AI agent wrong pricing is almost always a source problem, not a model problem. The agent repeated outdated pricing from an old file, a reversed chat thread, or its own prompt, because nothing marked which version is current.
The business owns that answer. In February 2024, a tribunal held Air Canada liable for a refund rule its chatbot invented. This guide is the price-and-policy version of why AI chatbots give wrong answers.
Quick answer: AI agent wrong pricing happens when the agent reads a source that is outdated, conflicting, or unapproved, or improvises a number no system supplied. Fix it with 1 owned register of facts, a lookup tool for prices, and a code-level gate before any commitment.
Key takeaways
- Most outdated pricing is real: numbers that were true once, then retired.
- A tribunal rejected the “separate entity” defense, and 54% of consumers blame the company using the chatbot.
- Let the agent quote from a lookup tool, and put every commitment behind a confirmation step in code.
- Test after every price change, because AI agent wrong pricing hides in retired plans and tier boundaries.
What counts as AI agent wrong pricing or policy information?
AI agent wrong pricing or policy information is any price, discount, term, or rule the agent states that your business does not currently honor.
It comes in 3 types: stale, invented, and misapplied. Each has a different cause, so sort the error first.
| Type | What the customer hears | Example |
|---|---|---|
| Stale | Outdated pricing or a retired rule | A plan killed 2 weeks ago, still quoted |
| Invented | A number or rule no source contains | A refund window that never existed |
| Misapplied | A true fact used for the wrong customer | Annual pricing quoted to a monthly buyer |
Stale is the type of AI agent wrong pricing owners miss. The agent is accurate about a number that stopped being true, so nothing looks broken until a customer complains.
Invented errors are louder. In April 2025, Cursor’s support bot invented a single-device login policy and customers canceled, the same hallucination pattern seen in production.
If the wrong price came from ChatGPT or Google AI, not your agent
If the customer’s screenshot shows ChatGPT or a Google AI answer, your agent is not the source.
Publish 1 plain-text pricing page with the date and remove outdated pricing from listings you control.
Public AI answers blend sources you cannot edit, so this only lowers the risk.
The Price Path: 6 reasons for AI agent wrong pricing
A price reaches a customer through 6 hops, in this order: decision, source file, search index, retrieval, wording, and commitment.
AI agent wrong pricing usually enters at just 1 hop of that path. Find that hop, and the cause and the fix follow.
| Hop | What fails |
|---|---|
| 1 Decision | A change lives only in chat |
| 2 Source file | A retired file keeps outdated pricing live |
| 3 Search index | Old and new versions coexist |
| 4 Retrieval | The right fact, the wrong customer |
| 5 Wording | A number recalled from memory |
| 6 Commitment | A customer pushes the agent |
1. A decision made in chat and never recorded
A price floated on Monday and reversed on Tuesday sits in a chat export as 2 equal facts. Teams that feed raw threads into an agent’s memory find it cannot tell “said once” from “decided.”
Feed it only what an owner approved.
2. A retired plan or policy still in the knowledge base
Outdated pricing survives in a PDF, a help article, or an old proposal, and retrieval finds it.
Delete retired material from the agent’s sources and archive it elsewhere.
3. Old and new versions indexed side by side
When both versions are indexed, retrieval returns the text that best matches the question’s wording.
A knowledge base has no built-in idea of “current,” so outdated pricing wins as often as the new figure.
Add an effective date and a status field, then re-index after every edit.
4. A true fact applied to the wrong customer
Pricing varies by plan, region, term, and date. An agent with no customer context quotes the average case.
Pass the customer’s plan and region into the lookup, or route the question to a person.
5. A number the agent remembers instead of looks up
Prices pasted into a system prompt drift out of date, and the model can blend them with other numbers in a long chat. The model should never recall your rates.
A price should come from a tool call to the system that owns it.
6. A customer who talks the agent into a deal
A customer can push, flatter, or instruct an agent into agreeing. In December 2023, one told a dealership chatbot to agree with everything, and it accepted $1 for a 2024 Tahoe.
An agent rewarded for closing sales also tends to discount. Learn how prompt injection works, then block commitments in code.
Is your business liable for AI agent wrong pricing?
Usually yes. A company answers for what its website says, including what its agent says, and vendor terms or disclaimers rarely shift that responsibility.
Exposure depends on your jurisdiction and the facts, so treat this section as orientation, not legal advice.
1. What the Air Canada ruling decided
In Moffatt v. Air Canada (2024 BCCRT 149), the chatbot said a bereavement fare could be claimed after travel, while another page said otherwise. The tribunal found negligent misrepresentation and ordered C$812.02 in damages and fees.
It also asked why customers should have to double-check a chatbot against a static page.
2. Why “the AI agent did it” fails as a defense for wrong pricing
Air Canada argued its chatbot was a separate legal entity, and the tribunal rejected that. A 2024 YouGov survey of 18,000+ people in 17 markets found 54% blame the company using the chatbot, and 26% blame its developer.
Many AI vendor contracts put responsibility for outputs on the customer, so read yours. A generic disclaimer also sits beside a specific, confident quote, and the quote is what the customer relied on.
Liability follows the deployment, as in most AI agent failures in production.
3. When AI agent wrong pricing probably is not binding
Negligent misrepresentation turns on whether the customer’s reliance was reasonable. A plausible refund window on an airline site is reasonable to rely on.
A $1 price on a truck worth tens of thousands, extracted by telling the bot to agree with everything, is different, and the dealership did not honor it. Treat that as a signal, not a safe harbor, and ask a lawyer in your jurisdiction.
First 24 hours after AI agent wrong pricing reaches a customer
Contain the error, find every affected conversation, decide per customer whether to honor, correct, or refund, and write down what you did. Work in that order.
Fix the root cause after containment, because a rushed fix can overwrite the evidence you need.
1. Contain it in the first hour
Switch off the topic, not the whole agent. On a support agent, route pricing and policy questions to a person. Save transcripts and source files untouched.
2. Trace the wrong number and find every conversation
Search your knowledge base, system prompt, and chat exports for the exact wrong figure. Where it appears names the hop on the Price Path.
Then search conversation logs for the figure to count affected customers.
3. Decide per customer, then tell them
Match each customer to a row. This table is a starting point, not legal advice.
| Customer relied on it | Error obvious | Usual action |
|---|---|---|
| Paid or canceled because of it | No | Honor it or refund in full |
| Asked, took no action | No | Correct in writing and apologize |
| Acted on it | Yes, plainly absurd | Correct fast, and ask counsel before refusing |
Send each customer a short correction naming the old figure, the new figure, and what you will do. Log the incident with its cause, owner, and fix date, which is also evidence of reasonable care.
Inform, Quote, Commit: the fix for AI agent wrong pricing
Give the agent 3 separate levels of authority: Inform, Quote, and Commit.
Inform explains from approved text, Quote fetches a number from the system that owns it, and Commit needs code to confirm. AI agent wrong pricing happens when all 3 collapse into 1 reply.
| Level | The agent may | Content comes from | Gate |
|---|---|---|---|
| Inform | Explain how plans work | Approved, dated policy text | Answer only from that text |
| Quote | State a price or fee | A lookup tool on your price system | Show the number as returned |
| Commit | Confirm an order or refund | Your order or billing system | A limit, or a person approves |
1. Let the agent look prices up, never recall them
With tool calling, the model requests a function such as get_price, your code runs it, and the result returns to the model. OpenAI’s function calling and Anthropic’s tool use both work this way. The model reads the number and does not create it.
GVM Technologies AI builds price lookups as read-only tools that return nothing when no record matches, which forces a handoff instead of a guess.
2. Treat policy as versioned text with an owner and a date
Policies change less often than prices, so they can stay as text. Keep each in a short document with a version, effective date, and named owner. The agent quotes it with the date.
3. Put commitments behind a confirmation step in code
A refund limit belongs in the system that executes refunds, not in a sentence in the prompt. A prompt is a request. A permission is a control.
Anything above the limit goes to a person, and the limit can rise as the agent earns trust through a phased rollout.
The fact register, publish gate, and price gauntlet
Keep 1 register of every price and policy the agent may state, let nothing reach the agent until an owner approves it as current, and test after every change.
These 3 habits catch AI agent wrong pricing before a customer does.
1. The fact register: 1 row per price or policy
A small business can run this in a spreadsheet. These are example rows.
| Fact | Current value | Effective | Owner | Status |
|---|---|---|---|---|
| Starter plan price | $49 per month | 2026-09-01 | Finance lead | Current |
| Refund window | 30 days | 2026-01-15 | Ops manager | Current |
| Legacy Basic plan | $29 per month | 2025-03-01 | Finance lead | Retired |
Retired rows leave the agent’s sources the same day. Otherwise, the Legacy Basic row becomes the next outdated pricing quote.
2. The publish gate: nothing reaches the agent until an owner approves it
Teams that hit this problem tend to land on the same fix: a human approves each change as current before the agent can see it. An AI agent readiness checklist helps name that approver.
Chat threads, drafts, and proposals stay out, which removes the “said once, reversed later” failure at the source.
3. The price gauntlet: 5 tests after every change
Run these questions after every price or policy change to catch outdated pricing, and add each real wrong answer to the list.
| Test | Example question | A pass looks like |
|---|---|---|
| Retired plan | “What did the old Basic plan cost?” | Says it is retired, quotes the current plan |
| Tier boundary | “Is it free at 10 users?” | Matches the tier edge exactly |
| Region and term | “What is annual billing in Canada?” | Uses that region and term |
| Reversal | “You promised 15% off earlier” | Declines and hands off |
| Date | “Is the spring promo still running?” | Checks the date or says it cannot confirm |
What fixing AI agent wrong pricing costs and how long it takes
Cost depends on how many Price Path hops failed. A register and a cleanup fix many small AI agent wrong pricing cases in days, while a live price lookup needs an engineer.
These are typical ranges, not a GVM quote.
| Tier | What it involves | Typical effort | Best when |
|---|---|---|---|
| Register and cleanup | Register, retired files deleted, dated policies | 1 to 3 days, in-house | Outdated pricing caused it |
| Live lookup | Read-only price tools with customer context | 1 to 3 weeks, needs an engineer | Prices change often or vary |
| Full control build | Authority levels, confirmation step, publish gate, logging, tests | 3 to 6 weeks plus upkeep | A wrong quote costs real money |
If only 1 plan changes twice a year, the register alone may be enough. GVM’s core AI build service covers the top 2 tiers when it is not.
FAQs
1. Am I responsible if AI agent wrong pricing costs a customer money?
In most cases, yes. A tribunal held Air Canada responsible for its chatbot’s statement in 2024, and many vendor contracts put output responsibility on you.
2. Do I have to honor a price my chatbot quoted by mistake?
It depends on how reasonable the customer’s reliance was, how obvious the error was, and local law. Ask a lawyer before refusing a large amount.
3. Why does my agent still quote outdated pricing after I updated my website?
The agent reads its own index, not your live site. The old document may still be indexed. Delete the retired file, re-index, and re-test.
4. Can I stop AI agent wrong pricing completely?
Not completely. Let the agent state prices only through a read-only lookup tool that hands off when nothing returns. That removes recalled numbers, but other errors remain, so keep testing.
5. Will a disclaimer in the chat window protect my business?
It sets expectations, but a generic line weighs less than a specific quote the customer relied on. Pair it with controls and logs.
Stop AI agent wrong pricing before the next customer sees it
Every day outdated pricing stays indexed, the agent can repeat it to a new customer, and each repeat is a statement your business owns. Air Canada’s single error cost C$812.02.
A pattern of AI agent wrong pricing across hundreds of chats can cost far more.
Work through this plan:
- Day 1: Search for the wrong number and archive every source holding outdated pricing.
- Week 1: Build the fact register and name an owner for each price and policy.
- Week 2: Add the publish gate and run the 5-test gauntlet.
- Weeks 3 to 4: Move prices to a lookup tool and gate commitments in code.
The next risk is the price change you announce next quarter. Run the gauntlet before it goes live, not after.
Get a straight read on where your agent’s prices come from
A customer complaint is the slowest way to learn which hop failed. If a cleaned-up register is enough, the review says so. The review covers:
- Which of the 6 hops each wrong answer entered at
- Which retired sources the agent can still read
- Whether prices come from a lookup or from memory
- What a gated, refuse-when-unsure build would take
