What Happens When AI Gets It Wrong?

What happens when AI gets it wrong? You own it. If your chatbot quotes a policy that does not exist, your email tool sends a customer someone else's invoice, or your marketing copy states a warranty you do not offer, the mistake is legally and commercially yours, not the vendor's. That is not a theory. A Canadian tribunal has already rejected the argument that a company's chatbot was a separate entity responsible for its own statements.

The good news is that AI mistakes are predictable in type, which means they are catchable. This guide covers the failure modes, who is on the hook, which mistakes are expensive versus merely annoying, the verification layer that actually gets followed, and exactly what to do in the first hour after something goes out wrong.

What Kinds of Mistakes Does AI Actually Make?

AI systems fail in a small number of recognizable ways: they invent facts, they use stale information, they get arithmetic wrong, they follow instructions hidden in the material you gave them, and they express every one of those errors in the same confident tone as the truth. The confidence is the dangerous part, because it removes your natural instinct to double check.

A hallucination is when a model states something that is not true as though it were fact. It happens because a language model predicts plausible text rather than looking up verified answers. It is not lying and it is not broken. It is doing exactly what it does, in a situation where you needed a lookup instead of a prediction.

The recurring failure modes for a small business:

• Invented specifics. Policies, hours, prices, deadlines, part numbers, staff names, citations, and links that look right and do not exist.
• Stale knowledge. The model answers from what it learned during training, so it may confidently describe last season's rates or a service you discontinued.
• Arithmetic and unit errors. Square footage, tax, proration, discounts. Language models are unreliable calculators.
• Wrong record, right format. An automation merges the wrong customer into a template, and because the template is perfect, nobody proofreads it.
• Instruction following from untrusted text. If you paste in an email, a review, or a web page, any instructions buried in that text can influence the output. Treat pasted material as data, not orders.
• Tone misses. The message is factually fine and reads like a corporation apologizing to a stranger, which is its own kind of wrong.

Rates of adoption have outrun rates of review. Stanford HAI's 2025 AI Index Report found 78 percent of organizations reported using AI in 2024, up from 55 percent a year earlier, and the same report notes that responsible AI practices in industry are uneven. Adoption is easy. Checking is the work.

Who Is Liable When AI Gets It Wrong?

You are. Regulators and courts have consistently treated AI output as your business's statement, not a third party's. The vendor's terms will almost always disclaim responsibility for the accuracy of what the model produces, and your customer has a relationship with you, not with the model.

The clearest illustration is Moffatt v. Air Canada. In February 2024, British Columbia's Civil Resolution Tribunal held Air Canada liable for negligent misrepresentation after its website chatbot gave a passenger incorrect information about bereavement fares, and awarded damages. The airline argued that the chatbot was in effect responsible for its own statements. The tribunal rejected that outright and found the company responsible for the information on its own website, whether it came from a static page or a bot.

On the advertising side, the Federal Trade Commission announced Operation AI Comply in September 2024, a set of enforcement actions against companies making deceptive AI claims or using AI in deceptive schemes. The framing in that announcement is the sentence to remember: there is no AI exemption from the laws already on the books. Claims made by your AI, about your AI, or with your AI are still claims you have to substantiate.

Two related rules matter for local businesses. The FTC's Rule on the Use of Consumer Reviews and Testimonials, which took effect in October 2024, prohibits fake reviews and testimonials, including generated ones, and prohibits buying fake indicators of social media influence. Generating a glowing review with a chatbot and posting it is squarely inside that rule. And existing advertising law still requires substantiation for any claim you make, however it was drafted.

Which AI Mistakes Are Expensive and Which Are Just Annoying?

Sort every AI use into three tiers by what happens when it is wrong, then set the review level to match. Most businesses either review everything (and quit within a month) or review nothing (and find out from a customer).

Tier One: Reversible and Internal

Meeting summaries, internal notes, first drafts, brainstorms, and research starting points. If it is wrong you notice and fix it. Review is casual. Let people move fast here.

Tier Two: Customer-Facing but Correctable

Social captions, email newsletters, blog drafts, review replies, service descriptions. A mistake is public and embarrassing but fixable. One human reads every one before it goes out, and anything with a number or a claim gets checked against the source.

Tier Three: Money, Law, Health, or Safety

Quotes and estimates, contracts, invoices, refunds, insurance and benefits explanations, medical or dental instructions, hiring decisions, and any statement about warranties or guarantees. AI may draft. A named person approves, and the approval is logged. Nothing in this tier goes out automatically, ever.

If you are considering a customer-facing bot, the tiering matters even more, because the bot answers when nobody is watching. Our post on whether your business should use an AI chatbot covers scoping one so it cannot wander into tier three.

How Do I Build a Verification Layer People Actually Follow?

Make verification a short, specific checklist attached to the task, not a general instruction to be careful. "Check the AI's work" gets ignored. "Before you send a quote, confirm the square footage and the material price against the estimate sheet" gets done.

1. Ground the tool in your own documents. Give it your rate sheet, policy page, and service list in the prompt or as an attached file, and instruct it to answer only from that material and to say "not in the source" otherwise. This single change removes most invented specifics.
2. Require a source line for factual claims. Ask the tool to name where each fact came from. Then spot check. If a citation or link does not resolve, treat every other claim in that output as suspect.
3. Never let a model do the math. Calculate totals, tax, and proration in your actual system (your point-of-sale, your estimating software, a spreadsheet) and let the AI write the words around the number.
4. Add a numbers-and-names rule. Any output containing a dollar figure, a date, a person's name, or a legal or medical statement gets read against the source before it leaves.
5. Keep a human name on every message. Accountability follows attribution.
6. Log approvals for tier three. A shared sheet with date, item, and approver takes ten seconds and answers the only question that matters after an incident: who signed off.

The National Institute of Standards and Technology's AI Risk Management Framework organizes this kind of work into four functions, Govern, Map, Measure, and Manage, with a companion Playbook of suggested actions. You do not need to adopt a federal framework at a nine-person business, but the sequence is the right one: decide who is responsible, list where AI touches customers, decide how you will check it, and fix what breaks.

What Does an AI Mistake Look Like at a Local Business?

Picture a roofing contractor in Delavan with four crews and a website chat widget the owner set up in an afternoon. The widget was trained on the website and given a friendly instruction to be helpful. In week three, at 9 p.m. on a Sunday, it told a homeowner that the company offered a 25-year workmanship warranty and that a tear-off on a house that size would run about $9,000.

Neither statement was true. The actual workmanship warranty is 10 years, and the price came out of nowhere. The homeowner took a screenshot and arrived at the estimate appointment with a number in hand and a warranty expectation the company does not offer. The crew leader had to explain the gap, the homeowner felt misled, and the job went to a competitor.

Nothing was hacked. The bot did what a language model does: predicted a plausible answer where a lookup was required. Here is what the fix looked like:

• The bot was restricted to answering only from a short, approved document of services, service area, and process, with an explicit instruction that it may not state prices or warranty terms.
• Any question about price, warranty, timeline, or insurance now triggers one line: "I can get you an exact answer from our estimator. What is the best number to reach you?" That turned the failure point into a lead capture point.
• The warranty language on the website was corrected and made prominent, since the bot had been paraphrasing an ambiguous sentence.
• The owner set a Monday habit of reading the previous week's chat transcripts. Fifteen minutes, and it caught two more drifting answers in the first month.

The lesson is not that chatbots are bad. It is that the bot needed a boundary and a handoff, and the business needed someone reading the transcripts.

What Do I Do in the First Hour After AI Gets Something Wrong?

Stop the system, fix the customer, then fix the cause, in that order. Most small businesses do it backwards and spend the first hour arguing about the tool while more wrong messages go out.

1. Turn it off. Pause the automation, the bot, or the scheduled send. You can always turn it back on. You cannot recall an email.
2. Find the blast radius. How many customers received it, over what period. Pull the list before anything else.
3. Contact affected customers directly, in plain language. Say what was wrong, what is correct, and what you are doing about it. Do not use the phrase "a technical error occurred." Say the message was generated by an automated system, it was wrong, and here is the right answer.
4. Decide whether to honor it. If your system quoted a price, sometimes the cheapest resolution is to honor it once and fix the cause. Get advice before promising anything on a warranty, a benefits question, or a health statement.
5. Write down the root cause in one sentence. "The bot had no price source, so it guessed" is a cause. "AI is unreliable" is not.
6. Change one thing that prevents recurrence, then turn the system back on with that change in place and watch it for a week.
7. Tell the team what happened. Silence teaches people to hide the next one.

If the mistake involved customer data going somewhere it should not have, that is a different and more serious problem. Our post on whether it is safe to put customer data into AI tools covers the prevention side.

Do This This Week

1. List every place AI text currently reaches a customer without a person reading it first. Include chat widgets, auto-replies, review responses, and scheduled email.
2. Sort each one into tier one, two, or three, and turn off anything in tier three that runs unattended.
3. Write your four never-guess categories on one line: price, policy, deadline, availability. Give every AI system a live source for those or forbid it from answering them.
4. Add the numbers-and-names rule to your team's checklist for customer-facing work.
5. Read one week of chatbot transcripts or AI-drafted replies end to end. Highlight anything you would not have said.
6. Write a five-line incident plan: who turns it off, who pulls the recipient list, who contacts customers, who fixes the cause, where it gets logged.
7. Correct any ambiguous page on your website that an AI system is likely to paraphrase, starting with pricing and warranty language.

Where to Go From Here

The realistic goal is not zero AI mistakes. It is that mistakes stay in the tier where they are cheap, that a person stands between the model and anything involving money, law, health, or safety, and that you find out from your own review rather than from a screenshot in a customer's hand. Ground the tools in your real documents, forbid guessing on prices and policies, and read the transcripts.

If you want a second set of eyes on where your automations could embarrass you, our AI automation work starts with exactly this kind of review, and the AI Opportunity Audit maps every unattended AI touchpoint in your business and ranks the risk. If you are earlier than that, the free Local Visibility Audit on our homepage is a lower-commitment start. Either way, decide now who checks the numbers, because the answer cannot be nobody.

Sources and Further Reading

1. BC Tribunal Confirms Companies Remain Liable for Information Provided by AI Chatbot. American Bar Association Business Law Today, February 2024.
2. FTC Announces Crackdown on Deceptive AI Claims and Schemes. Federal Trade Commission, September 2024.
3. The Consumer Reviews and Testimonials Rule: Questions and Answers. Federal Trade Commission, 2024.
4. Federal Trade Commission Announces Final Rule Banning Fake Reviews and Testimonials. Federal Trade Commission, August 2024.
5. AI Risk Management Framework. National Institute of Standards and Technology, January 2023.
6. NIST AI RMF Playbook. National Institute of Standards and Technology, 2023.
7. The 2025 AI Index Report. Stanford Institute for Human-Centered AI, April 2025.