By 2027, conversational AI is expected to handle 70 percent of customer service interactions, up from about 50 percent today, according to Gartner.
The driver? Generative and agentic AI. The catch? We've already seen two years' worth of high-profile AI misfires, and they are a warning shot of what's to come.
Some bot blunders made headlines for laughs, like Virgin Money's getting flustered over the word "virgin" or DPD dubbing itself the "worst delivery company in the world," but others cut deeper.
For instance, one ChatGPT-powered agent on a car dealer's site agreed to sell a $75k Chevy Tahoe for just $1, a "deal" that was treated as binding, but the next three examples had far bigger consequences, each one worse than the last.
1. Air Canada Gets Sued for a Serious Bot Blunder
In February 2024, Jake Moffatt, grieving the loss of his grandmother, turned to Air Canada’s chatbot for information about bereavement fares.
The bot incorrectly told him that he could purchase tickets at full price and then apply for a refund within 90 days after travel. Trusting this, Moffatt bought the tickets.
When he later tried to claim the refund, Air Canada denied it, explaining that bereavement fares don’t apply to completed travel. Frustrated, Moffatt sued and won, collecting $650.88 plus interest and fees.
As part of the ruling, a Civil Resolution Tribunal member noted:
“In effect, Air Canada suggests the chatbot is a separate legal entity that is responsible for its own actions. This is a remarkable submission.”
While $650.88 may sound minor, the real damage wasn’t financial; it was reputational. Stories like this spread quickly, raising questions about trust, accountability, and brand reliability.
The case underscores a critical truth: if your AI isn’t trained and monitored properly, it can misinform customers, damage brand credibility, and turn what should be a helpful tool into a liability.
2. Just Go Ahead and Break the Law, Says NYC’s Official Chatbot
From a lawsuit to lawbreaker: New York City’s “MyCity” chatbot misled small business owners with dangerously inaccurate advice and, at times, effectively instructed them to break the law.
The bot told shop owners that they could go cashless, contradicting a 2020 law requiring NYC stores to accept cash. It also gave a flat “no” when a landlord asked if they had to accept tenants using rental assistance, despite it being illegal in New York to discriminate based on sources of income.
A local housing policy expert called the tool “dangerously inaccurate,” while critics blasted the approach as “reckless and irresponsible.” Even Mayor Eric Adams faced backlash after defending the bot in a tense press conference.
Here’s the hard truth that they don’t tell you: If an untrained AI doesn’t know the answer, it doesn’t stay quiet, it makes one up. When that happens in high-stakes contexts like housing or compliance, the fallout isn’t just confusion; it creates legal risk, public outrage, and long-term damage to brand reputation.
3. A 400-Word Prompt Tricks Lenovo’s Virtual Agent Into Leaking Sensitive Data
In August 2025, Lenovo’s AI chatbot “Lena” was tricked into exposing sensitive company data with nothing more than a 400-character prompt.
Researchers at Cyber News discovered that this tiny input could leak live session cookies, which were enough for an attacker to bypass logins, hijack active chats, and sift through past conversations. The exploit combined an innocent product query with a hidden HTML switch, a fake image link, and a push to display it, effectively turning Lena into an insider threat.
Lenovo patched the flaw quickly, but the damage could have been far worse. A breach of this kind could have opened the door to massive data exposure, compromised customer trust, regulatory investigations, and an enduring black mark on Lenovo’s reputation.
The lesson is clear: AI chatbots aren’t just helpful, they’re vulnerable. Their likelihood to comply can be weaponized, and without rigorous safeguards, what starts as a customer service tool can spiral into a brand’s biggest liability.