AI-Wiki

How AI Is Changing Strategy—and 4 Skills Leaders Need Next

8 min read

Mollick — How AI is changing strategy, and four skills leaders need next (HBR IdeaCast, Oct 2026)

▶ Watch on YouTube · Harvard Business Review · 32:38

AI is moving beyond productivity gains—and forcing leaders to rethink how their businesses actually work. In this episode of HBR IdeaCast, Wharton professor and AI expert Ethan Mollick explains why leaders need to move beyond automating existing tasks and start redesigning organizations around what AI makes possible. He discusses how AI agents are changing the role of managers, where humans can create competitive advantage when everyone has access to the same AI models, and why companies need to experiment rather than wait for someone else to figure out the future. He also shares four capabilities leaders should cultivate in an AI-driven world: deep knowledge, wide knowledge, taste, and agency.

— channel description, Harvard Business Review (excerpt)

TL;DR

A ~33-minute episode of HBR IdeaCast on the Harvard Business Review channel, published 9 October 2026. The guest is Ethan Mollick (Wharton); the lower-third calls him “author of Co-Existence”, and the host refers to “your new book”. The host is not named in the episode. It is a webcam conversation with no slides. The captions are auto-generated; quotes are cleaned for names only.

1. What changed: the curve, and agents (0:36–1:12). “The exponential curve continues”, and “agents are real”, AI tools that “go off and do work on their own autonomously.”

2. Productivity alone gets eaten by the organisation (1:12–2:15). “People are getting ROI from AI … The problem is … those productivity gains get eaten up by the organization” because “AI is mostly a single player entity and you end up with more of everything. So if you don’t change your KPIs people … produce more PowerPoint, they produce more code.” And a default of automation means not thinking “hard enough about where humans fall into this process.” See micro-productivity-trap.

3. Jobs blur; the organisation’s purpose comes into question (2:15–3:22). AI can “regularly do the kind of work that would take a human two to 18 weeks” (he later attributes the figure to Epoch). “Coders do design work, designers do coding, product managers do everything. … What was the point of the PowerPoint in the first place?”

4. Leadership, lab and crowd (3:22–5:05). Leadership decides how AI is used and incentivised, not just “use AI”. The crowd gets access to current models (inferior access makes people “deluded about what AI can and can’t do”). The lab is a dedicated team, “not primarily technical”, that benchmarks systems and builds “systems of the future”. Individually, leaders must be “willing to eat risk” and use the tools themselves.

5. From centaur and cyborg to conjurer (5:05–7:30). In a year AI went from about half an hour of work per instruction to “16 plus hours of work out of a single AI instruction.” So you assign rather than prompt: “a conjurer is really about conjuring a team into existence when you need it … almost like writing an RFP.” “AI skills look like management skills”; he cites an Anthropic study in which managers, not coders, did best at first-pass coding “because they do a good job specifying what they want.”

6. Twilight factories, not dark ones (7:30–8:47). A software dark factory (he names StrongDM: “No human could look at code. No human can write code”) is “a dangerous approach.” Instead, “twilight factories”, where “the AI proactively reaches out to people and brings them in” for their variance, expertise, sign-off or interest. Automation as default “removes all the differentiation … you’re going to have the same [Claude]-based company everyone else does.” See automation-vs-augmentation.

7. Leaders regain span of control (8:47–10:05). With “a swarm of a hundred agents” you cannot approve each action; you get reports. But leaders “take back control over their company because they can now directly ask the AI”: C-level leaders “diagnosed problems of plants that no one brought up to them.”

8. AI already makes strategic decisions (10:05–13:29). Any long-running agent work “involves tons of strategic decisions and judgment calls.” “The question for you is where do you want to make decisions and where not”: ask for five options with defences, or tell it to build and test a prototype. Because the systems are jagged, the limits are found by the lab’s experiments, not by rule: “Nobody knows anything … The vendors are all making this up. This is a three-year-old technology.”

9. Four capabilities (13:29–14:53).

  1. Deep knowledge — expertise lets you judge output at a glance: “You don’t have to read those 400 pages of AI content because you know right away.” Juniors lack it.
  2. Wide knowledge — knowing a little about many fields, to give AI better feedback as work crosses boundaries.
  3. Taste — choosing among options, including “company taste … What strategy feels like the kind of strategy our company would do?”
  4. Agency — “the willingness to take action under uncertainty”, the hardest to teach.

See durable-skills.

10. No AI natives; experience pays (14:53–17:18). “It doesn’t work the same way with AI” as with the internet; “because it rewards expertise … BCG consultants with less experience were worse at using AI.” Senior leaders “actually have the best tools available.” Get a second opinion from a paid frontier model (“Claude Fable 5.1 … and GPT-6” at recording) on everything.

11. Incumbents and startups (17:18–19:19). Boards that decide “we’ll ride this out” give “an insane answer. You have 100,000 people working for you.” Startups gain because AI covers everything they were bad at.

12. Cybersecurity: “apocalyptic” in the near term (19:19–22:52). Distinguish future unguarded models from current guarded ones. He reports OpenAI moving “20% of their staff off of research” to hardening and financial institutions moving “50% plus” of security staff (second-hand, unsourced). Much of the attack surface will come from benign agents: one asked to price a sweater “relentlessly attacked” a wool mill to get futures pricing.

13. Two strategic mistakes, four questions (22:52–25:25). Mistakes: anchoring on today’s capability (“everyone’s spending too much on tokens” will pass as costs fall), and trusting vendors or consultants: “if they have successful cases, it’s for GPT-4.” Build capacity in-house. Citing Tushman’s ambidexterity, he asks: What have you stopped doing because of AI? What impossible thing are you building? What are you building that doesn’t work yet? Do you have leadership, lab and crowd?

14. Culture decides whether use surfaces (25:25–26:15). With good structure, culture and rewards, people bring use cases forward. “If you’re not, then everyone’s hiding all their uses … because they’re scared if they tell you they’ll get fired.” See ai-knowledge-hiding.

15. Not a normal technology (26:15–29:00). Treating AI as an IT project with “a six-month RFP process” produces something “obsolete the moment you launch it”, without “the dynamic capability to change” (see dynamic-capabilities). What he got most wrong: agents were easier to coordinate than he expected, “100,000 agent swarms that are self-organizing”, because agents share goals and knowledge. “People do what they’re incentivized to do.”

16. Jobs and risks (29:00–32:20). On employment, “nobody does” know; junior hiring may be squeezed, while insourcing may add jobs inside firms. “The AI labs … are aiming for the job apocalypse.” He worries about automation over augmentation and people whose “only job is verifying the results of AI content.”

Dynamic-capabilities reading

  • digital-sensing/digital-mindset-crafting — leaders using frontier models themselves and resisting the belief that they “can’t change”; the four capabilities as the individual side of sensing.
  • digital-seizing/rapid-prototyping — the lab as a standing experimentation unit; “build something that doesn’t work yet”.
  • digital-transforming/redesigning-internal-structures — leadership/lab/crowd, blurred job boundaries, twilight factories, and a wider span of control.
  • strategic-renewal/organizational-culture — KPIs and incentives that reward outputs rather than volume, and a culture in which people disclose AI use.

Linked entities and concepts

Debates and supersession

  • Do the less experienced gain more or less from AI? Mollick: “BCG consultants with less experience were worse at using AI.” His own jagged-frontier study reports the largest quality gains for the bottom half of performers (+31% vs +11%). The two claims measure different things (skill at judging AI output vs uplift on a set task), but the episode states the first without the second. Chesky, in this batch, makes the amplifier claim too; see his page.
  • Second-hand figures. The 2–18 weeks of work (Epoch), the 16+ hours per instruction, the Anthropic managers-as-coders study, and the OpenAI and bank security staffing shares are cited from memory, without references.

What was actually ingested

The full 32:38 episode from YouTube’s auto-generated English captions, fetched with yt-dlp after the transcript panel failed to render. The Gemini stills scan (manifest, 0 stills) returned no visuals, and an eleven-frame probe confirmed webcam footage throughout; one extra frame was checked for the guest’s lower-third.

Relationships

Supports

Graph View

Backlinks