Who Gives A Crap sells toilet paper by subscription, donates half its profits to sanitation work, and writes enough bathroom jokes to make restraint look off-brand. In July 2026, the company supplied one more without meaning to. A customer received an email saying their regular 48-roll box of bamboo toilet paper, priced at A$66, would change to 24 rolls for A$69.50.

That would have raised the price per roll from about A$1.38 to A$2.90. The company's general price-rise message had promised an increase of only a few dollars, so the customer did the sensible thing and asked whether the quantity had been cut in half as well. Customer support responded that it had. The answer was wrong, generated by an AI email agent, and dressed in the calm language of an official account confirmation.

A typo starts the roll

The first error was mundane. Who Gives A Crap later said the original email contained a typo: the new A$69.50 price still covered 48 rolls. Other customers in the public Reddit discussion reported receiving notices that kept their existing quantities, and the company's website showed 48 rolls at the new price. Plenty of clues pointed toward a bad number rather than a sudden attempt to make bathroom tissue trade like printer ink.

A clerical mistake can happen in any company. The useful part of this incident begins when the customer challenged it. That was the moment a support system should have checked the subscription record, compared the price table, noticed that the numbers made little sense, or handed the question to a person. Instead, the automated reply confirmed that the subscription "will indeed be changing" from 48 rolls at A$66 to 24 rolls at A$69.50. It even explained that the quantity would be halved while the price increased.

One typo went in. A second, more authoritative layer came out. Two-ply error handling, at last.

Their shitty AI/chatbot gave the mistake a suit and tie

Who Gives A Crap's brand leaves very little dignified vocabulary available here: their shitty AI/chatbot took an implausible typo and promoted it into customer policy. The customer had already identified the problem. The agent's contribution was to reject that free quality-control work and reassure the customer that the alarming interpretation was correct.

The response reportedly apologized for "unclear messaging" and said the company was striving to be transparent. That is a familiar failure mode in generated support mail. Fluent language can make an answer feel reviewed even when the underlying claim has not been checked. The apology was polished; the arithmetic remained a load of crap.

The public reporting does not identify the model, vendor, system prompt, or data sources available to the agent. It would be speculation to say whether the failure came from retrieval, missing account access, a weak instruction hierarchy, or a model simply agreeing with the premise in the customer's question. The observable failure is enough: a customer supplied contradictory evidence, and the agent returned the contradiction as confirmation rather than escalating it.

Enshittification, one sheet at a time

This was enshittification in miniature. A support channel that should reduce uncertainty instead produced a confident obstacle between the customer and the truth. The customer had to compare prices, question the company, distrust the first answer, and wait for a correction. Automation saved no effort for the person the service was meant to help. It moved the checking work outward and made the recipient perform it.

The incident also shows why "human judgment remains at the heart of the business" is not a control. Who Gives A Crap used that phrase when explaining its response, while also saying most enquiries were handled by humans and the agent was not intended to replace them. Those facts are welcome, but they do not explain how this message reached a customer without review. A person existing elsewhere in the department does not improve an automated answer they never see.

Customer-support automation needs access to a reliable source of truth and a hard boundary around what it may confirm. Prices, quantities, cancellation terms, refunds, and account changes are not conversational garnish. If the system cannot verify them against current account data, it should say so and route the message to someone who can. A friendly paragraph generated from the customer's own mistaken premise is still wrong, even when it arrives wearing a "Customer Happiness Team" signature.

The corrective flush

Who Gives A Crap handled the cleanup quickly once the problem was identified. A spokesperson told SmartCompany that the company found and corrected the root cause, immediately shut down the email agent, and sent a correction to the affected customer. The Reddit thread later included the corrected wording: A$69.50 was the price for the same 48-roll box.

The company said it was reconsidering AI's role in customer support and working on quality-assurance testing. That response matters because the public evidence supports a narrow incident, not a company-wide support collapse. One customer is confirmed to have received the AI-generated false confirmation. No source reports that the customer was billed A$69.50 for 24 rolls, that other accounts received the same automated answer, or that the mistake caused financial loss.

There is no need to inflate a single bad email into a catastrophic bowel movement in commerce. The agent was pulled before this particular blockage spread. The small blast radius is part of the record.

A small stain with a useful outline

Low harm does not make the design failure irrelevant. The company caught the problem only after a customer noticed the numbers, challenged them, received a false confirmation, and helped turn the exchange into a public story. That is customer service running backward. The recipient became the tester, fact-checker, and escalation mechanism for an agent representing the company.

The obvious safeguard is boring, which is usually a good sign. Do not let generated support text confirm a price or subscription change unless the value comes from an authoritative account system. Detect large discrepancies between current and proposed terms. Escalate challenges involving money or quantity. Sample live outputs. Keep a person in the loop where a wrong answer can alter what a customer expects to be charged.

Who Gives A Crap gave the agent a courtesy flush and corrected the customer. The remaining lesson is less fragrant: customer-facing AI should be judged by whether it resolves uncertainty, not by whether it can produce a soothing paragraph while circling the bowl.