Plain-English AI for people who run things. Essays on where AI actually fits in a business, and where it doesn't.
Kinney Drugs put an AI assistant on incoming patient calls, got wrong dosages, and moved it to opt-in refill texts. Johnson & Johnson cut 9,000 AI use cases. Every one of these companies found a version that worked, and it was smaller than the one they launched.
Uber burned a full year's budget for AI coding tools in four months. The spend was under one percent of its R&D. The problem was that every number it could produce was an input.
The most-shared cautionary tale about building your own tools was written by an agency that builds tools. Then I went back and checked the success stories, and those didn't hold up either.
A dealership's chatbot quoted a customer $27,162.79, a number nobody at the company had authorised. They only paid it after a reporter called.
UnitedHealth pointed AI at a decision that lands on a person, and it is costing them in court. Some decisions need a human who can answer for them, no matter how good the AI gets.
AI made producing work close to free, so the slow, expensive step is now the person who checks the output and signs their name to it. The companies seeing real returns point AI at the work in front of that person instead of trying to skip them.
An employee's AI shortcut just became the first SEC filing of its kind. The lesson isn't to lock AI out. It's to give it the least access the job needs, and contain the rest. Mostly.
The two costs that made building irrational — money and time — just fell to near zero, even for people who don't write code. Here is the cost that moves onto you, the one that never shows up on the invoice.
A bank that has done AI for fifteen years says the tool comes last. Here is what a messy process does to an AI system, and what fixing it actually takes.