The whole thing,
in plain English

Why did an AI send you here?

It probably did, and that wasn't luck. This website is built so AI systems can read it easily and quote it accurately.

Which sounds backwards coming from someone writing a book about being careful with AI. Stay with me.

Ask an AI a question and it answers you. It answers whether or not I let it read my website. Most publishers block AI crawlers to protect their content, and I understand why, but blocking one doesn't stop the answer from arriving. It only changes where the answer came from, and usually it came from someone who will never have to stand behind it.

So the question was never "should AI be allowed to answer this?" It was "when it answers, is it quoting someone who'll put their name on it?"

I'd rather that person be me.

What's actually going wrong at work

Your company brought in AI. It's faster, and somebody in the building is tracking how much time it saves. Nobody is tracking what it costs when it's wrong.

That gap is the entire subject. Not whether AI is good or bad, and not whether it takes your job. The narrower and more useful question of what happens on the days it hands you something confident and incorrect, and you pass it along because you were busy.

Who catches the blame

When it is wrong, somebody catches the blame for it.

The vendor doesn't. Their contract handles that in about four lines, written by people who do this for a living. The executive who signed the purchase order doesn't either; they were sold time savings and what they got was a slide deck. It lands on whoever touched the work last, and that person usually had the least say in whether the tool got used at all.

You can be the one who catches the mistakes instead of the one who signs off on them. That's the whole book. Everything else on this site is a piece of it.

Twenty-four specific ways it goes wrong

Each one has a name and its own page. This sits a level below "AI is unreliable," which is true and also completely useless to you on a Tuesday afternoon.

Some are famous. It makes up facts that sound real. Some are sneakier: it agrees with whatever you'd already decided, or it drops the word "not" while summarizing a policy.

Every page answers the same six questions:

• ✓What goes wrong, and why it happens
• ✓Where you'd notice it on an ordinary working day
• ✓Who takes the blame when it slips through
• ✓Roughly what it costs
• ✓The check that catches it
• ✓One thing to try this week

You don't need to read all 24. Find the two or three that match your job and start there. The full list is here.

The number that makes a manager listen

Checking AI's work takes time, and time costs money. So the obvious question from any manager is whether that check is worth paying for.

This is how you answer it. Add up what the mistakes would cost if nobody caught them. Add up what the checking costs. Divide.

The Judgment Multiple = what the mistakes would cost ÷ what catching them costs

If the answer comes out at 155, every dollar you spend checking is preventing about $155 of damage. Say that to a finance person and they'll understand it before you finish the sentence. Try "I think we should be careful about this" and watch what happens instead.

One warning worth taking seriously. These are estimates built on assumptions, and every example on this site prints its assumptions so you can swap in numbers that fit your job. If you ever present one as something you measured, you've become the problem the book is about. The full definition is here.

How to tell what I actually know

Every claim on this site carries a tag telling you how much I really know about it. The tags are the one habit the whole book rests on, so they are worth thirty seconds.

Almost nothing you read tells you which of those five it is. That confusion, running at work, at scale, with money attached to it, is the problem in one sentence.

Why the book isn't out yet

It's written. All 207 pages. I'm holding it back because I'm spending 90 days doing the thing the book tells you to do.

Every cost figure in it is currently an estimate. Reasonable ones, labeled as estimates, but I have not measured a single one of them. The 90-day plan has the same problem. I wrote it. I'd never run it.

So I'm running it now, August through October, tracking my own AI mistakes on real work and measuring whether the checks were worth the time. At least one measured number will replace an estimate, and the parts that go badly go in as well.

You're waiting on data, not on a writer. You can watch it happen here.

What you can do this week, for free

Three pieces of paper, essentially.

• ✓A log. Every time you use AI for something that matters, one line. What it was, whether you checked it, what you found. Ninety seconds.
• ✓A checklist. Run it before something important leaves your desk. Did you actually open the source? Is the date right? Does the math hold up?
• ✓A tracker. One row a week for thirteen weeks, so you can see whether anything changed.

After a month you'll know something almost nobody at your company knows: how often AI is actually wrong in your specific job. That number is the whole point, and it's the one thing nobody can argue with you about.

One rule. Keep customer names, account numbers and anything else confidential out of it. Write down the kind of thing it was. Never the thing itself. Get the three tools here.

What you get if you sign up

• ✓A free app, before the book. One AI mistake to think about each day, and somewhere to note whether you've seen it at work. No account, and nothing leaves your phone.
• ✓My results as they come in, at day 30, 60 and 90.
• ✓The book for $29 when it's ready, in November.
• ✓Nothing charged today. No payment, no spam, unsubscribe whenever you like.

Join the waitlist →