AI-Wiki

The CEO's Agenda for Thriving in the Agentic Age

8 min read

Catlin, Durth, Hämäläinen & Montard (McKinsey) — the CEO’s agenda for the agentic age

▶ Watch on YouTube · McKinsey & Company · 53:30

Generative and agentic AI are moving from experimentation to enterprise-wide transformation. For CEOs the challenge is no longer simply where to deploy AI, but how to rethink strategy, technology, and organization to create lasting competitive advantage. Productivity gains may be the starting point—but they are not the full opportunity. In this episode of Inside the Strategy Room, McKinsey experts Tanguy Catlin, Sandra Durth, Lari Hämäläinen, and Antoine Montard examine why AI represents a general-purpose technology with far-reaching implications for business. They explore the strategic choices facing CEOs, from reimagining workflows and operating models to navigating workforce transitions and building new organizational capabilities. The discussion offers a perspective on what it takes to move beyond incremental efficiency improvements and capture AI’s long-term value.

TL;DR

The 300th episode of Inside the Strategy Room on the McKinsey & Company channel, 53 minutes, host Sean Brown. Guests: Tanguy Catlin (senior partner, director of the McKinsey Global Institute), Sandra Durth (partner, people and organisation), Lari Hämäläinen (senior partner, leads gen AI and agents in QuantumBlack) and Antoine Montard (capability director, strategy). Six claims:

  1. Productivity is where you lose. “Right now, most organizations are obsessively focused on productivity … that’s where you lose the battle, not where you win the battle,” since competitors have the same models (Catlin). Montard calls it table stakes.
  2. Three waves. Productivity, then innovation (new products and business models, where “growth and value creation will be the most acute”), then lower transaction costs, which threaten any business built on aggregation, brokering or intermediation.
  3. The scale is unprecedented. About $7 trillion of US capital expenditure from 2026 to 2030, five times the rest of the G6 and China combined. Data centres go from about 2% of US electricity last year to 10–13% by 2030.
  4. The return so far is not from enterprises. Hämäläinen says gen AI moved the internet’s backbone (recommendations and so on), then programming, now half of workloads and all of it in the past six months. “Will enterprise AI be the next step function?” is open.
  5. Deploying tools gets 20 cents on the dollar. Tool deployment gives about 20% productivity; redesign gives multiples of that (two-week sprints compressed to one day). But “most organizations have not codified those workflows. It’s all in the head of the employees” (Catlin).
  6. Who controls your opex? With 50,000 employees and 500,000 agents three years out, in “a viable scenario,” agents could cost twice the payroll. “Somebody else would control your opex and I don’t think that’s an acceptable scenario” (Hämäläinen). Technology architecture becomes a CEO decision.

Key claims

Models

Compute going into training grows four to five times a year and algorithmic efficiency three times a year, and fewer models get trained. The performance gap between OpenAI, Anthropic and Google models “has halved in the last year” on his account. Differences come down to things like “tonality”, so choosing one is “almost like hiring an employee.” Almost any benchmark has been passed.

The developer reversal

Hämäläinen names no study. He says that from the organisations he works with, developers were benchmarked and productivity “actually declined by 20%” against an expected gain. The same group, measured this February, showed +20%, a reversal he puts at nine months, and those running the study “had real trouble” because developers would not work without AI. His attribution is a new model: task length went from 20 seconds (2023) to 1–2 minutes, then 20–30 minutes, and about 10 hours this year, crossing about five hours last autumn. “Nobody actually knows” why that threshold mattered. His lesson: when your organisation says “we tried this, it didn’t work,” try again. Catlin objects that there was a second driver: “the system is adjusting.”

History

Montard runs through steam (real gains came with urban factories and locomotives), electricity (factory redesign, then light bulbs and refrigeration), Solow’s 1987 line on computers, and e-commerce, which took about 20 years to reach 10% of retail. Technology arrives quickly, and economic impact takes “quite a bit of time and a lot of innovation.”

A disagreement inside the episode

Montard cites Keynes’s prediction of a 15-hour week: productivity lowers costs, demand expands, jobs follow. Hämäläinen pushes back: “There might be a more difficult future ahead because development is not linear … If you’ve been a call center agent for 30 years, are you the one who’s going to drive innovation?” Durth adds that the leading academics range “from black to white” because they assume different substitutability, compute costs, redeployment speed and government responses.

Work and organisation

  • 30%. Durth: 30% of current work could be done more economically by technology, spread across “everything” rather than whole jobs, and citing Harvard, 60% of managers spend over half their time on automatable tasks.
  • No-regret moves. “Investing into human capabilities”: AI fluency, while “not giving up on building expertise and institutional knowledge.” “I don’t think you teach AI, you do AI.”
  • Wrong things, faster. Automating existing processes makes organisations “wonderfully efficient at doing the wrong things.”
  • Design for adaptability. Today’s organisations were designed for command, control and predictability. Fluid organisations capture value more easily, though she has not seen it at scale; some tech firms redeployed 70% of their workforce within nine months.
  • Power users and skunk works. About 1% of users consume far more AI and find the new ways of working. Hämäläinen would put about 20 of them in a new team to build an AI-native process, “skunk working a new organization.”
  • Beyond employees. Catlin says productivity gains may first cut outsourced work and captive centres.
  • Judgment. Catlin cites Agrawal on cheap prediction raising the value of data and judgment, and Rigobon’s EPOCH framework (empathy, networking, visioning).

Closing advice

Durth: “Don’t wait on your no regrets and your learning moves.” Montard: be ambidextrous across time scales. Hämäläinen: make the technology architecture the CEO’s decision, and run bold breakaway experiments. Catlin: bring the team to learn together, because “it’s hard to imagine tomorrow differentiating the role of the CHRO and the head of technology.”

Dynamic-capabilities reading

  • digital-sensing/digital-scenario-planning: the three waves, the historical analogies and the open question of an enterprise step function.
  • digital-seizing/balancing-digital-portfolios: not depending on one provider, and choosing frontier models or others against an agent cost base.
  • digital-transforming/redesigning-internal-structures: end-to-end workflow redesign, flatter organisations, skunk works and new roles for overseeing agents.
  • digital-transforming/improving-digital-maturity: AI fluency for all staff as a no-regret move, and learning by doing.
  • strategic-renewal/business-model: lower transaction costs undermining aggregation and intermediation, and new AI-enabled products.
  • contextual/external-triggers: capital expenditure, energy demand and geopolitics as the frame CEOs read.

How it connects

What was actually ingested

The full episode (53:30, length_seconds: 3210) from the auto-generated track via the yt-dlp fallback, 192 lines. The captions garble names, corrected here against the description: Lauri Haapalahti, Larry and Lory are Lari Hämäläinen; Agarwal is taken as Ajay Agrawal (University of Toronto) and Rigabone as Roberto Rigobon (MIT). No stills: a frame probe showed one static title card throughout.

Linked entities and concepts

Dangling (single-source mention, deferred): Sean Brown, Sandra Durth, Antoine Montard, Lari Hämäläinen and Tanguy Catlin (two sources each, never author:), QuantumBlack, Ajay Agrawal, Roberto Rigobon.

Scope and reliability

A consultancy’s podcast addressed to the CEOs it sells to, closing on advice that matches what it sells. The figures come without sources. Two are shaky inside the episode: Catlin refers to “this 75%” of the workforce after Durth has said 30%, with no 75% stated anywhere; and Montard dates Keynes to 1920, while the essay usually cited is from 1930. Much of the value is that partners disagree on the record, on whether productivity creates jobs and on what drove the developer reversal.

Relationships

Supports

Graph View

Backlinks