I don’t just play with AI. I build systems.
I kept records of my AI work and sorted it into kinds of work. Over time I learned the names the wider industry uses for the same things. The table below lists my five biggest kinds.
My work, mapped to industry terms
Pick one to open it.
Workers spend about 1.8 hours a day searching for and gathering information. That is about a fifth of the workweek.[1]
I route every session’s work into one shared vault of notes. Every rule lives in one place that every tool reads.
Only 27 percent of organizations using gen AI review all of its output before it is used.[2]
I set written standards every session must follow. An independent critic agent reviews the work before it reaches me.
AI legal research tools gave false information in at least 1 out of 6 test questions. Wrong information passed off as fact.[3]
I check every claim against its source before trusting it. Nothing counts as done until I check it in the real, live thing.
95 percent of company gen AI pilots showed no measurable return.[4]
When something breaks, I trace it to whatever should have caught it and fix that, so the same failure does not come back.
97 percent of organizations that had an AI-related breach lacked proper AI access controls. No controls means the door stays open.[5]
I set hard stops no AI tool may cross: never sending anything to anyone, never deleting anything, nothing runs without my sign-off.
These five are some of the more than a dozen kinds of AI work I track.
See the October snapshot on my homepage chartThe case studies below are specific examples of this work.
These are small snapshots of my hands-on work with AI. I check what the tools actually hand back, and I keep them from acting without anyone’s sign-off.
Verification and Proof
Checking AI’s Homework
2 hours checking what AI called finished.
The AI called the app finished. I spent 2 hours testing it hands-on and found 76 errors and pieces of false information.
Read the case studyOpens on its own page in a new tab: one moon for each error, plus the full writeup.
Still to Come
Six cases still coming: quotes I never wrote, automations without sign-off, site migrations, governance rules, a downstream break, and a plan rejected in seconds.
Eight Percent
An AI research tool gave me 187 quotes as my own words. I checked every one. Only 15 matched my words exactly. Most of the rest could not be found in my writing at all (hallucination).
Consent for Automation
Automations were sending messages without anyone’s approval. One rule I wrote now covers several AI tools and keeps a person in the decision (human in the loop).
Two Sites into One
Merging two websites means every old web address can break. I used AI tools to keep them all working (redirects) so nothing got lost.
AI Governance Starts with One File
I run several AI tools. I wrote one rule set they all read first. Fix it once. It lands everywhere. That’s AI governance in practice.
What Else Reads This List
I added one category to a shared list. Basic impact analysis: I checked what else read it. One scheduled job still worked from the old count. Caught it.
Rejected in Three Seconds
An AI tool built me a complete plan. Its first reply broke one of my rules. I rejected the whole thing in 3 seconds. That’s user acceptance.
There is always more to write.
These six are just a start. I have more than 3,000 session transcripts, and I work across tools including Claude, Claude Code, ChatGPT, Codex and Gemini. Just last weekend I held 60 sessions across my AI tools, and that is a typical weekend for me.
Sound familiar? Let’s talk shop.
Let’s compare notes on this
If you care about operations, process design, or where AI fits into both, I enjoy comparing notes with people working on the same problems.