Being Good at AI Is a Management Skill
I never knew my years of management experience had prepared me for AI.
When I think about what it takes to be good at using AI, I keep landing on the same comparison: it’s a lot like being a good manager.
You’ve probably heard AI compared to a super-smart intern. Cheap, available the second you need it, willing to take a swing at anything you hand it. Also capable of making a real mess if you’re not paying attention.
A good manager knows how to get the best out of an intern like that. Turns out the same instincts apply to AI, for mostly the same reasons.
Give clear instructions
A good manager doesn’t just assign a task, they define the outcome, the constraints, the deadline, and they make space for the intern to ask questions if something’s unclear.
AI needs the same thing. “Generate me a Q3 report” and “Generate me a Q3 sales report in Excel format, pull the latest numbers from our CRM connector, flag our biggest wins and losses from the quarter, and ask me anything you need clarified about how to present it” are not the same prompt. The gap in output quality between those two is enormous, and most people are still writing the first one.
Review the work
Not everything needs a second look. A good manager has a feel for which tasks are low-stakes enough to trust and which ones need a closer read, based on how reliable the person doing them has been.
Same with AI. It’s genuinely capable, and just as quick to get things wrong when nobody’s checking.
Wong’s story
I use AI to draft JIRA stories for my team’s technical work, and in the name of moving fast, I stopped reviewing the details before publishing them. A few stories went out with scope that wasn’t actually accurate. My team picked them up, started building against requirements that weren’t real, and we burned real hours before anyone caught it. The fix took five minutes. Not catching it cost us a day.
That’s the trade-off in one example: AI saved me time writing the story and then cost me more time than it saved, because I skipped the one step that would’ve caught it.
Know your employee
A good manager knows what someone’s good at, where they’re weak, and how to talk to them in a way that actually lands.
That applies at two levels with AI. Which product you reach for is the first one. Claude is weak at image generation but strong at producing business-style documents and writing that reads like a person wrote it. Gemini has the deepest reasoning scores and the best native handling of video and huge documents. ChatGPT plugs into more of the tools companies already use, things like Microsoft 365 and Slack. None of them wins everything, and most people doing real work end up using two or three regularly instead of picking just one. (Worth its own post.)
Which model tier within a product is the second. This one cost me actual money to figure out.
Wong’s story
I had Opus read through a few hundred documents to produce an analysis, and it burned through tokens doing it. The next time, I had Opus fire off Haiku subagents to handle the reading and fetching, then had Opus focus only on the analysis once the legwork was done. Same output, a fraction of the tokens. It’s the same instinct as managing people: you don’t put your most expensive person on the grunt work.
Train them
A good manager spends real time developing people instead of just handing off tasks and hoping. With AI that means using memory features, building out skills, and giving it the context it needs to do the job your way consistently, not just the one time you happened to explain it well.
Know when to escalate
A good manager has a sense for which work is too important, too sensitive, or too far outside someone’s wheelhouse to leave unsupervised, and steps in personally when it matters.
Wong’s story
I don’t work in sales, but my team is building AI automation for our sales group right now. The AI qualifies leads and follows up with customers when something’s missing from their info. Once the lead is fully qualified, it gets handed to an actual sales rep to close. The AI does the screening. A person still does the selling.
That’s really the whole job: doing the parts only you can do, and trusting the rest to someone, or something, you’ve actually trained well. Most of us were bad at managing people before we got decent at it. AI just gives you a much cheaper place to practice the same muscle.