It’s Thursday afternoon and you’ve been in meetings since nine. The 1:1 that starts in ten minutes is with someone you like and haven’t actually thought about in a week. You pull up their name, skim the last Slack thread, and walk in cold. “So, how’s it going?” “Good. Busy.” Twenty-five minutes of read-outs you could’ve gotten from a standup, and then you both go back to the pile. Nobody decided to let that happen. It just happened, one overloaded week at a time.
That’s the real bottleneck in supporting your reports. It’s showing up prepared and present, week after week, for six or eight (… and no more? right??) humans, while your calendar’s a wall and every 1:1 is quietly losing to a fire somewhere else. When presence runs low, the conversation collapses into status.
The industry tells two stories about where AI fits here, and I think both are half-true. The first is the line every product pitch repeats: ✨ AI frees you up to spend more time on the humans ✨. That’s only true if the reclaimed hour goes to a person instead of another meeting, and in my experience, mostly it doesn’t. The second is the product wave: a dashboard, a knowledge graph of your engineer’s growth, a health score. That one mistakes a map of a person for the person.
I build AI tooling and manage engineers for a living, these days for the systems that deliver SNAP and Medicaid benefits, and before that leading engineering at Glitch. Delivery work and management keep handing me the same question: which parts get better with AI, and which just pick up new failure modes?
A manager’s bottleneck isn’t knowledge. It’s the time and attention to assemble that knowledge before a conversation instead of during it (I know we’ve all been prepping for a meeting in the last 5 minutes of the prior one). That’s the kind of work a model is good at!
The first is walking in with real context. Instead of “so, what have you been up to,” you’ve got a synthesis of what someone shipped (the PRs, the doc they wrote, the thing they flagged in the last three 1:1s) and a first-pass agenda.
The second is memory across months. One 1:1 is a single data point; six months of them is the shape of a career. Is this person stuck on the same edge they named in the spring? A summary across time surfaces a drift you’d both otherwise feel vaguely, and it’s what lets you prep a career conversation with specifics instead of vibes: pulling “three times you took an ambiguous problem and made it legible for the rest of us” out of six months of notes is tedious, and the first thing to go when you’re slammed.
None of that is the relationship! All of it is scaffolding that lets you show up as the version of yourself who remembers everything and prepared like it mattered. Used this way, AI doesn’t replace the manager. It buys back the attention the job runs on and lets you give more of it to the person directly in front of you.
The danger isn’t AI doing the prep. It’s AI creeping from the prep layer into the relationship layer, and that can be a very subtle creep.
Of course, this is where the AI companies have every incentive to talk you out of thinking about. Every vendor wants the same thing from you: pour it all in. Every 1:1 note, every offhand observation, every “keep an eye on this,”, slack message, notion page, fed to the model, because more data is the product they’re selling.
But a lot of what you know about your reports was told to you in confidence, or picked up in a moment that was never meant to be a record. “I’m going through something at home.” “I don’t think I want to be a manager after all.” “I’m worried I’m falling behind.” That isn’t data to be warehoused. It’s trust, and trust is a currency that is very quickly spent and hard to earn back.
So your reports should know what you are and aren’t feeding into a model about them. “I keep private notes on how you’re doing” and “I pipe our 1:1s into a system I don’t fully control” are very different relationships, and pretending otherwise is a betrayal you’re choosing not to look at. Using AI for prep feels okay, but make sure you have a conversation with your report about the boundaries.
Maybe you both feel comfortable letting the AI note-taker be in these personal and sometimes difficult spaces, or maybe your report wants to keep something from the robots in the case of the looming robot apocalypse.
AI can help you see the shape of someone’s growth over time: the trend line, the plateau, the drift, the thing they keep circling. But only a person can care whether it’s the right shape for this particular person, whether the direction they’re growing is the one they want and not just the one that’s legible to the org.
I had someone at Nava who’d hit their ceiling with us, not for lack of talent but because of what the work was. Every engineer on those government contracts was as much a project manager as an engineer: hundreds of tasks to keep straight, far more writing and talking than coding. That wasn’t the direction they wanted to grow, which was deep and technical. We spent a few weeks on it, landed on the honest answer that no upcoming contract had a real spot for that, and I helped them look. They found a great role somewhere else. We’re still in touch, we’ve worked through gnarly technical spots together since, and they’re the person I message when I have a Rust question. No dashboard tells you to help your best-fit-on-paper engineer leave; that came from knowing them.
A model can tell you someone is growing. It can’t want good things for them. That wanting, and acting on it, is the job we signed up for as managers.