The most dangerous thing AI hands you is a confident wrong number. Not a bad paragraph. You can see a bad paragraph. A wrong number sits inside a sentence that reads fine, and it rides along into the client deck, the proposal, the email to a prospect. The moment you hit send, it's yours.
For 25 years at Level 3, Microsoft and Meta I ran programs where the status report went up the chain. A number travels, and by the time it reaches a VP nobody remembers it started as an estimate. That's where I learned to ask where a number came from. The same habit now runs how I work with Claude, and in the past few weeks it has earned its keep three different ways.
Three ways it was wrong
My status file, the one Claude reads at the start of every chat, said the blog had three placeholder posts. There were four. The same file said scorecard signups were tagged by stage. They weren't; the stage lives in a field. Claude caught both by opening the real systems instead of trusting my notes. The notes were mine. The AI was the one checking. Hold that thought.
On my other business it went the other direction. The first day I pointed a chat at that site, it miscounted the items on a page. A few days later it told me some pages were missing content I knew was there. The content was embedded from another service, and the AI reads text, not what a browser draws on the screen. Both reports read as fact. Both were limits of what the tool could see, written in the voice of someone who had looked.
Then the one that mattered most. I asked Claude to write a listing for a service on that other business, and I asked for two phrases: "no AI" and "by hand." It declined. Part of that process runs on an AI agent between my checks, and the claim would have been false. We settled on wording that said what was true: a person makes the calls and a person signs off on every job. I was the one wrong that time. The tool was the check.
What "check" means
People hear "fact-check the AI" and picture reading the output slowly. That's not it. Checking is a short list you run every time, the way a program manager runs it on a status report before forwarding.
Names. Every person, company, product and page it mentions. AI is fluent with names and careless with them. A few weeks ago it opened the wrong one of three websites on a single account, because I'd said "finish the blog page" and not which site. It noticed and backed out. I still count that as a name check I failed.
Numbers. Ask where it came from. If the answer is "your status file" or "your earlier message," the number came from me, and my notes drift. If the answer is "I opened the dashboard," it's worth more. "Where did you get that" is the most useful five words I send in a day.
Sources. If it says something is on a page, it saw the text. If the thing is an embed, an image or a chart, it may not have. If it quotes an article, open the article. If it cites a figure from memory, the figure is a guess with good posture.
Make it tell on itself
Two lines, or something close to them, go into nearly every brief I write now.
"Check the live system, not your notes." Notes are a cache. Caches go stale. Mine went stale twice in one week, and I'm the one who maintains them.
"Flag anything you're not sure is true." It will. You'll get a bracket next to a date, a count, an assumption it had to make. Those brackets are the most valuable thing in the draft, and most people never ask for them because a clean page feels like a finished one.
Every post on this blog arrives the same way. Under the draft sits a table: each example in the post, where it came from, and a column for my call. The first post had nine rows. Two of them needed my OK before they could be public, and one I sent back with "no names." The post you're reading had its own table, and the sentence about my other business went to me with a question mark on it before it went to you.
And when it declines to write something you asked for, stop before you overrule it. The one time it refused me outright, it was right and I was the one who needed checking.
Who checks whom
Here's the uncomfortable part. In every example above, the error could have gone either way. My file was wrong and the AI caught it. The AI's read of a page was wrong and I caught it. I asked for a false claim and the AI caught it. Checking isn't something you do to the tool. It's something that runs between the two of you, and it only works if you treat its confident sentences and your own with the same suspicion.
Monday's update from Claude two weeks ago was accurate in every line, and I still didn't act on it until Wednesday. A correct report nobody acts on is the other way to fail the check.
Before anything with a number in it leaves your hands, ask where the number came from. If the answer is "the draft," you haven't checked it yet.