Principle + Expert Voice
Draw a clear line for each kind of task: AI handles drafting, scaling, enumerating options, boilerplate; humans handle judgment, trade-offs, accountability, edge cases. The manager's job isn't the all-or-nothing "should we use AI"—it's designing this dividing line for each type of work: what goes to the machine, and where a human must take over.
"The Centaur approach involves a strategic division of labor between human and machine, switching between AI and human tasks depending on which is better at what."
— Ethan Mollick, "Co-Intelligence" (2024)
Scene: From "All-or-Nothing" to "Draw the Line"
Context: Your team needs to write a design doc for a new service. A senior engineer insists "AI output can't be trusted, ban it entirely"; a junior pastes an entire AI-generated architecture section straight into the doc without a close read.
✗ Two extremes
Total ban: "No AI on design docs."—You give up exactly what it's best at: drafting and listing options.
Total handoff: A generated section goes in unverified—you've outsourced the very judgment that should stay human.
✓ Draw the line (centaur split)
"Here's the split: AI drafts the skeleton, lists the pros/cons of three storage options, generates the interface boilerplate; the human decides the architectural trade-offs, reviews edge cases and failure modes, and signs off on the final call."
Give the team one repeatable rule: "AI can draft; a human must sign." Make the line a default habit, not an argument every time.
Division-of-Labor Design Checklist
- In this task, which part is "generate/enumerate" (fit for AI), and which is "judge/trade-off/own" (must stay human)?
- Am I giving an all-or-nothing "use AI or not," or a real line of "which parts, where"?
- Have I explicitly ringed the high-stakes zones (security, money, data, external commitments) as "human must take over"?
- Can the team repeat this line in one sentence—or does it live only in my head?
Common Mistakes + Female Lens
- Substituting "attitude" for "design." "Embrace AI" is a slogan, not a line; without task-level specifics, everyone does their own thing.
- A frozen line. Models shift each quarter; work once "human-only" gets pushed across the line—re-draw it periodically.
- Splitting without aligning standards. If "what counts as a good AI draft" is unstated, humans clean up after it—slower, not faster.
This Week's Exercise + Reflection
Exercise: Pick one recurring task type. Split it into two columns—"generate" and "judge"—and explicitly write which column goes to AI and which stays human. Share it as the team's default.
Reflection: Is my team's current AI usage a division of labor I designed, or a set of private workarounds each person improvised?
Principle + Expert Voice
AI flattens the "can it even be built" barrier, and scarcity shifts to what it still can't do well: taste, problem definition, verification, cross-domain synthesis. Reskilling your team isn't teaching a few prompt tricks—it's pushing each person from "the keyboard end" toward "the judgment end." Execution is being commoditized; judgment is appreciating.
"Consultants using AI finished 12.2% more tasks, 25.1% faster, and produced 40% higher quality results—with the largest gains among the lower-performing consultants."
— Dell'Acqua et al., "Navigating the Jagged Technological Frontier" (Harvard/BCG, 2023)
Scene: Re-position the "Flattened" Value
Context: A senior engineer confides, anxious: "Now that Copilot exists, the juniors' code looks about as good as mine. My years of craft aren't worth anything."
✗ Empty reassurance
"Don't worry, you're still better than they are."—Neither specific nor credible. He knows you're brushing him off, and next time he won't tell you what he really thinks.
✓ Acknowledge + re-position the value
"You're right—the pure code-writing part really did get flattened. But where you're now more valuable has shifted: judging where this AI code breaks under high concurrency, turning a vague requirement into the right problem, and that 'smells wrong' taste in code review—AI still can't give you those."
"Next quarter I'm moving your time off the keyboard: lead architecture reviews with two juniors, run a postmortem. That's your moat."
Reskilling Execution Checklist
- Is my team's reskilling focus "better at the tool" or "better at judgment"? (The former gets flattened again in six months.)
- Have I freed up my senior people's time for the high-judgment work—verification, problem definition, mentoring?
- Are juniors using AI to skip the fundamentals? Their reps are being taken by AI—how do I make that up?
- Am I modeling "AI drafts, judgment gatekeeps"—or "dump it on AI and don't look"?
Common Mistakes + Female Lens
- Reskilling as attendance box-ticking. Assigning a prompt course to be watched—judgment isn't grown from videos, it's grown in real tasks.
- Upgrading the tool but not the role. People still do the flattened work, just faster—eventually the whole slice disappears.
- Ignoring the junior "practice paradox." AI does their foundational reps, so judgment is harder to grow.
Female Leader's Note
Multiple surveys show women report using generative AI at markedly lower rates than men—partly because women are scrutinized more harshly for mistakes and so are warier of "trial-and-error" use in public, tending to use it quietly and privately (what Ethan Mollick calls "secret cyborgs"). Countermeasure: managers should openly normalize AI use and write it into the formal workflow, making it an explicit team rule rather than an underground skill—once a tool goes underground, existing adoption gaps only widen.
This Week's Exercise + Reflection
Exercise: For each direct report, write one line: "the single judgment skill you should deepen most in the AI era" (architectural trade-offs / problem definition / verification / mentoring). At your next 1:1, agree on a real task to practice it.
Reflection: If AI flattens another half of my team's execution work next year, am I growing "skills that will get flattened again" or "judgment that's harder to flatten"?
Principle + Expert Voice
"Will AI replace me?" isn't melodrama—it's real fear. Empty reassurance ("you'll never be replaced") backfires: people don't believe it, and they learn to stop leveling with you. The honest move gives two things together: acknowledge which tasks will genuinely disappear, while pouring resources into concrete actions that help them transition. Panic is useless; a path is useful.
"This is the worst AI you will ever use."
— Ethan Mollick, "Co-Intelligence" (2024)
Scene: Acknowledge the Truth + Give a Path
Context: In a 1:1, a report says, "I've seen AI write tests and first-draft docs now—am I about to become useless?"
✗ Two failure modes
Denial: "Stop overthinking, your role is safe."—He doesn't believe it, and feels you won't tell him the truth; the door closes.
Panic without an exit: "Honestly, yeah, you'd better hustle."—Leaves anxiety, offers no direction.
✓ Honesty + a transition commitment
"Some work really will be automated—boilerplate tests, first-draft docs. That's true, I won't lie to you. But that doesn't mean you get replaced."
"Next quarter we move you toward what AI still can't do well: failure diagnosis on complex systems, mentoring a junior, leading a cross-team design. I'll give you the specific projects and timeline. Panic doesn't help; transition does—and I'll walk this path with you."
Honest-Anxiety Checklist
- Am I giving a real assessment, or the "you're safe" reassurance he can see through in a second?
- Have I separated "which tasks disappear" from "you as a person won't be replaced"?
- After acknowledging the fear, do I give a concrete transition path with projects and a timeline—or just "hang in there"?
- At the team level, have I openly discussed AI's changes, or let rumor ferment into panic?
Common Mistakes + Female Lens
- Dodging via silence. If you don't discuss it, the fear doesn't vanish—it goes underground into attrition and quiet quitting.
- Over-promising. "I guarantee no one will lose their job to AI"—the day the org actually cuts headcount, all your credibility is wiped at once.
- Only threat, no opportunity. AI also frees people to do higher-value work; talking only about displacement is one-sided fearmongering.
This Week's Exercise + Reflection
Exercise: Find the person on your team most likely to be anxious and open a conversation: say honestly which work will change, and together write down one concrete first step toward higher-judgment work.
Reflection: Last time I discussed AI's impact on the team, did I give reassuring truth—or a comforting falsehood that just made people go quiet?
Principle + Expert Voice
AI-augmented teams have new failure modes: hallucinations reaching production, over-trusting output, diffused accountability ("the AI wrote it" becomes an excuse). The "jagged frontier" means you can't tell by intuition where AI is reliable and where it quietly fails. Managers must build two things: a verification culture and clear ownership—AI is the tool; the byline and the accountability are always human.
"Be the human in the loop... You are responsible for the output. The AI is a tool, and you are the one who signs off on the work."
— Ethan Mollick, "Co-Intelligence" (2024)
Scene: Head Off "The AI Wrote It" Before It Becomes an Excuse
Context: In a PR, AI-generated code calls an API that doesn't exist (a hallucination). It passes review and reaches staging before anyone catches it.
✗ Two extremes
Excuse: "Oh, the AI got it wrong, not your fault."—Accountability outsourced, no one learns, same error next time.
Overcorrection: "From today, no AI for the whole team."—Abandoning all the gains over one failure.
✓ Set the rule + build verification
"Let's be clear: whoever submits, owns it—regardless of whether AI wrote it. AI drafting for you doesn't change that you're the signer."
"Two supporting mechanisms: (1) AI-generated critical code/conclusions get a human check before submission—especially whether external dependencies actually exist; (2) in high-stakes zones (security, money, data), AI output requires a second reviewer."
Human-in-the-Loop Verification Checklist
- Does the team share an explicit norm—"whoever submits owns it, independent of whether AI generated it"?
- Are high-stakes zones (security / money / data / external commitments) covered by a red line: "AI output must be human-verified / second-reviewed"?
- Does everyone know AI fabricates confidently (hallucination)? Are they wary of output that "looks right"?
- Am I encouraging labeling "this was AI-drafted," so reviewers know where to look twice?
- When I sign off on AI-drafted work myself, have I actually read it—or am I also passing "looks right"?
Common Mistakes + Female Lens
- Verification as bureaucracy. Not everything needs a second reviewer—spend the cost only on high-stakes zones; let low-stakes go, or the team routes around the process.
- Diffused accountability, no owner. "AI + three people touched it" = everyone can pass the buck when it breaks. Ownership must land on one named person.
- Trust rising with fluency. The smoother and more confident the AI output, the lazier people get about verifying—exactly when it's most dangerous.
Female Leader's Note
Blame for "the AI got it wrong" is often assigned unevenly: in murky, many-hands situations, women and underrepresented people are more likely to bear heavier consequences for the same error (the likability / competence double bind, extended into the AI era). Countermeasure: make "humans own AI output" a clear, situation-not-person process—which zones must be verified, who signs by name, verification records kept—so blame isn't assigned by impression and identity.
This Week's Exercise + Reflection
Exercise: Write a one-page "AI use & accountability agreement" for the team: sign-off rules, high-stakes red lines, verification requirements. Send it this week and walk it through at the next team meeting.
Reflection: If an AI hallucination quietly reached production next Monday, would today's process catch it—or could I only assign blame after the fact?
Going Deeper
Won't the "centaur split" fail as AI grows stronger—the line pushed further and further toward the human, until there's nowhere left to retreat?
The line is indeed moving, but "failure" is premature. Every wave of automation has pushed humans toward higher abstraction: from writing assembly to writing high-level languages, humans didn't disappear—they moved to judgment that's harder to specify. The point isn't holding a fixed line but migrating upward continuously—today humans own architectural trade-offs and accountability; tomorrow, maybe problem definition and value judgment. The real risk isn't "the line gets pushed to nothing," but a manager who never re-draws it, leaving the team stuck at a layer that will eventually be swallowed. Treat re-drawing the line as a quarterly routine, not a one-time setup.
If "judgment" is the scarce thing, but juniors grow judgment precisely by doing the foundational work AI now takes over—how do you solve this "practice paradox"?
This is the most real management dilemma of the AI era: AI does the juniors' foundational reps, so judgment is harder to grow. The fix isn't sending them back to hand-writing boilerplate, but changing how they practice: make juniors the reviewers and verifiers of AI output—read the AI-generated code, find where it breaks, explain why. Reviewing exposes you to the "good vs. bad" contrast faster than writing from scratch—exactly the nutrient of judgment. Pair it with clear feedback, and you migrate practice from a "production act" to a "judgment act."
Won't honestly saying "some work will disappear" actually create panic and accelerate attrition?
There's that risk, so honesty must come in pairs: in the same breath that names the threat, give the path. Saying "your work will disappear" alone is intimidation; "it will disappear + here's my concrete plan to help you transition" is leadership. People can bear bad news; what they can't bear is uncertainty plus silence. If you don't speak, the fear is still there—it just goes underground into quietly sending out résumés. Transparency actually reduces attrition: those who stay know you didn't deceive them, and know you're investing in their next step. What really accelerates attrition is reassurance that gets exposed as false.
Does "humans own AI output" still hold as AI grows more autonomous (agents executing multiple steps on their own)? How should accountability scale with the degree of automation?
The principle holds, but its anchor moves up with autonomy. When AI goes from "giving suggestions" to "executing multi-step actions on its own," humans can no longer review line by line—so accountability shifts from "review every output" to "design the guardrails and acceptance criteria": define what the agent may touch, when it must stop and find a human, how results get accepted. The analogy: you don't audit a report's every keystroke, but you're accountable for who you hire, what permissions you grant, how you evaluate. The higher the automation, the more human accountability retreats from the "operational layer" to the "policy-and-boundary layer"—what exits is execution; what can't exit is being accountable for setting those boundaries.