The AI Counterculture Is Your Procurement Manual
Josh Tyrangiel’s AI for Good reads like a warm story about teachers and bureaucrats. Read it as a buyer and it turns into something sharper: the method that separates the AI rollouts that ship from the ones that end in a press release and a lawsuit.
The AI projects that succeed and the ones that implode run on the same tools. What separates them is method. AI for Good documents six repeatable moves the working teams share: define the problem before buying, keep a human deciding, audit by hand, pilot before scaling, refuse what is not ready, and own the judgment instead of renting it from a vendor.
A general explains a logistics miracle
A retired four-star general walks an audience through how the country got COVID vaccines into all fifty states at once. No magic. Standardized data, one clear view of the problem, decisions made fast.
A school district launches a mascot
A district unveils an AI assistant with a press conference and a six-foot inflatable mascot. Six million dollars. Inside a year: furloughs, a whistle-blower, bankruptcy, and the founder under indictment.
Same technology. Opposite outcomes. The difference was never the model. It was the method.
The book follows a quiet group Tyrangiel calls a counterculture. Teachers, doctors, logisticians. Most arrived with no software background. They are easy to file under inspiring and then forget.
That would be the expensive mistake. What they share is not warmth. It is a repeatable way of putting AI into a system that already works without breaking it. If you buy or deploy AI for a living, that is your manual. Here are its six moves, each set against the loud version that keeps failing.
Two camps, one decision
Scroll each decision. The panel shows how the loud camp plays it, how the counterculture plays it, and what each one shipped.
Define the problem before you buy the product
Most failed deployments begin by purchasing a solution to a problem nobody has stated out loud. The counterculture states it first, in plain language, then finds the tool that fits.
Keep the human as the decider
The systems that work put AI next to the expert, not above them. The expert can overrule it, and the overrule is not waste. It is how the model learns.
Audit by hand, relentlessly
Nobody tunes a model by watching a dashboard. They tune it by reviewing real cases one at a time, including the ones it got wrong and the ones it never flagged at all.
Pilot before you scale
Scale is where hype goes to die. The wins start small, watch the data nightly, and expand only after the thing has survived contact with real users.
Decide what not to do
Refusal is part of the method. The strongest operators say no to features that are not ready and tools that solve nothing, even when the hype is loud.
Own it. Do not rent your judgment
The people in this book learned enough of the technology to bend it, test it, and sometimes ignore it. They treated the system as theirs, not the vendor's.
The six moves at a glance
| Decision | Loud camp | Counterculture | What each shipped |
|---|---|---|---|
| Define vs buy | Buy the product, announce the future | State the problem first, then find the fit | $6M lost / vaccines to 50 states |
| Who decides | Let the system make the call | Human overrules; the override trains it | ignored alerts / a model that earns trust |
| Tuning | Watch the top-line dashboard | Audit every case by hand | generic misreads / a model fit to your mix |
| Scaling | Roll out to everyone on day one | Pilot small, watch nightly, then grow | bankruptcy / a rollout users defend |
| Restraint | Adopt every feature on offer | Refuse what is not ready | a dead tool / the right tool, later |
| Ownership | Follow the vendor roadmap | Learn it enough to argue with it | a contract you cannot leave / a system you control |
Keep a human in the loop is not a free pass.
Honesty check, because that phrase has hardened into a slogan, and slogans stop earning their keep.
The hospital’s sepsis model was reliable on low-risk flags, exactly where the humans were already reliable. At the highest-risk level, the one place it mattered most, it had not cleared ninety percent, and supervising it was eating a team’s hours. At some point amplification is just expensive babysitting, and the honest move is to automate the part the machine has earned and take the human out of it.
Amplify, don’t replace can quietly become a way to protect headcount and call it ethics. It can also cap your return. Sometimes the win is removing the person from the loop, not flattering them by keeping them in it.
The book’s point is not that you never automate. It is that you earn automation through the audit loop. Replacement without the method is just the same failure at a bigger budget.
Score your last rollout
Pick the line that matches what you actually did. Not what the deck said. Six decisions, then read the result.
FAQ
What is AI for Good by Josh Tyrangiel about?+
It documents a quiet group of teachers, doctors, and government logisticians using AI to fix specific, tangible problems. Tyrangiel argues the useful work happens away from the loud debate between AI boosters and doomers, in the hands of practical people who treat the technology as a tool rather than a miracle.
Why do most enterprise AI projects fail?+
They buy a product before defining the problem, ship to everyone at once instead of piloting, and watch a dashboard instead of auditing real cases. The book’s failed example, a school district’s six million dollar chatbot, did all three and collapsed inside a year.
Should you keep a human in the loop when deploying AI?+
Usually yes, because human overrides are how a model improves. But there is a limit: when supervising a model costs more than it returns and the model has earned reliability, the honest move is to automate that step and remove the human. Keeping one in place can quietly protect headcount and cap your return.
What separates the AI loud camp from the AI counterculture?+
The loud camp, boosters and doomers alike, argues about the far future and sells hype. The counterculture is made of practitioners shipping working systems now. Their method is repeatable: define the problem, keep a human deciding, audit by hand, pilot first, refuse what is not ready, and own the judgment.
How should a company evaluate an AI vendor?+
Make the vendor define your actual problem before pitching a product. In the book, a general gave each consultant one hour to do exactly that and hired the only team that could. A vendor who reaches for the contract before the problem is the warning sign.