Brilliant in the Basics
Last week I ran across a quote from Jim Mattis’s Call Sign Chaos that I had underlined in my Kindle (it showed in my Readwise review that I do daily). Mattis names three things a leader owes the people who follow him (competence, caring and conviction), and he puts competence first:
“The first is competence. Be brilliant in the basics. Don’t dabble in your job; you must master it.”
Mattis was writing about leading Marines, where not knowing your job gets people killed. In a business the stakes are usually smaller, which makes the rule easier to let slide, and AI makes it easier still now that more of the work landing on a leader’s desk starts as an AI draft.
Most of what’s written about AI and skills is about people just starting out. Stanford’s Digital Economy Lab tracks jobs for 22-to-25-year-olds , and in the jobs AI is best at, employment for that group has fallen about 2.5 to 4.5% a year since ChatGPT arrived. In the jobs AI is worst at, it has grown. The worry is that the entry-level work where people learn the basics is the first work AI takes over, so the next generation never gets the reps.
I think that worry is right. I think a version of it is happening one floor up.
Leaders have always signed off on work they didn’t do themselves. What’s changed is how wrong work looks when it arrives. A shaky forecast from an analyst used to come with tells: hedged language, a number that didn’t really ‘feel’ right or someone who got quiet when you asked how they built it. AI output has none of those (although if you ask questions, you learn quickly who at least read the work before handing it off). The contract summary, the churn analysis and the board deck all read clean and confident whether they’re right or not, so the cues a leader used to rely on to spot a problem have mostly gone away.
That leaves the basics. If you know how a number gets built, you can tell when it’s wrong. If you’ve read enough of your own customer contracts, you’ll notice when the summary of one says something they never say. Without that grip, an AI’s summary is just words you have to trust.
In 2025, Deloitte Australia delivered a 237-page report to the Australian government under an A$440,000 contract. The report cited academic papers that don’t exist and quoted a federal court judgment with words the judge never wrote. Deloitte later disclosed that a generative AI tool built on Azure OpenAI was used in producing it, and agreed to refund the final payment on the contract. The department accepted the report and published it on its own website. Chris Rudge, a health and welfare law researcher at the University of Sydney, caught the errors after the report was already public.
Checking a citation is about as basic as this kind of work gets, and two organizations had the chance to do it: the firm that wrote the report and the department that paid for it. The person who caught it was someone who knew the field well enough to see that the papers weren’t real.
Anthropic ran a study on AI and skill formation earlier this year. They had 52 mostly junior developers learn a new programming library, some with AI help and some without, then quizzed them on what they’d just used. The AI group scored 17% lower on average. Inside that group, how people used the AI mattered more than whether they used it. The few who used the AI to ask follow-up and conceptual questions about the code held onto what they learned, scoring between 65% and 86%. The ones who handed the work over and moved on scored under 40%. The groups were small, and these were junior level people learning something new rather than executives keeping what they know. The study didn’t test executives, but I’d bet on the same pattern at the top: the people who kept asking how the code worked are the ones who still knew it afterward.
So what does being brilliant in the basics look like for a CxO who reviews AI output all day?
I’d start with one test for anything you sign: could you catch a wrong answer here without asking another AI? Where the answer is no, the fix is staying close enough to the raw material to notice when something is off. That might mean having someone walk you through how one key number gets built each month, from the raw data to the dashboard, and asking where it could break. It might mean reading one actual customer contract or support ticket a week instead of the summary. And for anything going to the board, a regulator or a customer, it should mean reading the AI’s draft line by line at least once and clicking the sources. None of that takes much time.
Look again at the end of Mattis’s line: “Don’t dabble in your job.” For a leader right now, dabbling looks like approving a summary you couldn’t have caught an error in. Mattis put competence first because the people who follow you are counting on you to catch what they miss. With AI drafting more of their work, your signature is often the last check before that work reaches a customer, a regulator or the board, and staying brilliant in the basics is what makes that check something that is meaningful.