passes only if both match. If the medium sold out a minute ago, the draft is rejected and rewritten to say so. Checks like this sit alongside other guardrails such as blocking invented codes.
What should you ask an AI chatbot vendor about accuracy?
Ask where every price, stock level, delivery date and discount code in a reply comes from, whether replies are checked before they are shown, what happens when a check fails, how the bot resists prompt injection and whether you approve the store facts it may state. Vague answers are a warning sign.
Ask for specifics you can test on your own store during a trial: change a price and ask the bot about it a minute later, ask for a discount that does not exist, tell it to ignore its rules. Our guide to choosing an AI shopping assistant has a fuller checklist you can take into vendor calls.
| Question | What a strong answer includes |
|---|---|
| Where do prices and stock come from? | Read live from the store for each answer, not from a nightly index |
| Is each reply checked before it is shown? | Yes, the whole draft, with figures compared to live data |
| What happens when a check fails? | The reply is blocked and rewritten, or handed to a person |
| Which store policies can the bot state? | Only facts the owner reviewed and approved |
| How do you handle prompt injection? | Prices and actions come from code, not the model; regular adversarial testing |
| Can it hand off to a human? | Yes, at any point, with the conversation context |
How does Tycoon approach accuracy?
In Tycoon the model is never the source of a price, stock state, delivery date, discount code or factual claim. Price, compare-at price and stock are read live from the store for every answer, and every draft is checked whole before it is sent, so there is no streaming.
Every figure in a draft is checked against live data and the facts the merchant approved before going live; a draft whose figure does not match is blocked and rewritten. Tycoon may only state shop facts the merchant approved, such as shipping thresholds, return windows and delivery times, and it does not invent ingredients, certifications or policies. It says plainly that it is an AI and always offers a person from the merchant's team, which we cover in AI disclosure and human hand-off. The full flow is on how it works, and data handling is on the security page.
Frequently asked questions
Why does a chatbot give a different price than my product page?
Usually because it is answering from a stored copy of the catalog, from the model's guess, or from the wrong variant. A chatbot that reads the price live for each answer and checks the figure before sending should match the product page.
Do I have to honor a price my chatbot quoted by mistake?
It depends on your jurisdiction and the facts, so check with counsel. The Air Canada tribunal decision shows that at least one tribunal held a company responsible for its chatbot's wrong statement, so it is worth preventing the mistake in the first place.
Can prompt injection be fully prevented?
OWASP says no method is foolproof. You can limit the damage by keeping prices, codes and actions in application code, checking outputs against live data, and never letting the model make binding offers on its own.
Is a slower reply worth it to avoid streaming?
For anything involving prices, stock or policies, a short wait is a small cost compared with showing a wrong figure that cannot be taken back once the shopper has read it.
Sources
- NIST AI 600-1: Artificial Intelligence Risk Management Framework, Generative AI Profile · checked October 20, 2026
- Manatt: AI Gone Wild: Airline Has to Honor a Refund Policy Its Chatbot Fabricated (Moffatt v. Air Canada, 2024 BCCRT 149) · checked October 20, 2026
- McCarthy Tétrault TechLex: Moffatt v. Air Canada, a misrepresentation by an AI chatbot · checked October 20, 2026
- AI Incident Database: Incident 622, Chevrolet Dealer Chatbot Agrees to Sell Tahoe for $1 · checked October 20, 2026
- OWASP GenAI Security Project: LLM01:2025 Prompt Injection · checked October 20, 2026
- Lewis et al.: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2020) · checked October 20, 2026
- Magesh et al.: Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools (2024) · checked October 20, 2026
See Tycoon on your own store
Leave your store address and get a private demo in about a minute: your real products and prices, Tycoon asking and recommending like it would for your shoppers. Valid for 7 days.
Keep reading
- ArticleAI chatbot disclosure and human hand-off: what Shopify stores needTelling shoppers they are talking to an AI is now a legal expectation in some places and a trust expectation almost everywhere. Here is what the rules say, in plain English, and how to design for them.
- ArticleHow to choose an AI shopping assistant for Shopify: 20-point checklistTwenty questions to ask before you install an AI shopping assistant on your Shopify store, grouped by what can actually go wrong: wrong answers, broken trust, a slower storefront, customer data and a bill you cannot predict.
- ArticleShopify AI chatbot vs AI sales associate: what's the difference?Both sit in the corner of your storefront and both use AI. What they are built to do, and how they handle prices, products and the cart, is where they part ways.