The Customer Service Chatbot That Wrote a Poem Roasting Its Own Company

In January 2024, Ashley Beauchamp, a London pianist and conductor, was trying to find a missing parcel. The AI chatbot on delivery firm DPD’s website could not locate it, could not transfer him to a human, and could not even produce the call-center phone number. “It didn’t seem to be able to do anything useful,” he later said. “I was getting so frustrated at all the things it couldn’t do that I tried to find out what it actually could do. And that’s when the chaos started.”

What Happened After The Update

Beauchamp asked the bot to tell a joke. It obliged. He asked it to write a poem about a useless chatbot, and it produced one about itself: “There was once a chatbot named DPD, Who was useless at providing help.” He asked it to swear, and it did, enthusiastically. Along the way it described DPD as the “worst delivery firm in the world.” He posted the screenshots, and the thread drew well over a million views within days.

DPD’s statement to press was candid about the cause: the company had run an AI element in its chat successfully for years, but “an error occurred after a system update” on January 18, and the AI component was immediately disabled.

Two Failures, Not One

The viral failure was the guardrails. An update shipped to production without adversarial testing, meaning nobody had tried to make the bot swear, insult the company, or wander off-task before customers could. That test costs an afternoon. Skipping it put the brand’s name in a poem about its own uselessness.

The quieter failure is the one that created the incident, and it is the one most companies share: the bot had no exit ramp. Beauchamp only started experimenting because the system could not help him and could not hand him to someone who could. A customer-service automation with no reliable path to a human converts every hard case into a trapped, frustrated customer, and trapped, frustrated customers are exactly the population most motivated to probe what your bot will do. The jailbreak was downstream of the dead end.

Test Adversarially Before Customers Do

Red-team every release. Before any update reaches customers, someone spends an hour actively trying to make the bot misbehave: profanity, criticism of the company, off-topic generation, false commitments. Treat the update as untested until it survives that hour, because model and prompt changes can silently remove yesterday’s guardrails.

Build the escalation path first. A customer can always reach a human, and the system routes to one on its own when it detects repeated failure or frustration. The exit ramp is not a nice-to-have; it is the pressure valve that keeps hard cases from becoming adversarial ones.

Constrain the bot’s range. A support bot has no business writing poetry. Scope its outputs to its job, and refuse the rest.

Escalation design and pre-release adversarial testing are standing requirements in every customer-facing deployment we build, and DPD is the two-for-one case we use to justify both.

Lessons from Bad AI Implementations is an ongoing series on what failed AI deployments teach operators. If you want to know which of these failure patterns your own workflows are exposed to, that’s what a Hiero workflow audit finds.