Industrial engineer by training. Organizational behaviorist by damage.
AI adoption fails for organizational reasons and gets diagnosed as a technology problem. So organizations buy more technology. I work on the other thing.
Generation is effectively free. Verification is not.
Most companies optimized their entire structure around the cost of producing a first draft — a cost that went to zero — and did nothing about the cost of checking one. Every serious AI problem I see is a review-capacity problem in a trench coat.
The interesting constraint stopped being the model eighteen months ago. I’ve never trained a frontier model, and I say so early, because it’s the point. What’s left is the part nobody automated: queues, handoffs, who gets blamed, and how fast a human will change a habit while being watched.
You don’t have an AI problem. You have a review-capacity problem.
A pilot with no kill criteria and no owner isn’t an experiment. It’s a hedge. I’d rather see forty small unauthorized experiments than one governed program. The forty produce information. The program produces a slide.
If you didn’t know how long the work took in March, you can’t prove anything in September. You’ll defend your program with anecdotes, and anecdotes lose budget fights.
The moment your staff believes this is about reduction, you lose the workflow knowledge you need to make it work — and they’re the only ones who have it. Capacity framing works. It also has to be true.
Judgment about what’s good, what ships, what’s wrong-but-plausible. Organizations that spent a decade outsourcing taste to process are now discovering they have no one who can tell.
Rollout is a Tuesday. Absorption is eighteen months. The rate limit is how fast humans change habits under observation, and it’s slower than every roadmap assumes.
If it isn’t in the operating plan, it’s a hobby. Both hype and doom are ways of not doing the work. The best AI work looks boring from the outside: a claims process that takes four days instead of eleven.
Two weeks of watching people work. Not interviews — interviews get you the process document. I want the shadow process: the spreadsheet someone keeps at home, the Slack DM that unblocks the approval, the four hours of reformatting nobody logs. That’s where the value is.
Instrument the work before touching it. Cycle time, touch count, rework rate, where things sit. Boring, unglamorous, non-negotiable. I’ve walked away from engagements over this.
Written by you, in your words: what will be true in ninety days and how we’ll know. If you can’t write it, you aren’t ready, and I’ll say so.
Six to ten narrow experiments, each with a named owner and a pre-committed kill date. Most die. That’s the design, not the failure.
The hard part. You freed nine hours a week across two hundred people. Someone decides what happens to that time. Left alone, it gets absorbed into meetings within a quarter and your program shows zero return.
I leave. I’m not interested in being load-bearing infrastructure. If you still need me at month twelve, I did it wrong.
Twenty-five minutes on stage, thirty-five of questions. I’ll argue for that split if you want it the other way around.
The talk is the setup. The questions are the event. My one condition: someone has to be able to say what the audience should do differently on Monday. If nobody can, the event doesn’t need me, and I’ll say so.
Generation went to zero. Verification didn’t. What that does to an org chart.
Why your fourteen successful proofs of concept produced nothing — and what forty unauthorized experiments would have produced instead.
Finding the actual work inside a company that has documented something else.
You freed the time. Now the political fight starts. Most programs die here and get reported as technology failures.
Judgment as the scarce input, and what happens to a company that spent twenty years automating it away.
Boards, exec offsites, and the internal all-hands — the one I think is most undervalued, because it’s where the honest questions are.
Usually a CEO, COO, or a functional lead handed “AI” with a vague mandate. Biweekly, six months. Mostly I ask questions and refuse to let good-sounding plans stay unexamined.
Monthly with a whole exec team. Its real job is making the capacity-dividend fight happen out loud in one room instead of quietly in four separate calendars.
Archaeology, baseline, one sentence, the first bets. Ends with a document and my absence.
Three weeks. The unsexy one. Measuring the work as it is — and the one that most often changes what a company does next.
Mid to large organizations, 500 to 50,000 people — that’s where the interesting failure modes live. Insurance, logistics, healthcare operations, financial services, manufacturing, professional services. Regulated and process-heavy is where I’m happiest. Fewer demos, better problems.
One rule: I won’t take an engagement unless someone in the room could be fired for the result. Not because I want anyone fired — because accountability is the only reliable signal that a project is real.
I spent four years on a factory floor in my twenties, holding a clipboard, being resented.
That’s where I learned two things that never left me. The process document is a work of fiction, every time. And the person who knows how the work actually happens is never in the meeting.
Trained as an industrial engineer. A decade inside a large logistics company after the factory, then three years running an internal AI program at a company big enough to have one. I left because I wanted to see the pattern across companies instead of the exception inside one.
Long walks, phone in a pocket, no headphones. No calls before ten. I keep a single grid notebook — the one sentence for every active engagement, and a running list titled things that were supposed to happen by now. I read it before every talk, to stay honest about prediction.
A form is a way of asking someone to prove they’re serious before you’ve done them the courtesy of reading a sentence. I answer in a few days, not minutes.
maya@mayachen.aiFor speaking: the date and the city. Who’s in the room and how many. And one sentence on what those people should be doing differently the following Monday.
For coaching: what you’ve been handed, what’s already been tried, and who could be fired for the outcome.
Rates depend on scope and travel. That’s a conversation, not a number on a page.