The Professionals Getting the Best AI Results Are Working Harder, Not Less

If you started using AI this year hoping it would let you work less, the professionals getting the best results from it are proving the opposite. The gap between people who get real value out of AI and people who get mush isn't which model they're paying for. It's how much supervision they're willing to put in.

What happened this week

Four separate things converged on the same point. Simon Willison, who has spent two years writing in public about coding agents, said on September 24 that the more time he spends working with them, the more convinced he is they make software engineering "even harder" — they demand more discipline and domain knowledge, not less. The same day, Ethan Mollick announced a new book, "Co-Existence," built around four things AI still can't do for you: deep knowledge, wide knowledge, taste, and agency. A day later, Cognition's coding agent Devin crossed $1 billion in annualized revenue, with Ruby creator DHH now writing only 3% of his own code by hand — his point wasn't that coding got automated, it's that reviewing AI-written code became the harder job. And on September 23, Stripe reported that sales staff using its AI platform saw 39% higher revenue and 83% weekly active usage within two weeks — but only where the rollout came with structured training, not a chatbot handed over with no instructions.

The skill implication

None of these four are saying AI doesn't work. They're saying the payoff scales with how much you invest in checking, correcting, and scoping what it does — not with which tier you're subscribed to. Mollick's four advantages translate into habits you can start this week: taste means deciding what "good" looks like before you generate anything, agency means scoping the task tightly enough that the AI can't wander, deep knowledge is what lets you catch the wrong-but-plausible answer, and wide knowledge is what tells you when a different tool would do the job better. Skip all four and you land in the other statistic from this week: McKinsey's finding that only 37% of organizations report AI contributing positively to their results, flat from last year.

Two ways this plays out

Take a marketing manager at a 40-person SaaS company who has an AI agent drafting the weekly customer newsletter. Treated like a vending machine — prompt in, copy out, send — the newsletter reads fine and converts at whatever the median is. Her actual practice is different: before each draft she spends ten minutes writing down what "good" means for that specific send — a named CTA, a tone, a product fact she'll check against the changelog before it goes out. The draft she gets back needs less rewriting each week, because her corrections are specific enough for the agent to act on next time.

A project manager at a 200-person logistics company hit the opposite problem. He pointed an agent at a stack of vendor contracts with one instruction — "flag anything unusual" — and got back a tidy summary that missed the clause that actually mattered, because "unusual" was never defined against his company's real risk list. The fix wasn't a better model. It was writing the task as a checklist of what counted as a flag before he handed it off, the same "agency" habit Mollick points to. The next batch, same agent, caught the clause.

The takeaway

The professionals pulling ahead this year aren't the ones with sharper prompts or pricier subscriptions — they're the ones treating every AI output as a first draft that has to earn trust one verified answer at a time.