Not AI theory, not hype — the practical version, written down.
A field guide to the anti-AI identity: the contradiction, the self-sealing logic, failure theater, the artisanal coders, and the one question that tells you whether you are talking to a skeptic or a belief system. With the receipts — the Codeberg vote, the Torvalds line, the car wash test — and a copyable checklist for telling governance from ideology.
Seven months after opening Claude Code for the first time, my business runs on plain text files, 34 scheduled jobs, and four rules — and most of what I built along the way is dead. The orchestrator, the outreach autopilot, and the frameworks all died for the same reason: complexity feels like progress while you're building it. Here is the machine that survived, the actual rules file it runs on, and how to start yours in about an hour.
There is no outside body evaluating AI against your production line, and the national testbed that just launched is not pointed at manufacturing. So the acceptance test has to be one you build: a frozen baseline, a read-only run across a full cycle, and a number written down before you look.
Measure AI agent ROI after review, exceptions, maintenance, failures, recovery, provider cost, and the value of useful new capability.
Near-universal AI adoption alongside almost no reported mission impact isn't a slow rollout — it's a sector measuring private habits and calling them organizational capability. Here's the difference, and the four things that actually close it.
Industrial AI succeeds or fails on whether equipment, sensor, control, operator, and business systems can hand each other a fact they agree on. A smarter model does nothing for a plant where they can't — and the fix is a vocabulary you own, not a license you rent.
Map the trigger, inputs, steps, exceptions, owner, output, and current cost of one workflow before deciding what an AI agent should do.
Agents drift because the governing layer is auto memory and a single context file. Auto memory is a suggestion, not a system, and one file has a hard ceiling. The fix is a maintained operating standard that already knows your organization — and the fix has a maintenance bill.
Agentic fundraising arrived as a category in one week. The useful question is not whether to use it — it is who holds the donor record while you do. If the context only lives inside the platform's agent, changing platforms stops being a migration and becomes amnesia.
Human approval is not one universal boundary. It gets discovered workflow by workflow, stays permanently hard on anything that cannot be recovered, and loosens everywhere else as the agent earns trust the way an employee does. Here is the loop I would actually run.
Build a company AI brain alongside one useful workflow so real work defines the context, controls, evidence, and reusable foundation.
A real agent runs by itself with little downtime and still shows a net gain after maintenance, review, confusion, mistakes, and recovery are subtracted. Here is the bill most deployments never run, and the boring week-long test I would apply before anybody calls it ROI.
Why owners can't name where AI should go first, why the savings question produces a blank stare, and the North Star I'm operating from now: remove the boring, monotonous, crappy work so people get more time being human.
Why changing today's jobs is not the same as ending work, why software is shifting first, and how lower barriers can turn AI into a larger field for human ambition.
Why I nearly skipped last week's biggest interface change, what dispatch-across-threads actually changes about a multi-project day, and what Opus 5 has been like in my own work now that model choice is a routing decision instead of a ranking one.
An AI opportunity assessment should produce measured workflows, ranked opportunities, a no-build list, and an actionable first-project brief.
I'm not afraid of AI. Smarter does not automatically mean murderous, refusing to engage does not give you control, and a lot of this panic sounds like the same person complaining about parking whenever somebody tries to build something downtown.
America cannot lead the AI era while refusing to build the power and compute infrastructure it requires.
A concrete, evidence-backed checklist for deciding whether an AI agent is ready to run real business work—not just a good demo.
A modern business website should explain the business deeply, capture real inquiries, preserve search equity, and stay monitored after launch.
Why do most AI projects fail? MIT found 95% of organizations piloting generative AI saw zero measurable P&L return. What the research from MIT, McKinsey, RAND, Gartner, S&P Global and BCG says about how corporate AI programs die, and what the successful 5% do differently.
What agentic AI is, how it differs from chatbots and workflows, what its parts do, where it fails, and how to judge a real business use case.
Pick a first AI project by scoring business value, workflow fit, available evidence, measurable outcomes, risk, and ownership—not novelty.