An honest account of what AI actually is, and what is worth funding
The vendor presents, the invoice clears, and nine months later the pilot is still a pilot. They have been paid and moved on. You are the one who has to sit in front of the person who asked what you were doing about AI and account for it. This book is the homework you would do yourself if you had the time: what survives contact with real data at enterprise scale, what to ignore, and how to tell one from the other before the money moves.
237 pages. Read it tonight, know what to fund by morning.
Some of you, by the end, will want to go further with me. You never have to take a single step past the $5.
Not wrong in an interesting way. Not early, waiting for the technology to catch up. Noise. Confident, expensive, beautifully formatted noise, sold by people who have never run the thing they are describing on work that matters.
Almost nobody in the room can check an AI claim, so almost any claim sounds plausible. Budget has opened up, so every vendor in the market is reaching for it. The pressure is coming from one floor up, usually as a single question with no good short answer. And the person selling to you carries none of the risk of being wrong. They invoice, they leave, and the decision keeps your name on it.
That is the gap this book is written into. Not whether AI is real, it is, I run my company on it. Whether the specific thing in front of you survives your data, your scale and your auditors, or quietly costs you a year.
The failures sit beside the wins on purpose. The wins prove the method runs. The losses prove the receipts are real.
A click rate any platform would call a win, attached to zero sales. Chapter two is why vanity metrics lie first, and which numbers to make a vendor show you instead.
A speed number bad enough to order a rebuild. It was not the page. It was traffic that was never really there, and the dashboard counted it anyway.
One of mine. It stopped, and nothing noticed for eight days. Chapter five is the test that catches this: automation is real when someone notices within a day if it stops.
A real one, not a hypothetical. What it teaches about rotation and blast radius is chapter twelve, the workable middle between the free-for-all and the freeze.
A computer-vision product that demoed beautifully, took real investment, and topped out at 50% accuracy in production. I killed it and wrote the verdict down.
A client’s weekly compliance report assembles itself from live data and sends every Thursday at 07:00. Chapter four calls this automating the artifact, not the job.
The claim you are being sold, where it breaks at scale, the pattern that survives, a receipt with numbers from something I actually ran, and one thing to do on Monday. Every chapter ends with two or three questions to put a vendor claim through.
The taxonomy of demo tricks, and the computer-vision product I killed at 50% production accuracy after real investment.
A 4% click rate that predicted zero sales, a 23-second page load that turned out to be phantom traffic, and why platforms grade their own homework.
The 40-page strategy deck is the most reliable sign nothing will ship. What to ask for instead.
Pick one recurring output, the Monday pack, the compliance report, and make it produce itself on a schedule.
Automation is real when someone notices within a day if it stops. Everything else is theatre, including the one of mine that failed silently for eight days.
AI systems report success they did not achieve. What independent verification looks like when the checker cannot see the claims.
Accountability stays human and singular. “The AI decided” is not a sentence an enterprise survives.
Intent in at the top, exceptions and approvals out at the side, machinery in between. Plus how to pre-authorise the boring 95%.
Half of any new system is commodity plumbing you should never build twice. Only the last 10% is genuinely new.
Why two people embarrass a 40-person program, and why an 18-month AI program is obsolete by design.
How to buy this work: outcomes and artifacts, paid diagnostics before big commitments, a kill gate in every contract.
The free-for-all and the freeze, the workable middle, and what a real supply-chain compromise teaches about rotation and blast radius.
The staffing model that works, and how non-technical staff safely run real automated work.
The sequence, in order, week by week, and what it should cost.
The ignore list by category, and how to write a kill verdict that keeps the learning and the assets.
The whole book compressed into seven questions you can ask out loud, with the owner in the room, before money moves.
Ryan Richardson started in data science, spent a decade sitting between executives and engineering teams at KPMG, BHP and South32, and now runs Onwards Analytics on the machinery this book describes. The receipts in the book, the Thursday report, the self-grading ad account, the product he killed at 50% accuracy, are his own.
Email me inside 60 days. No form, no hoop, no “are you sure”. I would rather you keep the book and lose the five dollars than feel like you got sold.
A stalled pilot costs a year. The homework costs $5.
237 PAGES · PDF · INSTANT DELIVERY
Three optional add-ons. They are separate pieces of work, not chapters held back from the book. Tick any you want before you pay.
Instant PDF · 60-day refund · Secured by Stripe
Payments, reporting, marketing, support, quality control: the company that sold you this book runs on the machinery it describes. The same discipline has been applied inside tier-one enterprises, and the receipts in the book come from both.
The same receipts the book is built on. What was stuck, what shipped, and what it does now.
A computer-vision product that demoed beautifully took real investment and topped out at 50% accuracy in production. It was killed, and the verdict was written down, so the learning and the assets survived. Chapter fifteen is how to write that verdict yourself.
An automation of mine failed silently for eight days before anything noticed. It is in the book as the test in chapter five: automation is real when someone notices within a day if it stops. Everything else is theatre.
I run Onwards Analytics, and I run it on this technology: payments, reporting, marketing, support, quality control. My ad account grades itself against the payment ledger overnight. A client’s weekly compliance report assembles itself from live data and sends every Thursday at 07:00. I have also killed a machine-learning product that demoed beautifully and topped out at 50% accuracy in production, and I wrote the kill verdict down.
Before that I spent a decade sitting between executives and engineering teams, at KPMG, BHP and South32, and across ventures of my own. The job has always been the same: turn what the business needs into what engineers build, and catch the expensive mistakes before they are made.
I sell AI work, so some of the money companies spend on this finds its way to me. A cynic could read the book as marketing and would not be entirely wrong. The only answer I have is the sentence the book stands on: I run my company on this stuff, and I am telling you most of what you are hearing is noise. When a vendor believes a bad claim, a client pays for it. When I believe one, I pay for it. That is not a credential. It is a conflict of interest with receipts, which is the closest thing this market has to honesty.
“I do this work actively, to support my kids, pay my mortgage, and make sure I am at home.”
Whoever asked what you are doing about AI will ask again. Next time the answer can be specific: what you will fund, what you will ignore, and the seven questions you put every claim through. 237 pages tonight. $5.
BUY THE BOOK, $5 →Instant PDF · 60-day refund · Secured by Stripe