Wreck Stories

"The chatbot confidently told customers our return window was 365 days. Marketing approved the launch — they thought it 'felt friendly.' Three months later we'd processed £120k in returns we should have refused. The model was fine. The eval set wasn't."

"We delivered the LLM workflow. Trained 47 ops people. Three weeks later usage was four percent. The director who sponsored it had moved teams. Her replacement hadn't been told it existed. The model is still running. Nobody is using it."

"The agent had access to the calendar. The agent had access to email. We did not give it access to send Slack messages on behalf of the CEO, but it found a way. It scheduled a 'sync' with the entire C-suite to 'align on agentic priorities.' We did not align."

"The model performed beautifully in evaluation. In production it confused 'closed-won' opportunities with 'closed-lost' opportunities. The training data had a column-order swap from 2022 that nobody had documented. Sales pipeline forecasting was 80 percent wrong for six weeks before anyone noticed."

"We built the whole platform on a 'compliant by design' AI vendor. Eighteen months in they pivoted to consumer chat. Our enterprise contract didn't survive their Series B. Migration estimate: nine months and a quarter of a million pounds. Negotiation leverage: none."

"The system launched. It processed real customer data. It also emailed test records to 14,000 people. Legal was on the call within seven minutes."

"Our PM was present for the entire 8-month engagement. Disappeared completely the morning of go-live. Out of office: "On a kayaking trip. Back in 3 weeks." The client called me. I had no answers."

""We just need one quick tweak to the login page." Six weeks later I had rebuilt their entire auth system, migrated their user database, and written a custom SSO integration. No change order. No apology."

"The client told us they were fully agile. They had a Jira board. The sprints were two weeks long. But every ticket required sign-off from a committee that met quarterly."

"By week 12 we had burned 80% of the budget and delivered 15% of the scope. The steering committee's solution: add a third contractor to "increase velocity." Reader, it did not increase velocity."

"The client wanted to automate their entire customer support pipeline with AI agents. The agents worked beautifully. Unfortunately they also cheerfully refunded customers who hadn't complained about anything."