Context
Molly Graham spent 20 plus years in leadership roles at Google, Facebook, Quip, and the Chan Zuckerberg Initiative, and 13 years ago she wrote a widely read essay called "give away your Legos," advising people to hand off their projects and responsibilities as their company scaled instead of clinging to them. In this conversation with Lenny Rachitsky, she revisits that advice for an AI era and admits parts of it no longer hold cleanly. She also shares fresh survey data on burnout and happiness in tech, and lays out what she now thinks people should refuse to delegate to AI. The episode matters to PMs and product leaders because most of them are already managing AI systems on top of their teams, without a clear model for what that kind of delegation should look like or cost.
The Big Idea
Giving away your work to a robot is not the same as giving it away to a person, because you can automate the doing but you cannot automate the oversight, so the psychological weight of ownership stays with you even as your output grows.
Graham argues this changes her old advice in a specific way: still give away everything you can, but recognize that some things (judgment, trust, taste, accountability) were never really "Legos" to give away at all, they were always yours to keep.
Key Insights
Delegating to AI still leaves you holding oversight
Graham's original "give away your Legos" advice meant fully handing a project to someone else and mentally letting go of it. Delegating to AI does not work the same way: you get more done, but you can never fully release the thing, because you are still the one who has to catch the AI's mistakes and stand behind the final output. She compares it to a Lego tower where a small robot keeps wandering through and knocking pieces over, and you have to keep saying "no, that's not what I said" the way you would coach a new intern. The practical result is that people who feel like they are producing more than ever are often also carrying more mental tax than ever, since oversight cost never drops to zero.
Treat AI as a bad intern, not a genius hire
The current narrative, that a company just hired the smartest employee it has ever had and should pour all its knowledge into it, sets people up to fail, according to Graham. Her replacement framing: AI needs context, onboarding, and correction the way a junior, often lazy intern does, and you would never take an intern's first draft and forward it straight to your boss. She ties this directly to the AI slop problem, arguing that slop happens when people treat AI output as already trustworthy and skip the iteration and editing they would automatically apply to a human's rough draft.
Burnout jumped from 44% to 55% in one year
Lenny's annual tech survey found burnout rose 10 percentage points between 2025 and 2026, while roughly half of respondents also reported feeling the happiest they have ever been in their careers. Graham reads this as two real things happening at once: emotional exhaustion from constantly re-learning how to work (one colleague at OpenAI told her it took six months to swing from "use AI everywhere" to "wait, is this actually working"), and a genuine sense of opportunity for the people who feel amplified by the tools. The strongest predictor of happiness in the survey was people who said AI had amplified their own work; the least happy group was designers, who described watching "everyone's a designer now" once AI made passable design accessible to anyone.
Smaller teams and strong managers predict happiness
The survey found people on smaller teams, and people with more authority over their own work, reported meaningfully higher happiness than people at larger companies. Graham connects this to a separate finding: the single strongest lever companies actually have over employee happiness is the quality of a person's direct manager, more than any AI tool or policy. She warns that the current trend of stripping out management layers to cut cost is a mistake she expects to backfire, because a good manager is what makes people feel seen during a period of rapid, disorienting change, and there is no data suggesting management matters less now.
AI branded layoffs mostly are not about AI
Graham says she has "a lot of beef" with layoffs publicly attributed to AI, arguing most of them are really companies that over-hired admitting the mistake under a more palatable label. She contrasts this manufactured fear narrative (your replacement is trained and coming for your job in six months) with the actual evidence available right now, which does not show AI broadly eliminating roles. She points to engineering as the visible precedent: the job has completely changed in two years, from writing code to orchestrating agents, yet demand for engineers has gone up, not down.
Reframe "will my job disappear" as "how will it change"
Graham cites a conversation with journalist Manoush Zomorodi, who has spent 30 years in an industry repeatedly declared dead, as the source of a reframe she now uses constantly: instead of asking whether your job will go away, ask what you would do if you believed it would always exist but look completely different every few years. Applied to journalism, that meant the shift from a newsroom paycheck to every journalist effectively running their own business, not the job vanishing. Graham thinks the same question applies to product management, design, and engineering, none of which have fixed, permanent definitions to begin with.
Mental Models & Frameworks
The centaur and the reverse centaur
Borrowed from writer Cory Doctorow, a centaur is a human head directing an animal body, in this case a human directing AI tools toward goals they chose. A reverse centaur flips it: the AI (or algorithm) directs a human body, the way gig-economy dispatch systems already route Uber and DoorDash workers. Use it as a design check on any AI-enabled workflow: if the human is setting direction and the AI is executing, that is a centaur; if the AI is setting the pace and the human is just filling in the gaps it assigns, that is a reverse centaur, and Graham and Lenny both treat that as the outcome to actively avoid in knowledge work.
The human sandwich
A framework Lenny raises for structuring where AI belongs in a workflow: the human sets the direction and the vision at the top ("here's where we want to go, here's the idea"), AI does the bulk execution in the middle, and the human reviews and iterates at the end before anything ships. The point of the sandwich shape is that both ends stay human even as the middle gets automated, so direction-setting and final judgment never get outsourced along with the grunt work.
Which Legos to keep versus give away
Graham's updated version of her original framework: still give away everything that is repeatable execution, because oversight cost is the price of getting more done and it is worth paying. But hold back anything that requires judgment you cannot yet define ("if you don't know what good looks like, you can't hand that to a summer intern"), anything that requires trust or is fundamental to a relationship, and anything where you are the one who has to set the vision for what the end result should even be. She frames the test as: could this be outsourced to a competent but inexperienced intern, and would you actually want it to be?
Trade-offs & Nuance
Productivity is up, but efficiency may not be
Graham and Lenny both cite an engineering research finding that AI coding tools have made developers measurably more productive by output, while the amount of code that later has to be rewritten has risen roughly eightfold, alongside a rise in security incidents. Graham's framing: productivity asks "can you generate more," efficiency asks "did that actually move things forward," and right now teams are optimizing hard for the first while the second quietly gets worse. She points to companies literally tracking token-usage leaderboards as a modern version of counting lines of code or slack messages, a metric she considers a known-bad way to manage people.
Lean into change, but grief is real and needs space
Graham's core message, that you should lean into change rather than resist it, still stands, but she now insists on pairing it with permission to grieve what is being lost. She describes a conversation with an engineer who missed the flow state of hands-on coding and now spends his time reviewing agent output instead, and argues leaders should let people mourn a role they loved (she references entrepreneur Chip Conley's idea that some changes deserve a literal funeral) rather than only pushing them toward excitement about what comes next. Both things, grief and opportunity, are true at the same time, and treating only one as acceptable to express makes the transition harder, not easier.
Common Mistakes
Mistake: shipping AI output without owning it
Graham describes CEOs presenting strategy memos that were visibly written by AI, without editing or real engagement, as a direct example of leaders role modeling that it is fine to outsource thinking and skip accountability. The harm compounds because whoever receives that unowned output then has to do the work of catching what is wrong with it, a hidden tax on the rest of the organization. Her fix is simple: treat every AI draft the way you would treat a first pass from a new intern, expect to edit it, and never forward it without having actually engaged with whether it is good.
Mistake: assuming AI can be managed like a senior hire
Several people in the conversation describe frustration managing "a swarm of agents" that constantly ping for approval or produce low-quality first attempts, the same fatigue that comes from managing genuinely junior human employees rather than experienced ones. Graham's point is that people expected AI delegation to feel like handing a project to a strong senior engineer who runs with it; instead it behaves like onboarding several interns at once, which requires far more oversight bandwidth than most people budgeted for, and that mismatch between expectation and reality is a real driver of the burnout in the survey data.
Practical Application
Ask "can AI help with this" before starting
Lenny describes deliberately building the habit of pausing before any task to ask whether AI could help with it, comparing it to the gap between stimulus and response that meditation practice builds. The payoff compounds as tools improve: people who have built the reflex of checking in with AI first are positioned to take advantage of every capability jump, while people who default to doing things the old way by habit fall behind even though the tools were available to them the whole time.
Write down which Legos you will not give away
Before delegating more work to AI, make an explicit list of what should stay human: work you do not yet know how to define "good" for, work that depends on trust or an ongoing relationship, and work where you are setting the vision rather than executing against someone else's. Graham's test is whether you would hand the task to a real but inexperienced summer intern; if the honest answer is no, that is a signal it should stay with you regardless of what the AI is technically capable of producing.
Name the grief before asking for the excitement
If you manage people through a period of AI-driven change, explicitly acknowledge what is being lost (a role someone loved, a way of working that felt good) before pivoting the conversation to opportunity. Graham suggests normalizing this the same way her original Legos essay normalized the discomfort of scale: telling people plainly that what they are feeling is real and shared, rather than only offering reassurance, is itself what makes people more willing to lean into the change afterward.
Protect management instead of cutting it
Given that manager quality was the survey's strongest lever on happiness, resist the instinct to flatten management layers purely for cost or headcount efficiency during an AI transition. Graham argues the cost of removing managers shows up later, in disengagement and burnout that is hard to trace back to the decision, so treat manager capacity, including their time to actually check in on people, as a protected resource rather than the easiest place to cut.
Questions to Consider
- Which parts of my own role would I genuinely be uncomfortable handing to a capable but inexperienced intern, and am I currently letting AI take on any of those anyway?
- Where on my team have I skipped editing or engaging with an AI draft before passing it along, the way Molly Graham describes CEOs forwarding AI-written strategy memos unedited?
- Is my current AI workflow a centaur, where I set the direction and the tool executes, or a reverse centaur, where I am mostly reacting to what the tool has already produced?
- Have I made space for people on my team to grieve a part of their job that AI has changed, or have I only offered them reassurance about the upside?
- Am I treating token usage, output volume, or similar activity metrics as a proxy for real progress, the way Graham describes companies tracking token leaderboards?
Bottom Line
Give away everything you can to AI, but recognize that judgment, trust, vision, and accountability were never Legos to give away in the first place, they were always the human's job, and the biggest mistake right now is treating AI as a genius replacement instead of a junior intern that still needs your oversight.