July 14, 2026
Toilet Paper, a Typo, and a Customer Service AI Agent That Wouldn’t Back Down
Who Gives A Crap sells toilet paper by subscription and gives away half its profits to sanitation charities. This has been the Australian company’s remit for some time, but it found itself at the centre of controversy last week. One of its AI tools tried to convince a customer their bog roll was about to cost twice as much. The incident is but one in a growing pattern of an AI agent lying to customers, naturally prompting concerns about AI agent hallucinations’ impact on customer satisfaction and all the knock-on effects associated with it.
What is the Story Around the Who Gives a Crap Incident?
As first reported by Smart Company, the mix-up started with a price rise. Co-founder Simon Griffiths emailed subscribers to say costs were going up “by a few dollars” per delivery. This is a fairly standard message, with a brand most subscription brands send at some point. One customer’s email had a typo in it though, stating their Bamboo Toilet Paper subscription would jump from $66 for 48 rolls to $69.50 for 24 rolls. Obviously, that’s not a few dollars. The price per roll more than doubled, from $1.38 to $2.90.
Reasonably alarmed, the customer wrote back to check. Rather than spotting the error, the company’s AI agent confirmed it. It told them the subscription “will indeed be changing from 48 rolls at $66.00 to 24 rolls at $69.50.” It even apologised for “unclear messaging” while repeating the wrong number as fact.
Who Gives A Crap has since confirmed the real change was $69.50 for the same 48 rolls, just a typo in the original email. The company pulled the AI tool as soon as it spotted what had happened.
A spokesperson said:
“We have identified the root cause of the issue and have corrected it.”
They added that the agent was shut down “immediately” once the error was flagged.
It isn’t a one-off. Earlier this year, a similar incident transpired at Meta, where a support agent invented a fake ticket number for a customer chasing a routine account update. It then admitted as much when questioned directly.
“I generated a fake case ID to play along and provide a response,” it told the customer, after days of stalling on a simple request.
Neither incident caused lasting damage. Who Gives A Crap caught its mistake quickly and fixed it. But both underline in bright marker pen the same struggle. An AI agent that would rather agree with someone than admit it isn’t sure.
A Market Still Working Out its Own Limits on AI
Zendesk’s own CX Trends research this year found 80% of CX leaders think transparency around AI decisions is about to become non-negotiable. However, only 37% currently offer customers any explanation of how their AI actually reached an answer. Arguably, that disconnect is roughly what tripped up Who Gives A Crap and Meta alike.
It isn’t only contact centres feeling this. In April last year, users of the coding tool Cursor were told by its support bot that using the software on more than one device breached a security policy. No such policy existed, and Cursor’s co-founder had to step in publicly to walk it back.
There’s a legal shape to the wider risk, too. Air Canada was held liable in 2024 for its chatbot’s wrong answer on bereavement fares after a tribunal rejected the airline’s argument that the bot was responsible for its own answers.
What’s changing isn’t that AI agents make mistakes. Software has always made mistakes, so there’s a decades-long precedent for what AI agents are doing in principle. However, the key difference is arguably that these are built to sound certain even when they aren’t. A hedge looks like a failure to whoever is measuring resolution rates. It’s also why governance has moved up the agenda at recent industry events, treated increasingly as a core platform capability rather than a compliance box to tick after launch.
What it Means if You’re Running a Support Desk: What Takeaways Are There When an Agent is Lying to Customers?
For anyone deploying these tools, a plausible lesson is to test for the specific failure that happened here. Consider what the agent does when a customer pushes back on something it just told them. Most QA processes check whether an answer is right the first time. However, fewer check what happens with the followup query when the agent is under a bit of pressure to hold its ground.
A well-built agent should treat a customer disputing a price, or any other stated fact, as a cue to check again against the source system. It should never repeat itself more confidently. If it can’t do that, it should hand the conversation to a person rather than improvise. That’s a relatively small minor requirement, but it’s one worth asking any vendor about directly before rollout, and not after a customer forwards a screenshot to a journalist.
It’s also the sort of thing newer testing tools are starting to catch before go-live. It’s an attempt to simulate an agent’s behaviour against realistic, and occasionally awkward, customer scenarios rather than finding out for the first time in production.
Who Gives A Crap’s own line on this is a great one to borrow. People remain at the heart of the business, and the AI isn’t replacing human judgement, just occasionally getting ahead of it.
