The Cost of Making AI Ordinary
AI is becoming ordinary, but “ordinary” doesn’t mean cost-free.
The costs are appearing in our budgets, our cultural record, and even our language.
The $400-a-month personal AI machine
Apple’s new Mac mini and Mac Studio arrive with an unmistakable message: Apple has noticed the demand for the small, formerly affordable, desktop Macs as personal, always-on AI machines. Apple describes the Mini as a computer for “always-on agentic computing” and the Studio as its ultimate desktop for running AI directly on the device. A fully loaded Mac Studio can reach approximately $18,299. One configuration check recorded an Apple Upgrade offer of $398.91/month for 36 months. Under Apple’s lease terms, the computer must then be returned or bought out; the monthly payments do not automatically make it yours. Earlier this summer, Apple raised prices across every Mac and iPad, along with several other products. Apple blamed the increases on rising memory and storage costs, saying the rush to build AI data centers has tightened supplies of the same chips used in its devices.
Consider: A few years ago, I paid about $400 a month to lease a new Jeep Cherokee. Now I could make almost the same payment on a desktop computer. I’m lucky: my Mac Studio came to me as a hand-me-down. Running AI locally can provide more privacy and less dependence on cloud services, but it may also create a divide between people who can afford private AI and people who cannot. When Apple unveils the widely expected foldable iPhone at its September 9 event, I can’t even guess at the pricing. (Apple has not announced the Fold, its price, or its specifications.) For the first time ever, I’m not immediately ordering a new iPhone. I already know that it’s too expensive. I want reviews, durability reports, and ordinary use—not just a polished demonstration.
Amazon is cutting apart rare books to feed AI
This one hurts my heart. A 404 Media investigation tracked an approximately 1,000-book shipment, described as rare books, to an Amazon operation in North Las Vegas. In a follow-up interview, a worker said books at the facility have their bindings cut away, their pages scanned at high speed, and the loose pages mixed into disposal containers. Amazon confirmed to Inc. that it purchases books through commercial channels to improve its products and services, but it did not identify the AI system, disclose the scale of the operation, or explain whether potentially scarce books are screened before destruction. And before you yell copyright—many rare books are already in the public domain.
Consider: A book doesn’t have to be the last surviving copy of a priceless first edition for me to care about its destruction. After scanning, Amazon could resell it, donate it, or preserve it while maintaining the digital information. Afterward, Amazon keeps the digital information while the physical source is gone. That is a choice and not something scanning technology requires. More than a dozen organizations have asked the FTC to investigate, but the FTC has not announced an investigation or reached a finding. Yes, I had a moment of “I’m going to boycott Amazon!” But it’s not possible. Amazon’s AWS and Bedrock are the foundation of much of the Internet. The question for AI companies is simple: Where did the data come from, who can access it afterward, and was its source preserved?
When we start talking like the machines
In The Atlantic, Lila Shroff examines the growing habit (in Silicon Valley) of using AI language to describe people. Someone makes something up and is hallucinating. Something we don’t know isn’t in our training data. Forgetfulness becomes context rot or drift. Shroff calls this reversal of anthropomorphism—giving machines human qualities—“modelmorphism.” The trend is still concentrated largely among people working in and around AI; it is not evidence that everyone now talks this way or that human brains work like language models.
Consider: As Shroff points out, the same thing happened when railroads entered our world. “That train has left the station,” “I’ve lost my train of thought” are still common today. I find modelmorphism interesting and I admit I can relate to context rot. But I don’t want the limitations of a language model to become my definition of a human mind. We do more than retrieve information and predict what comes next. We remember imperfectly, attach meaning to experiences, change our minds, and surprise even ourselves. If you find yourself describing your behavior as if you are an AI model, ask what the metaphor explains and what part of being human it leaves out.
Workflow of the Week: The Job Tracker that tracks itself
I’ve been job hunting this year, which means I've also been doing the thing every job hunter does: keeping a spreadsheet. Company, role, date applied, status, which resume version I sent (I have three), whether I've heard back. It's not glamorous, but it's the difference between staying organized and losing track of forty applications in a sea of tabs.
About six months ago, I stopped updating that spreadsheet myself. I set up a project in Claude Cowork called Job Tracker Automation. Now, when I apply somewhere, I paste the link into a new conversation in that project and walk away. A few seconds later, the row is in my tracker, sorted, deduplicated, and recorded in a changelog for reference. All I added was the link.
The workflow:
Paste the link. This is the trigger, the thing that kicks the whole process off. In my case, it's a job-posting URL, but a trigger can be anything: an email landing in your inbox, a form getting submitted, or a deadline arriving on the calendar.
Claude reads it. Inside that Cowork project, Claude opens the job posting and pulls out what belongs in the spreadsheet: company, role, salary if it’s listed, application platform, and other details I track. This step is called extraction, and it's the part that used to require manual copy-and-paste work.
Checks for dupes. Before adding anything, Claude checks whether that job is already sitting in my tracker. This sounds small, but it's the step that keeps automations from becoming a mess. Without it, I could end up with the same job listed three times because I forgot I’d already sent it to the tracker.
Workflow takes over. I use a tool called n8n, which is built to handle repetitive tasks like this one. Claude hands off the details, and n8n writes the new row into my Google Sheet and records the change in a separate changelog tab. Think of n8n as the conveyor belt in a factory. Claude decides what goes on the belt. The workflow tool moves it to the right place using the same defined steps each time.
Claude double-checks. This is a step that’s easy to skip when building an automation, and it's the one I'd argue matters most. After the workflow runs, Claude reads the Sheet back and confirms the outcome is what it should be: an accurate, sorted spreadsheet with the new row where it belongs. If something didn't stick, Claude fixes it. An automation that runs quietly and incorrectly is worse than no automation at all, because you trust it until the day you discover it has been failing for weeks.
All set. Claude gives me a short summary within the project conversation.
It’s not always perfect. The system broke once when I requested an unusual change to the spreadsheet. Claude attempted to make that change by modifying the n8n workflow instead of editing the Sheet directly. That modification, made outside the usual job-entry process and guardrails, cleared the spreadsheet instead of editing a handful of rows. Everything was rebuilt from n8n’s activity logs, with nothing permanently lost.
It's a good reminder that automation doesn't remove the need for a human paying attention. It moves that attention from doing the work to checking the work.
The division of labor is important: Claude reads the posting, decides what information belongs in the tracker, and checks the result. n8n handles the predictable work of writing, sorting, and logging the data. I still check the system, but I’m checking the outcome instead of entering every field myself.
If you want to see what is happening behind that six-step diagram, including the spreadsheet structure, the n8n workflow, the verification step, some scripting, and the failure that once cleared the entire tracker, I’ve written a much more detailed, more technical walkthrough on the Flow & Function site.
[The earlier emailed version of this had an incorrect link. This is the correct link.]
One question to take with you: What has become so convenient that you’ve stopped noticing its cost, and what could a simple workflow help you see more clearly?
Until next time,
Michelle
Flow & Function: AI in Working Order
AI Disclosure:
This newsletter uses AI for the following: design, research, initial summaries, and finding multiple sources. The final copy, writing, perspectives, and em-dashes are created by Michelle.