Most of my AI wasn't earning its keep

Hi {{first_name|friend}},

A power plant in West Texas got me thinking about my own AI bill this week. There is a question that links the two, and that same question is the thread in this issue.

A couple months ago I audited my AI usage. I killed 4 of the 5 automations I had running in n8n (an AI workflow automation platform), dropped that subscription plus a few others that weren't adding enough value and moved the work I'd been running through API apps into weekly routines in Claude Code and Codex. My bill dropped about 30% and the output I actually use went up.

I wasn't thinking about data centers, I was just asking myself: what decision or deliverable does this serve?

Then I read more news about data centers this week -- Amazon is backing a power plant in West Texas w/35 gas turbines that will mostly feed a single new data center. And there are many other companies doing the same (NYT counted at least 82 of these private, data center only power plants in the works across the US).

We pay a subsidized price for AI and treat it like it's free. Somebody else pays the rest, mostly whoever lives near one of those plants. And the subsidized price probably won't last. DeepSeek, the cheapest of the big AI providers, just announced a price increase of more than 4x. I'd expect others to follow.

My usage is of course a rounding error. But my business turned out to be a scale model of the same problem: AI running with nobody asking if it's worth it. My automations ran for months before I asked. If you worked in corporate tech in the 2000s, you probably remember shelfware. Companies bought hundreds of software licenses because everyone else was buying and audits kept turning up unused software. My automations were shelfware with a monthly bill.

What's running in your setup right now that you can't say exactly what it's for?

Brett

The ten minute AI audit that I ran, condensed to a short session.

Watch: the management skill that runs AI agents, in under a minute.

New research out of Memorial University (a review of 121 studies plus a survey of 361 workers) finds that 83% of Canadian workers want AI training, but badly timed, overly complex or unsupported programs add stress instead of skill for workers 55 and up, sometimes enough to push experienced people toward the exit. Why you should care: if a rapidfire AI course ever left you drained, it wasn't your ability but the training design that was at fault. Training that fits around a real workload and builds gradually is more effective. Read more

A Glean survey of 6,000 workers across the US, UK and Australia found 87% now use AI at work, saving around 11 hours a week. Then they spend 6.4 of those hours feeding it missing context, checking its output and cleaning up its confident wrong answers, a chore the study names botsitting. Why you should care: more than half the saved time goes back into supervision, so people should budget for checking answers as well. You've been reviewing other people's work for decades; that skill just became measurable. Read more

College students are logging AI agents directly into Canvas and Blackboard to complete quizzes, write papers, watch lectures and post in discussion forums while they do something else entirely. The big AI companies have mostly declined to add safeguards, and professors are left asking whether anyone real is in the room. Why you should care: if there's a student in your family, this is a dinner table conversation worth having this month and it's also a preview of a world where proving you actually did the work will become increasingly valuable. Read more

A developer built his son a private copy of Roblox populated entirely by AI bots posing as real players and didn't tell him. The internet's reaction ran from Truman Show comparisons to "dystopian," and here's an important detail: the kid played happily and never suspected. Why you should care: "is this real?" is now a skill kids have to be taught and the people who remember a world before the internet are unusually well equipped to teach it. Read more

What's one AI subscription or automation you've cut, or know you should? One word is fine, I read every one.