In September I spent an hour on the phone with the general manager of a 600-person aluminum mill. He is a metallurgical engineer. He last wrote code in Turbo Pascal and C++, so software doesn’t scare him. He had tried the obvious thing, uploading one of his planning spreadsheets to a chatbot and asking for help. “It’s giving me some Python code,” he told me, “and I go, okay, what do I do with this now?”
He wanted to know how to get the whole mill using AI. Most executives I talk to want the same thing, and many have already picked a method. I described it to him that afternoon, a little unkindly. “The board told us we need to do AI, so we’re going to do AI. And the way that we’re going to do that is we’re going to roll out [enterprise AI] licenses to everybody in the organization.” Then I said what I think of it: “That is a disaster.”
I still think so, though it took me a while to say why in one sentence. AI multiplies judgment. A careful person gets faster and so does a careless one, and the tool gives you no help telling them apart. A universal rollout hands the same multiplier to everyone on the same day, and what comes back looks equally polished whether the judgment underneath was good or bad. The better move, for now, is to arm a few people first, choose them carefully and teach them one at a time. That is an argument about sequence. Whether it eventually becomes an argument for everyone is the part I can’t answer yet.
Four quadrants
In May I showed a room of portfolio-company CEOs a two-by-two. One axis is judgment. The other is how much a person uses AI. Every company already has a bottom row, the people whose judgment you don’t trust. Without AI they are what I called dead weight, and you know who they are. Then I described what happens when you move them to the right. “If you give your entire organization AI, you’re going to be giving a bunch of people with bad judgment… tools that allow them to 10x their bad judgment. Which we call slop cannons.”
The top row is more pleasant. Smart people who won’t touch AI are steady hands. “They can still be reliable,” I told the CEOs, and I meant it.11 Nothing in the matrix says steady hands are a problem. Plenty of good companies run on them. They are also, in my experience, the slowest to agree to move right. Give those same people the tools and some help and you get what we call, only half joking, turbo brains.
A license moves a person along one axis. Nobody’s judgment improved because they got a login. Roll AI out to all 600 people at the mill and the whole chart slides right at once, which means the bottom-left quadrant empties into the bottom-right. Slop cannon output is also hard to spot. A bad memo used to look like a bad memo. Now it arrives formatted, confident and long, and someone with good judgment has to spend an hour finding out it’s wrong.
There is decent outside evidence for why judgment matters more with these tools, and it comes from people who are paid to have it. In 2023 a group of academics, Ethan Mollick of Wharton among them, ran an experiment at Boston Consulting Group that used “a full 7% of its consulting force (758 consultants).”22 Ethan Mollick, “Centaurs and Cyborgs on the Jagged Frontier,” One Useful Thing, September 16, 2023. On realistic consulting tasks that the AI of the day could handle, consultants using it finished 12.2 percent more tasks, finished them 25.1 percent faster and produced work rated 40 percent higher in quality.33 Fabrizio Dell’Acqua et al., “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality,” Harvard Business School Working Paper 24-013, 2023.
The researchers also gave consultants a task chosen to sit just outside what the AI could do well. It looked like the others. On that one, consultants with AI did worse. Without AI, 84 percent got it right. With AI, only 60 to 70 percent did.44 Mollick, “Centaurs and Cyborgs on the Jagged Frontier.” The working paper’s abstract reports the same result as a drop of 19 percentage points. The paper’s title calls the boundary between those tasks a “jagged technological frontier,” and the adjective carries the argument. What the model does well and what it botches don’t line up with how hard a person would rate the task. The line zigzags.
Nobody marks the frontier for you. Telling which side of it a task sits on, before you hand it over and again when the answer comes back, is the judgment the matrix is about. A person with good judgment notices when an answer is too smooth. A person without it forwards the answer.
An engineer I coach, a few weeks into a job as the AI lead at a venture firm, gave me the best name I’ve heard for the failure. The firm had told its staff to become AI-native. He had interviewed about half of them, and most either didn’t trust the tools or overrated themselves. “They think they’re good at AI, but in reality they’re casino coding. They’re just pressing the button hoping that they’re going to hit the lucky sevens.”
The unevenness shows up even among people who have signed up to learn. This fall Matthew and I helped teach a six-week course for CEOs of search-fund-backed companies. Before the first session, participants rated themselves on four areas of AI fluency, on a scale of one to four. Matthew summarized the heat map for the room: “Mostly ones and twos out of four.” A few weeks later an outside facilitator opened a session with an anonymous poll that finished the sentence “AI in my work feels like…” Six people chose “a fire hose I’m drinking from.” Three chose “a new hire I’m still training,” and three “a colleague I’m starting to trust.” Nobody chose “something I’m supposed to be doing.”55 Nine people didn’t vote. The facilitator generously assumed they hadn’t found the button. That is a room full of people who want in and mostly haven’t found their footing.
Credentials don’t predict fluency either. The engineer at the venture firm shared an office with a new associate whom colleagues had been calling the firm’s first AI-native hire. The associate had spent a decade as an engineer at a government space lab. He didn’t use the tools. Asked why, he gave two words: “confidence, trust.” He also gave some arithmetic. He could spend two hours building an automation that might not work, to save two hours of work he already knew how to do, or he could just do the work. He kept doing the work. His arithmetic is right for one task on one day and wrong over a year, and the gap between those answers is mostly confidence. The engineer drew a broader conclusion: upskilling everyone is “extremely difficult,” and the firm should probably “just hire for the talent that you’re looking for.”
Who gets the keys
So who goes first? My answer hasn’t changed much this year. At a September office hours with the same CEO cohort, I put it this way: “It’s the person who learned how to use Excel macros, not because they wanted to, but because they needed to figure it out to make their job better, that’s the person that you want to give unfettered, unlimited access to AI.” In most companies you can name that person without looking at an org chart.
The macro person brings two things. They know the work well enough to recognize a right answer, and they have already shown they’ll grind through tedium to make a tool behave. Today’s AI demands both. The people who have done well in our engineering role were obsessed enough to build things for their family, because so many small things are not yet solved for you and you have to push through them. That takes thousands of hours. It’s why my one hiring question for that role is “What have you built on nights and weekends with AI?” If somebody can’t answer that, they can’t be an operator engineer.
Matthew comes at the same trait from another side. “AI fluency is just delegation,” he told the cohort in its first session. “People who are great delegators are great with AI.” His version of good delegation runs in order: define what good looks like, set the standards it has to meet, supply the context and tools, then work alongside and improve. It is, he pointed out, “the exact same process of when you get a new employee.” The macro person and the good delegator share a habit. They hold a picture of the finished result in their head and check the work against it, which is what a jagged frontier requires.
Once you’ve found them, the advice gets specific. Give those two or three people unrestricted access to the best models from more than one lab, and let them switch as the frontier moves, because it “changes on a weekly basis.”66 I’m not naming models here. Whatever I printed would be stale by the time you read it. Everyone else can have a basic plan. Then, as I told the cohort, “parachute [them] in as a SEAL team to the rest of the company in prioritized order.” I gave them a picture of what that means: “send them into accounting, solve a bunch of problems there, and then when they’re done, send them over here, solve a bunch of problems there.” In my experience that works better than “a broad shotgun approach giving everybody AI.”
Sequencing also protects the first impression, which matters more than most executives expect. A skeptical employee who gets burned once by a new tool tends to stay away from it. A slop cannon in every department guarantees a lot of first impressions like that. A small team working down a priority list leaves behind something that works, built by a colleague people already know.
One at a time
Arming the few still means teaching them, and on how to do that I’ve already changed my mind once this year. We ran a workshop for about 20 people. “It was fine,” I told a peer in October. “People learned something.” Then I started sitting down with executives one or two at a time, “for two one-hour sessions in person if possible,” and found it “extremely valuable for getting executive-level people educated and producing.”
The difference is mostly logistics. As I explained to the mill GM, there is “often about an hour of those little gotchas to sort through,” such as the email connection that needs a ticket to IT first, or the phone that won’t talk to the laptop. In a room of 20, an hour of gotchas per person means most people leave before anything works. “We’re at a stage where it’s early enough that those kinks haven’t all been worked out, and you can’t just do a workshop for this. We’re not there yet.”
The other change is what people build first. My instinct, and every executive’s, is to start with a real business problem. I now start with a personal one. As I put it to the GM: “Make a tool for your wife, make a thing that gets you new insurance for your car… You don’t have to ask, you don’t have to build consensus, which is the hardest part of a lot of these projects. There’s no change management, there’s no risk on company data.” A personal project strips out everything that makes work projects slow, so the only thing being tested is whether the person can get the tool to do what they want. Companies push back. “We want to get to work” is the usual response, and I understand it. But Matthew’s observation to the CEO cohort matches mine: “everybody we know who is highly effective at AI starts with personal projects.”
My own position contains a contradiction, and I’d rather show it than tidy it. Four days after telling the mill GM to arm a few people, I told the engineer at the venture firm what I actually suspected, which is that the thing that “really will transform the organization” is “just leveling everybody up.” Then I added: “What I don’t know is how trainable that is.”
I’ve said nearly every possible thing about how trainable it is. I’ve told a room of CEOs that “anybody can develop this skill, we train for it all the time,” and in the same talk that it “just takes thousands of hours.” In September I told the GM that the combination I look for, technical acumen and business acumen plus some taste and enough charisma to sit next to someone and learn their job, is “impossible to train for.” A few minutes earlier I had told him that getting someone from the 30th percentile to the 80th or 90th is “a 4-hour exercise.” These can all be true if they describe different things. Four hours might get a willing person past the gotchas. Thousands of hours might be what it takes to become someone who builds things for their family on weekends. The charisma may not be teachable at all.
Some evidence points toward optimism. My wife is doing this now, and a year ago she wasn’t. My mother is starting to, and she is 80. Other evidence doesn’t. Companies come to us with a seven-person development team in which four people are “super AI-pilled,” one is “a resistor” and everyone else is somewhere in between, and they ask how to get the most out of all of them. “I don’t have a good answer for that,” I admitted in October. I still don’t.
So for now the advice is about order. Find the two or three people who already work like turbo brains, give them the best tools and a ranked list of problems, and teach the next few one or two at a time on projects that matter only to them. Keep a list of who asks to go next. When I last heard, the engineer at the venture firm had started sitting with the space-lab veteran to automate some of his daily work. I don’t know whether a few hours side by side will be enough. I’d like to find out before I tell another CEO who deserves a license.