Very insightful. My one quick comment (also as a former partner at McKinsey and Bain :)) is that consultants are also paid for making trust incentive-compatible. Repeat business, references, and the firm’s reputation put future rents at risk—an implicit bond that AI does not yet replicate.
The issue of trust is studied in a recent paper by Piotr Dworczak and Alex Smolin, “Robust Trust” (2026): https://arxiv.org/abs/2602.09490
[UPDATE: And I see Garett already commented! So I'll add… language-use context, it seems, will remain the hardest informational dimension for AI to capture. Even in a “purely private” corridor that elicits true preferences.]
TBF this was several years ago. I get the sense the current admin operates, um…differently.
Now retired, but for decades the formal and informal institutions around the NatSec Council's PCC process stayed pretty stable and surprisingly similar to what you described.
I speak on the topic of AI in Finance, and routinely come back to this concept of trust. AI does not have a life to lose, nor a reputation or mortgage etc. Humans do. As long as humans exist in organizations, humans will need to work with them.
AI is still very bad at persuading. It can't communicate in the modes you need to communicate effectively. Like it can't keep an email chain going over a few days with multiple parties, it can't participate effectively in a Zoom call with many people, it can't have a "quick chat" to answer questions, it can only really answer slowly at length, it can't stay focused on any issue over a timeframe of weeks or months.
Maybe this will change eventually, but for now, the humans have to do all this stuff.
Enjoyed this and strongly support the knowledge is distributed angle.
One thing to add though—tasks are neither atomic nor a stable unit of analysis. It’s not just that tasks are hard to disentangle from the job, more that a task doesn’t exist as a stable object prior to the doing of it.
‘Task’ in Hutchins’ sense (see Cognition in the Wild) is constituted by the assemblage—the people, the artifacts, the setting, the history of prior attempts—that carries it out. Change the assemblage and you’ve made a new task, rather than automated the same task with new tools.
Thanks! Agreed about tasks- but even more about Hutchins sense of cognition. T restate the key point from Hutchins closer to the spirit of the piece, cognition is not just a "cognitive" act (If i may put it like this) but the result of the organizational system; technology matters because the whole system changes. Thanks for pushing here.
This closely reflects what I see in cross-border market-entry projects. The analysis is often the relatively straightforward part. The harder work begins when headquarters, local stakeholders, professional advisers and potential partners interpret the same proposal differently, hold different information and have different levels of authority. In practice, progress depends on translating strategic intent into something the relevant parties can accept, authorize and actually execute. That combination of analysis, translation and intermediation is where human consultants continue to create value.
Great topic and I just had this discussion in my life recently. Funnily enough taking 100% perspective, but your post is making me leaning rather in the direction that agents can probably take over more than I first anticipated.
Specifically the argument I saw that agents do not have any issue to learn (maybe with harness or further post training alignment) to perform well in games such as diplomancy or other standard nash equilibrium problems. Noise can for sure make it more difficult with sparse data but the only moat humans then would have is a ticking timer.
I think that the reinforcement learning is the crucial one. Agents *can* indeed calculate many moves ahead. But they need ot know the game to learn to play it. My point here is not that they cannot do nash equilibrium; they clearly can. The point is that playing Nash requires knowing the rules and the payoffs. When teh game starts, they really have no idea. My argument on the consultants is that all of that emerges from the interactions between the players-- in all of those meetings new knowledge emerges and is created.
Great article. I wonder if it's possible to teach AI agents that people in organizations (whether a company or other non-economic organization) don't actually want things to be optimally defined, because they want the power to define what is and is not optimal. But only about a small subset of all of the issues that need to be worked out. That's why they argue.
I'm picturing the meetings you talk about in the article, and if there are 8-9 people in that room, there are at least 3-4 who want no part of it and will say whatever is necessary to get the meeting over and done with. They don't care about that issue, whatever that meeting was, but are very interested in other issues. How would an AI agent understand that if it's constantly working things out 'in the moment' because it has no memory?
Thanks Jess. There is stuff "in context" , all the conversations in the context window, and there is some "in-context learning" as a result, but I agree, probably not enough for this.
Also, ST just above pointed me to your work, but all i could find was a missing tweet in a broken MR link. Could you send me to what you have written on this point (workers making org capital)?
Actually it never went beyond tweets. Arnold Kling mentioned my key tweet many times, and it helped inspire his book and article, Patterns of Sustainable Specialization and Trade.
Here's the link to the article in Phelps's journal, my tweet is formally mentioned in a footnote, first time I had that happen:
I came to my view that most work is building organizational capital partly from the old Barro-Sala-i-Martin finding that alpha ~ 2/3 in growth regressions, from Brynjolfsson and Hitt's work on intangibles, Prescott's paper Organization Capital, and from summer jobs at various companies. The tooth-to-tail ratio looks pretty small in lots of modern, highly productive organizations.
In my opinion—and, above all, based on my own experience—your analysis is excellent.
Many so-called white-collar jobs are not just about analysis and deliverables. In many organizations, analysis is only one part of the work. Coordinating different stakeholders, understanding competing interests, navigating internal agendas, reading the company culture, and leveraging tacit knowledge can be far more important.
In fact, I would go even further: in many cases—perhaps even in the majority—the action that will eventually be proposed is already largely decided. What is needed is an analysis that helps validate, reinforce, or articulate a belief that has been formed through tacit knowledge and experience. Of course, this is debatable, and certainly not always the case.
However, I think the real impact of AI is not primarily in answering the question: How will AI affect existing jobs?
The more interesting question is:
How will AI change the very concept of the company?
My thesis is that AI may not simply be something that existing companies “adopt.” The bigger risk is that companies built around traditional organizational structures will be displaced through creative destruction by companies designed from the ground up around AI.
And that could fundamentally challenge the idea of hierarchy itself.
Imagine a company worth $1 billion consisting of a board of directors and 100 AI agents reporting to it. Reporting, internal memos, layers of management, and much of today's organizational bureaucracy could become relics of the past.
Perhaps the companies of the future will have, at most, two human layers: decision-makers and everyone else replaced by networks of agents and sub-agents.
There may be no interns climbing the corporate ladder—only people responsible for direction, judgment, accountability, and capital allocation.
Will companies with 40,000 employees still exist? Probably. In the same way that small traditional farms still exist today. But they may increasingly have to compete directly with companies of 10 people generating €40 billion in revenue.
The most radical possibility is that agents themselves could become economic entities with measurable track records, reputations, and even independent market value—not simply because of the technology behind them, but because of the experience they accumulate.
I'd argue that at the premium segment of the Consultant market (where Mck/BCG and a few others reside), the perceived value lies in executive validation rather than analysis or operational execution. In other words, exec management needs a trusted and prestigious 3rd party to rubber stamp a decision that they themselves should be responsible for.
It will be very interesting to see how the sector is impacted by AI given the potentially commodification of the analytical component.
That's certainly one commonly touted aspect of the value prop. However, having worked inside the premium segment, I would say this analysis is very on point. A lot of the consultant's work is mediation between stakeholders, each with conflicting views. Another aspect, which AI can likely replicate, is offering a third party "objective" lens (in reality quite subjective, but I would liken it to how the family behaves differently when a friend is over for dinner).
Very insightful. My one quick comment (also as a former partner at McKinsey and Bain :)) is that consultants are also paid for making trust incentive-compatible. Repeat business, references, and the firm’s reputation put future rents at risk—an implicit bond that AI does not yet replicate.
The issue of trust is studied in a recent paper by Piotr Dworczak and Alex Smolin, “Robust Trust” (2026): https://arxiv.org/abs/2602.09490
Thanks Kamierz, I agree. I will check it out.
[UPDATE: And I see Garett already commented! So I'll add… language-use context, it seems, will remain the hardest informational dimension for AI to capture. Even in a “purely private” corridor that elicits true preferences.]
Really tracks two things it reminded me of:
Garett Jones's claim most of what the modern workforce produces is “organizational capital” https://marginalrevolution.com/marginalrevolution/2009/11/the-wisdom-of-garett-jones.html
Selling the loyalty plan sounds almost exactly like what it's like day-to-day working in the White House
Love the point! I would love to hear what the day to day in the WH is like! Maybe you can call it a governatorial office and tell us about it!
TBF this was several years ago. I get the sense the current admin operates, um…differently.
Now retired, but for decades the formal and informal institutions around the NatSec Council's PCC process stayed pretty stable and surprisingly similar to what you described.
I speak on the topic of AI in Finance, and routinely come back to this concept of trust. AI does not have a life to lose, nor a reputation or mortgage etc. Humans do. As long as humans exist in organizations, humans will need to work with them.
AI is still very bad at persuading. It can't communicate in the modes you need to communicate effectively. Like it can't keep an email chain going over a few days with multiple parties, it can't participate effectively in a Zoom call with many people, it can't have a "quick chat" to answer questions, it can only really answer slowly at length, it can't stay focused on any issue over a timeframe of weeks or months.
Maybe this will change eventually, but for now, the humans have to do all this stuff.
Enjoyed this and strongly support the knowledge is distributed angle.
One thing to add though—tasks are neither atomic nor a stable unit of analysis. It’s not just that tasks are hard to disentangle from the job, more that a task doesn’t exist as a stable object prior to the doing of it.
‘Task’ in Hutchins’ sense (see Cognition in the Wild) is constituted by the assemblage—the people, the artifacts, the setting, the history of prior attempts—that carries it out. Change the assemblage and you’ve made a new task, rather than automated the same task with new tools.
Thanks! Agreed about tasks- but even more about Hutchins sense of cognition. T restate the key point from Hutchins closer to the spirit of the piece, cognition is not just a "cognitive" act (If i may put it like this) but the result of the organizational system; technology matters because the whole system changes. Thanks for pushing here.
Very much enjoyed it!
This closely reflects what I see in cross-border market-entry projects. The analysis is often the relatively straightforward part. The harder work begins when headquarters, local stakeholders, professional advisers and potential partners interpret the same proposal differently, hold different information and have different levels of authority. In practice, progress depends on translating strategic intent into something the relevant parties can accept, authorize and actually execute. That combination of analysis, translation and intermediation is where human consultants continue to create value.
Great topic and I just had this discussion in my life recently. Funnily enough taking 100% perspective, but your post is making me leaning rather in the direction that agents can probably take over more than I first anticipated.
Specifically the argument I saw that agents do not have any issue to learn (maybe with harness or further post training alignment) to perform well in games such as diplomancy or other standard nash equilibrium problems. Noise can for sure make it more difficult with sparse data but the only moat humans then would have is a ticking timer.
Much appreciated with the article in any case
Thanks David. Happy you enjoyed it.
I think that the reinforcement learning is the crucial one. Agents *can* indeed calculate many moves ahead. But they need ot know the game to learn to play it. My point here is not that they cannot do nash equilibrium; they clearly can. The point is that playing Nash requires knowing the rules and the payoffs. When teh game starts, they really have no idea. My argument on the consultants is that all of that emerges from the interactions between the players-- in all of those meetings new knowledge emerges and is created.
Your perspective*
Great article. I wonder if it's possible to teach AI agents that people in organizations (whether a company or other non-economic organization) don't actually want things to be optimally defined, because they want the power to define what is and is not optimal. But only about a small subset of all of the issues that need to be worked out. That's why they argue.
I'm picturing the meetings you talk about in the article, and if there are 8-9 people in that room, there are at least 3-4 who want no part of it and will say whatever is necessary to get the meeting over and done with. They don't care about that issue, whatever that meeting was, but are very interested in other issues. How would an AI agent understand that if it's constantly working things out 'in the moment' because it has no memory?
Thanks Jess. There is stuff "in context" , all the conversations in the context window, and there is some "in-context learning" as a result, but I agree, probably not enough for this.
Excellent post, tiny typo toward the end:
“What is missing is a cheap, portable data set that has not be strategically distorted”
Oops! Corrected thanks!
Also, ST just above pointed me to your work, but all i could find was a missing tweet in a broken MR link. Could you send me to what you have written on this point (workers making org capital)?
Actually it never went beyond tweets. Arnold Kling mentioned my key tweet many times, and it helped inspire his book and article, Patterns of Sustainable Specialization and Trade.
Here's the link to the article in Phelps's journal, my tweet is formally mentioned in a footnote, first time I had that happen:
https://scholar.google.com/scholar?cluster=17865056913866073106&hl=en&as_sdt=0,21
I came to my view that most work is building organizational capital partly from the old Barro-Sala-i-Martin finding that alpha ~ 2/3 in growth regressions, from Brynjolfsson and Hitt's work on intangibles, Prescott's paper Organization Capital, and from summer jobs at various companies. The tooth-to-tail ratio looks pretty small in lots of modern, highly productive organizations.
Glad to offer more if useful!
Wow! this is massive impact from an old style (hence short!) tweet, unprecedented? But thanks, I agree, and will follow up.
In my opinion—and, above all, based on my own experience—your analysis is excellent.
Many so-called white-collar jobs are not just about analysis and deliverables. In many organizations, analysis is only one part of the work. Coordinating different stakeholders, understanding competing interests, navigating internal agendas, reading the company culture, and leveraging tacit knowledge can be far more important.
In fact, I would go even further: in many cases—perhaps even in the majority—the action that will eventually be proposed is already largely decided. What is needed is an analysis that helps validate, reinforce, or articulate a belief that has been formed through tacit knowledge and experience. Of course, this is debatable, and certainly not always the case.
However, I think the real impact of AI is not primarily in answering the question: How will AI affect existing jobs?
The more interesting question is:
How will AI change the very concept of the company?
My thesis is that AI may not simply be something that existing companies “adopt.” The bigger risk is that companies built around traditional organizational structures will be displaced through creative destruction by companies designed from the ground up around AI.
And that could fundamentally challenge the idea of hierarchy itself.
Imagine a company worth $1 billion consisting of a board of directors and 100 AI agents reporting to it. Reporting, internal memos, layers of management, and much of today's organizational bureaucracy could become relics of the past.
Perhaps the companies of the future will have, at most, two human layers: decision-makers and everyone else replaced by networks of agents and sub-agents.
There may be no interns climbing the corporate ladder—only people responsible for direction, judgment, accountability, and capital allocation.
Will companies with 40,000 employees still exist? Probably. In the same way that small traditional farms still exist today. But they may increasingly have to compete directly with companies of 10 people generating €40 billion in revenue.
The most radical possibility is that agents themselves could become economic entities with measurable track records, reputations, and even independent market value—not simply because of the technology behind them, but because of the experience they accumulate.
That, to me, is the much bigger story.
AI may not just transform jobs.
It may transform the company itself.
I'd argue that at the premium segment of the Consultant market (where Mck/BCG and a few others reside), the perceived value lies in executive validation rather than analysis or operational execution. In other words, exec management needs a trusted and prestigious 3rd party to rubber stamp a decision that they themselves should be responsible for.
It will be very interesting to see how the sector is impacted by AI given the potentially commodification of the analytical component.
That's certainly one commonly touted aspect of the value prop. However, having worked inside the premium segment, I would say this analysis is very on point. A lot of the consultant's work is mediation between stakeholders, each with conflicting views. Another aspect, which AI can likely replicate, is offering a third party "objective" lens (in reality quite subjective, but I would liken it to how the family behaves differently when a friend is over for dinner).