Can you trust AI? Five real mistakes and who paid for them

AI tools are useful, and they get things wrong. For a company, the question isn’t whether its AI will make a mistake. It’s who catches the mistake before a customer, a client or the budget does. In the five cases below, nobody did, and the company that used the AI paid for it.

1. Air Canada: a refund rule the airline didn’t have

Air Canada’s website chatbot told a customer that he could ask for a bereavement fare after his trip. The airline’s policy didn’t allow that. When the customer took the airline to a tribunal in British Columbia, Air Canada argued that the chatbot was “a separate legal entity” responsible for its own actions. In February 2024 the tribunal rejected that argument and made the airline pay (ABA Business Law Today).

What would have caught it. Answers about money, refunds or policy come from a page a person approved, not from a model improvising. The chatbot can point to that page; it shouldn’t write the rule.

2. Deloitte Australia: a report with sources that don’t exist

In October 2025, Deloitte agreed to partly refund the Australian government for a report it had been paid AUD 440,000 for. The report contained a made-up quote attributed to a federal court judgment and references to research papers that don’t exist (AP via CTV News).

What would have caught it. Someone who opens every source before the document goes out. AI can draft; a person signs.

3. Meta: campaigns that spent the day’s budget in hours

In 2024, advertisers reported that Meta’s automated Advantage+ shopping campaigns were spending a day’s budget within hours, with costs up to 10 times higher and few or no sales. Meta said it had fixed the technical issues (eMarketer).

What would have caught it. Spend limits set by a person, and someone checking spend against sales every day, not only against the platform’s own report.

4. Google Ads: broad match by default

In July 2024, Google made broad match the default for new Search campaigns that use Smart Bidding (Search Engine Land). Broad match lets Google show your ad for searches it considers related, which can raise costs if nobody adds negative keywords. In December, some advertisers found their keywords switched to broad match when they changed bidding strategy, without being asked; Google said the issue was under investigation (PPC News Feed).

What would have caught it. Reading every setting the platform changed before it spends, and reviewing the search terms every week.

5. OpenAI: fluent and wrong

In OpenAI’s own tests, reported in April 2025, its o3 and o4-mini models made things up in 33% and 48% of their answers to questions about people (PCWorld).

What would have caught it. Treating a fluent answer as a draft, not a fact. If the answer goes out in public or moves money, a person checks it.

What we saw in our own account

In the Google Ads account of our martial arts academy in Mexico City, Google showed the search terms behind only 63 of 363 clicks over ten days in October 2026. The other clicks were paid for without showing which searches they came from. When we built the kids’ campaign, the ad copy Google wrote automatically had to be rewritten line by line.

Four rules that prevent this

1. The goal comes from your business. Your margins and the leads that close, not what the platform counts.
2. A person approves every change. No recommendations applied automatically, no machine-written copy published unread.
3. AI decides alone only where mistakes are cheap. If a mistake is expensive or public, AI assists a person.
4. Check every month against your own numbers. Your invoices and sales data, not the platform’s report.

You don’t need AI for everything. Where you use it, put an expert in charge.

Score a task with the free “AI or not?” tool, or book a diagnostic.