AI Doesn't Dream

When Bob Sternfels, McKinsey's Managing Partner, sat down with Harvard Business Review for McKinsey's centennial interview published this month, he shared that the firm now employs 60,000 people: 40,000 humans and 20,000 AI agents. And that number had grown from 3,000 agents in just eighteen months. ( https://hbr.org/2026/01/we-want-to-make-ourselves-better )

The financial press ran the lazy "AI displacement" angle. Business Insider went with: "CEO Bob Sternfels says McKinsey now has 60,000 employees: 25,000 of them are AI agents". The marketing and advertising trade press didn't run anything. The "AI displacement" angle requires zero original thinking. You do simple math, and write the doom headline. You don't have to understand consulting. You don't have to understand AI limitations. You don't have to understand business model transformation.

You just count things and predict job losses. Insert variables. Generate content. Collect clicks.

But it misses actual insight when something genuinely interesting is happening.

Every publication led with the count. Every headline focused on displacement. Every article missed what Philip K. Dick understood when he asked whether androids dream of electric sheep: the question isn't whether artificial intelligence exists, but whether it can aspire to something beyond its programming.

Sternfels just answered: It can't.

As, buried in the HBR interview, past the agent count that every publication breathlessly reported, Sternfels said this:

"One thing AI models don't do well is aspire; they're not good at setting the right level of aspiration. Great leaders set an aspiration and then get people to stretch."

Now this is interesting.

Last month I wrote that WPP hiring McKinsey for strategic advice revealed the problem with consulting-led transformation. "McKinsey can recommend structural changes, portfolio optimisation, and efficiency initiatives. But McKinsey can't change culture." I wrote.

I spent thirteen years at Publicis where we executed the industry's deepest transformation: The Power of One. It was underappreciated, even mocked, then, but it is respected now. We systematically dismantled the holding company model and rebuilt Publicis as an integrated company.

That transformation took three years of grinding cultural work. Every agency leader had to surrender P&L autonomy, walking away from the independent business empires they'd built over decades. Every creative director had to collaborate with former competitors, sharing talent and credit.

The resistance was fierce. Agency presidents fought to protect 'their' clients. Creative chiefs hoarded their best talent. Finance teams resisted shared service models that threatened their autonomy. We had to rebuild compensation structures, reporting relationships, and performance metrics, not just once, but repeatedly as we learned what worked.

McKinsey's blue-hued charts showing organisational restructuring couldn't have made that happen. Culture change can't be consulted into existence. You can't optimise your way to transformation.

But if McKinsey's own managing partner understands that aspiration, not optimisation, is what AI cannot replicate, maybe the WPP strategic review actually has a chance.

Because AI is commoditising analysis. The competitive advantage isn't in better presentations explaining current state. It's in the aspiration to reach different state and the judgment to get there.

If McKinsey is applying this insight to their own transformation, if they genuinely understand that aspiration cannot be automated, they might actually be the right partner for WPP's cultural transformation.

Not because they can optimise organisational charts (AI can do that). But because they understand transformation requires human aspiration that no model can generate.

Sternfels identifies specific limitations that matter more than his agent count. But to understand why these limitations are fundamental rather than temporary, you need to understand how large language models actually work.

"AI models don't do well at setting the right level of aspiration."

This isn't a temporary limitation that better models will solve. Large language models are trained on historical data of everything that's been written, created, and documented up to their training cutoff. They learn patterns from what exists. When you ask an AI to generate something, it's fundamentally predicting "given everything I've seen before, what's the most likely next word/idea/solution?"

That's interpolation, not extrapolation. It's finding patterns within known territory, not imagining beyond it.

Humans imagine beyond boundaries because we can conceive of states that have never existed. We can aspire to create categories that aren't in the training data. We can set stretch goals that have no historical precedent.

The breakthrough innovations, the ones that create new markets, transform categories and redefine customer expectations all require imaginative leaps that abandon linear extrapolation.

That's human work. No amount of computational power changes this fundamental limitation.

"There isn't truth in AI models; there isn't judgment. Humans need to impose those parameters." said Sternfels.

In a training manual, IBM wrote this:

"A computer can never be held accountable. Therefore a computer must never make a management decision."

And that was forty-five years ago, back when computers filled entire rooms and companies were just beginning to automate basic processes. The principle was simple: accountability requires agency. Systems that cannot be held responsible cannot be trusted with decisions that require responsibility.

Here's what's changed: The computers got smaller, faster, and vastly more capable. What hasn't changed: They still can't be held accountable. They still have no judgment. And they still cannot determine what you should do, only what you can do based on patterns in historical data.

Large language models have no objective function beyond "predict the next token correctly." They have no goals, no values, no strategic context. They don't want anything. They don't believe anything. They pattern-match at unprecedented scale and sophistication, but pattern-matching isn't judgment.

When you ask an AI for strategic advice, it's giving you the statistically most common answer given similar questions in its training data. It's extraordinary pattern recognition, not judgment.

AI can optimize tactics brilliantly. It can tell you what works based on correlations in historical data. It cannot tell you what you should do given your specific values, competitive position, and strategic intent.

Strategy requires understanding context that doesn't reduce to patterns: your specific competitive position, your organization's risk tolerance, your board's patience, your team's capabilities, your clients' unstated needs, your vision for where the market is heading.

AI tells you what most companies did. Judgment tells you what your company should do.

When an AI system makes a recommendation that fails, who's responsible? The vendor who built it? The data it trained on? The executive who approved it? The CMO who implemented it? This ambiguity isn't a bug, it's fundamental to the technology.

IBM understood in 1979 that decision-making authority requires accountability. You cannot delegate strategic decisions to systems that cannot be held accountable for outcomes. You can use them as sophisticated analysis tools. You cannot abdicate judgment to them.

AI tells you what most companies did. Judgment tells you what your company should do.

"AI models are great at a linear approach to problem-solving but not at making discontinuous leaps." opined Sternfels.

Here's the technical limitation: Large language models work by predicting "what comes next" based on probability distributions from their training data. Next word, next sentence, next idea. It's inherently sequential and incremental.

Even when AI seems creative, it's recombining elements from its training data in statistically probable ways. It's remixing, not inventing.

McKinsey is now "looking more at liberal arts majors, whom we had deprioritsed" because they need people who make conceptual leaps that abandon linear logic.

The classic example: Henry Ford said "If I had asked people what they wanted, they would have said faster horses."

An AI analysing transportation problems in 1900 would have recommended: better breeding programs, more efficient feed, improved horseshoes, faster training methods. Every next step would have been horse-optimisation because every historical data point involved horses.

Ford made a discontinuous leap: abandon horses entirely, mechanise transportation, mass-produce automobiles. No historical pattern predicted that solution. It required imagining a category that didn't exist.

AI can plot the next point on your current trajectory. Humans can choose an entirely different trajectory.

McKinsey is hiring for human capabilities that AI cannot replicate. They're deprioritising credentials (perfect academic marks) and prioritising resilience, collaboration skills, and creative thinking.

They're preparing for a future where AI handles optimisation and humans handle aspiration.

If you're a CMO or CEO reading these '20,000 agents' headlines and thinking "we need to deploy more AI agents faster!" you're optimising for the wrong metric. Please don’t get distracted.

When AI eliminates the tactical work, what remains is the strategic work that AI fundamentally cannot do: imagining futures that don't exist in the training data, judging what your specific team should pursue given context that doesn't reduce to patterns, and making discontinuous leaps that abandon linear extrapolation.

WPP's strategic review will reveal whether McKinsey understands this distinction in practice, or whether Sternfels's insights remain theoretical.

If the deliverable is organisational restructuring recommendations such as: consolidate these agencies, eliminate redundancies, implement shared service centres, deploy AI to reduce headcount, then McKinsey will have optimised but not transformed. They'll have delivered the same PowerPoint strategy approach that Sternfels says they're moving away from.

That would be the easy path. It's what boards expect from consultants. It generates clean charts showing cost savings and efficiency gains. It's defensible because "best practices" support it.

If the deliverable is aspiration for what WPP could become combined with cultural transformation roadmap, something that is messy, difficult, requiring years of grinding work that resists neat presentation, then McKinsey will have demonstrated they understand their own managing partner's insight.

WPP CEO Cindy Rose described the company's performance as "unacceptable." Revenue declining 5.5-6%. Two profit warnings in three months. Headcount down 7,000 to 104,000. Market cap deteriorating while competitors like Publicis pull ahead.

Optimisation would look like: make our current structure more efficient, cut costs deeper, consolidate overlapping agencies, deploy AI to reduce headcount further, report cost savings to the board, declare transformation complete.

Instead, aspiration would look like: reimagine what an integrated marketing services company could be in the AI era, build culture around that vision, transform over time with courage to make difficult choices that hurt short-term metrics, accept that the work takes years not quarters.

That's hard. It requires the human capabilities Sternfels described: aspiration about what's possible, judgment about strategic direction, discontinuous thinking about business model transformation.

The more powerful AI becomes at optimisation, the more valuable human aspiration becomes.

When everyone has access to the same optimisation capabilities, and they will, because technology standardises and commoditises, then competitive advantage shifts entirely to strategic imagination.

McKinsey's competitors aren't sleeping. BCG is deploying agents. PwC is deploying agents. Accenture is deploying agents. Deloitte is deploying agents. Every consulting firm has access to the same underlying AI technology.

If competitive advantage came from agent deployment, there would be no competitive advantage. The technology is available to everyone with capital to invest.

But aspiration about what's possible for clients? Judgment about strategic direction in moments of uncertainty? Discontinuous thinking that imagines business models that don't exist yet?

Those remain human domains. And they're more valuable now than ever precisely because AI has commoditised everything else.

The agents aren't replacing humans. They're eliminating the work that prevents humans from aspiring.

Not to replace aspiration. To free humans for the discontinuous thinking and strategic imagination that creates competitive advantage when everyone else has the same tools.

Brandflow is written by Justin Billingsley, who has spent his career on all three sides of the industry's table: senior client, global agency leader, technology founder. First published 19 January 2026 in the Brandflow newsletter on LinkedIn.