And for those who might be curious, in this earlier piece (further down in the middle part of the essay) I compare the SpaceX approach versus NASA when it comes to space exploration.
What your whole piece implies thus is the companies are themselves a bottleneck for results they are time clocks are set for human speeds for ai work utilization first they need to be redesigned grounds up in a special new way. The hard part is it only means one thing reduce clog and manpower for decision making implying first downside redesign then redeploy a company structure designed around ai speeds, implying 24 hrs working, faster decision making, rapido finance approvals and finance remodeling, rapid testing and rapid collecting data and recaliberating.
Very true, but I also think that for a traditional company, their competition is not a tech giant's clock speed. Rather, it is their peer-group competitor. I like to think that if an operating company is willing to invest in people to think systematically, it could stay slightly ahead and then it is winning. Thank you for pushing our thinking collectively.
My two cents: I have been thinking about this topic for some time now in the context of my own organization. What follows is where my thinking is today. It may well change as I learn more, so apologies in advance for the long comment.
This is a thought-provoking piece. I agree with the core warning: simply adding AI to existing processes will mostly speed up local activity without necessarily improving company-level outcomes. Faster code, memos, tickets, and decks do not automatically create customer value.
As the saying goes: “Every system is perfectly designed to get the results it gets.” If an organization is slow, bloated, over-governed, under-accountable, and full of low-value work, AI will not fix that on its own. The current system is producing those outcomes. The system itself has to be redesigned.
But I do not think the starting point should be “how do we redesign workflows?” I think it should be more basic: why does the organization exist, what customer and societal value is it accountable for, what risks must it control, and what operating model should sit around that?
For many established companies, especially in regulated industries, the “company of one” or ultra-lean AI-native firm is a useful provocation, but not a complete model. Banks, insurers, healthcare companies, manufacturers, telecoms, utilities, airlines, and public companies cannot simply optimize for speed or headcount reduction. They also have obligations around auditability, regulatory reporting, cybersecurity, safety, privacy, segregation of duties, model risk, operational resilience, board oversight, and customer protection.
So, to me, the real redesign unit is not the workflow. It is the organization as an accountability system.
Before deciding where AI agents go, leaders should probably ask some first-principles questions:
Why does this organization exist?
Which customers are we choosing to serve, and which are we not?
What do those customers actually value, based on behavior rather than slogans?
What regulatory and governance obligations are non-negotiable?
What risks are we willing to take, and what risks must we never take?
What capabilities must remain inside the firm?
What can be automated fully?
What should be automated with a human in the loop?
What should remain human-led, with AI support?
What should remain deliberately slow or human-only?
Who is accountable when an AI-enabled process fails?
What evidence must exist for auditors, regulators, customers, and the board?
This matters because workflow optimization can easily become cost-cutting in disguise. Many organizations may well have more labor than they would need in a better-designed operating model. But that does not mean the people are useless. More often than not, the system has created unnecessary work: approvals, handoffs, reports, reconciliations, committees, controls, exceptions, meetings, and manual bridges between broken systems.
In other words, many companies are overstaffed relative to the value they deliver, but understaffed relative to the complexity they have created.
That distinction matters. If leaders cut people before removing the work, they create brittleness, burnout, quality problems, and eventually rehiring. If they remove the work first, the organization can become smaller, clearer, and more capable.
The Nokia example is powerful because Nokia’s problem was not just slow software. It was slow learning and a broken organizational adaptation. The 48-hour compile time was a symptom of a deeper operating model issue. In the same way, today’s token budgets and AI copilots are symptoms, not strategy. They may increase individual throughput while the company remains constrained by governance, testing, approvals, legacy data, compliance interpretation, and decision rights.
The question should not be, “How do we make every department faster?” It should be, “What prevents validated customer value from flowing safely through the institution?”
The word “safely” matters. In many industries, some friction is not waste. Some friction protects fairness, safety, privacy, customer rights, explainability, resilience, and accountability. The point is not to eliminate all human review. The point is to distinguish between friction that protects value and friction that protects bureaucracy.
A practical AI transformation might classify work into four categories:
1. Fully automated — repeatable, rules-based, low-discretion tasks with embedded controls and audit trails.
2. Automated with human review — AI does much of the work, but humans review exceptions, approve consequential decisions, and remain accountable.
3. Human-led, AI-supported — judgment-heavy areas such as strategy, regulatory interpretation, crisis response, ethical tradeoffs, high-impact customer decisions, and board matters.
4. Human-only or deliberately constrained — areas where legal, reputational, safety, or dignity considerations require strong human authority.
This requires more than process mapping. It means redesigning decision rights, accountability, data architecture, controls, escalation paths, assurance, and governance.
The data layer is especially important. Many firms cannot responsibly automate because their data is fragmented, poorly governed, inconsistently defined, or not auditable. If sales, finance, risk, and operations all define “customer” or “exposure” differently, AI will not create a learning organization. It will create faster confusion.
There is also a political economy problem inside the firm. Every efficiency program quietly raises the question: who captures the gains?
Customers may get better service. Shareholders may get higher margins. Executives may get bonuses. AI vendors may capture more spend. Employees may get redeployed or eliminated. If the distribution of gains is not addressed honestly, employees will rationally resist. The people who know most about how to simplify the organization are often the same people whose roles are threatened by simplification.
So AI transformation needs a credible internal bargain. If employees help remove low-value work, are they rewarded, redeployed, reskilled, given mobility, or cut? If managers shrink bureaucracy, are they promoted for impact or punished for having fewer people? If all gains flow upward, complexity will be defended because complexity protects jobs, budgets, and status.
This is also why “customer value” alone is not enough. Companies say they are customer-centric, but most operate through internal proxies: departmental metrics, quarterly targets, risk avoidance, executive narratives, and political incentives. Support optimizes handle time. Sales optimizes bookings. Finance optimizes budget control. Engineering optimizes velocity. Legal optimizes risk reduction. Each metric may be rational locally, while the customer experience worsens overall.
A deeper design question is: what set of metrics prevents local optimization? A serious operating model should balance customer outcomes, quality, risk, cost, speed, learning, employee capability, resilience, and regulatory compliance. Otherwise, AI helps each function game its own scoreboard faster.
My takeaway would be this:
Do not design AI around the existing org chart. Design the organization around accountable value creation under explicit constraints. Then decide which work belongs to humans, which belongs to machines, and which must be governed jointly.
For regulated incumbents, the future is probably not “no one works here.” It is more likely to mean fewer manual processors, fewer coordinators, fewer reporting layers, more embedded controls, more data stewards, more model-risk capability, more customer outcome owners, stronger assurance, clearer accountability, and smaller teams with much greater leverage.
That is not just workflow redesign. It is a new theory of the firm in the age of AI.
The hard question for incumbents is:
If we were founding this organization today, with current AI capabilities and current regulatory obligations, what would we build, and what inherited work, roles, rules, and structures would we refuse to recreate?
The biggest question is: how many companies will actually be willing to reinvent themselves before they are forced to?
Nokia is a useful reminder here. Reinvention is easy to praise after the fact, but incredibly hard to do from the inside. It means challenging existing power structures, incentives, operating models, product assumptions, talent models, and even the identity of the company.
Most organizations will probably do the safer thing first: add AI to the current system, automate pieces of work, reduce some costs, and declare progress. Far fewer will ask the harder question: if we were building this company from scratch today, with current AI capabilities and current regulatory obligations, what would we refuse to recreate?
That is where the real transformation is. Not in adopting tools, but in having the courage to redesign the system that the tools are being placed into.
I think that's the reason we have human agency. Leadership is not merely to preside over the inevitable, but to change the course of history itself. Thank you for sharing this greater aspiration.
Every organization I have worked in has lacked sufficient capacity to pursue everything important. There are always deferred initiatives, underexplored customer needs, data problems no one has time to fix, products that could be improved, and moonshots that stay on the shelf because they are too expensive or time-consuming.
If AI increases a team’s intelligence and productivity, the first question should not be, “How many people can we remove?” It should be, “What new value can we create with this leverage?” That value should be considered in a broader order: society first, then customers, employees, and stakeholders. The real opportunity is not just to make the current business more efficient, but to build something useful, responsible, and durable, something that can outlast the individuals and leaders who create it.
As the saying goes, “The best way to predict the future is to create it.” AI gives organizations a chance to do exactly that by not optimizing yesterday’s business, but creating a better version of tomorrow’s.
That requires moving away from short-term thinking. Human resources should not be viewed only as a cost to reduce, but as a source of capability, imagination, judgment, and institutional knowledge that can help transform the industry in which the organization operates.
When an employee leaves, even if they are not a superstar, they often take tacit knowledge with them: process understanding, operational context, informal networks, and trusted relationships. Of course, persistent underperformance has to be addressed, but leaders should not assume AI can simply replace what walks out the door. In its current form, AI can assist, document, and accelerate work, but it cannot fully replace lived context, judgment, and human trust.
AI should shift the conversation from pure efficiency to possibility: new products, services, business models, customer experiences, controls, or operating models that were previously out of reach.
That does not mean the talent mix stays the same. If the nature of work changes, skills need to change too. Some roles may shrink or disappear, while new ones emerge. Leaders may need fewer people doing repetitive coordination or manual processing, and more people who can define problems, work with data, manage risk, design better customer experiences, govern AI-enabled processes, and pursue new opportunities.
So, I think that the goal should not be simple workforce reduction. It should be workforce reallocation and reinvention.
And if we focus too much on the cost cutting, this is where we will end up:
Ford had to rehire 350 engineers after its AI got vehicle quality wrong
The automaker's VP of hardware engineering says Ford mistakenly believed AI could replace experienced workers, a blunder it fixed just in time to top JD Power's quality ranking for the first time in 16 years
Thank you so much for giving so much thought to the piece, and for your critical views. I especially like the point that the one-person company concept is often a provocation rather than a direct label in many highly regulated industries. In that regard, the goal is not to become a one-person company, but relative to other bloated peer groups, you are one inch ahead in satisfying all the different non-negotiables.
What I really appreciate in your comment is that you highlight the multiple dimensions people need to think about systematically. It is not a single dimension, such as just thinking about the customer. Other companies cannot do that.
Thank you again for sharing your work in progress and your current thinking. This is how we get the truth out as the landscape continues to evolve. I am sure some of your input will serve as inspiration for my next piece in the future.
I spent some time looking at model routers and concluded that the cost differential between a subscription and API based usage was still too much. And the economics of context caching dramatically raises the cost of model switching for long-running work.
I can imagine applications backed by a model router where each session is small and well-bounded. And I see a lot of potential there for subagents and agent teams. But for long-running high-complexity sessions the cost differential remains huge.
I know, thank you for giving me additional thoughts as the situation evolves so rapidly. Have to keep thinking about and staying close to the phenomena.
Thanks. Amazing insight. It looks a lot like a typical lean project. Instead of improving processing speed or operation time (value-adding time), concentrate on eliminating non-value-added activities. Do the right thing before doing the thing right. The second insight is the bottleneck one, which again is often found in manufacturing activities. It does not make sense to improve a machine's OEE if the next machine is a bottleneck in the process. The transition to an AI agent world may well deliver a future where non-value-adding activities have been minimized or reduced to almost zero.
For that to happen, the concept of value stream would be important to emphasize and implement.
It is lean project all over again, which I find comforting because there is a dominant narrative saying AI is changing everything and all the old rules do not apply. As I go through more and more of the deeper thinking, we actually know a lot about how to cope with the upcoming challenge.
There are the lean startup methodologies, and there is the re-engineering playbook. It is just that we need to apply them much more broadly in context, not just living within a small group or within software engineering. It is the same philosophy, applied deeply across the entire organization or enterprise.
Thank you so much for the spot-on observation. I'm glad that piece has been helpful.
Thanks for the great historical account. I wasn't aware of the details behind the Nokia story, which goes alongside The Mythical Man-Month as a cautionary and instructive tale.
I buy the overall concept -- in chemistry this is analogous to the rate-limiting step of reaction kinetics. For an AI-centric business, though, I'm afraid we still lack transparency on cost to benefit. I'm a research scientist using a $20/mo Anthropic subscription, and Claude Code tells me I used an equivalent of $180 in tokens yesterday to accomplish some pretty routine code organization (much of which involved cleaning up its own mess). Its worked great for parts of this project *AND* I 100% could not afford to pay full freight here.
This is partly why Cursor is a cautionary tale: they're downstream of the model providers, and now Anthropic is under-cutting them on cost to a similar product by a factor of ~10x. Building a business (or a work process) on something you don't own or control seems like a big liability. And yes,
I can imagine a future where model compute becomes a commodity and prices drop, but that comes with it's own set of challenges.
There's a learning curve to new tools and tech, and I think we're still early days for AI use *optimization*. But we're also facing huge uncertainties regarding the funding and financing of model providers that adds existential risk to an over-reliance on some of these tools.
I like it. Thank you so much for pointing it out, almost from the ground up, with first‑hand observation.
Quite honestly, it feels like the push to deploy AI everywhere is driven by a pricing policy that U.S. Frontier Labs could have changed overnight. The discussion about token expense wouldn’t exist if they hadn’t switched from a monthly seat to token billing. But the broader point you describe is spot on. Thank you again.
And, critically, their switch to token pricing wasn't accidental. It was driven by IPO demands of opening their books, showing revenue to cover the mountains of debt. In this regard, I think Ed Zitron is spot-on in his financial analysis of the so-called frontier labs.
In my limited testing, Anthropic retains a clear lead on Chinese open-source models in terms of the models themselves and the overall usability of the ecosystem. Claude Code is a great product, and Opus is solid (and Fable was blocked by the government -- a whole separate cautionary tale there!). But OpenAI and Anthropic prices are being heavily subsidized by investors, and those costs are not sustainable.
I think this is a sharp reminder of what drives this: price switching, which, for some reason, was less discussed among executives in my classes or in the media. As you said, the blocking of Fable warrants another extensive essay for executives worldwide to reduce reliance on a single service provider.
The token pricing or "correct model for the job" conversation is being addressed by OpenRouter. Your AI disruption article explains this very clearly Howard.
Per the "No One Works Here" thing...I was just reading a short blurb about "1 person companies" in China. I forget the actual number, but it was in the hundreds of thousands...maybe millions...whatever, it was a huge number...of one person companies utilizing AI, mostly DeepSeek, that have been formed in China in the last 12 months.
Couldn't find the piece, but a quick DeepSeek query here....
"One-person companies (OPCs) in China have moved from a niche concept to a national strategy, fueled by AI and new regulations . By the end of 2025, nearly 7.32 million new OPCs were registered—a 42.3% year-on-year increase .
What Defines an OPC?
· Sole Shareholder, Not Sole Employee: An OPC has only one shareholder (natural or legal person) , but can hire staff and appoint directors .
· Limited Liability: The core advantage. You are only liable up to your capital contribution, protecting personal assets—unlike sole proprietorships which have unlimited liability .
· New Flexibility: Revised Company Law allows a natural person to own multiple OPCs .
Why AI is the Catalyst
AI dramatically lowers startup costs. Entrepreneurs now use AI tools to handle design, operations, and marketing tasks that once required a whole team . This "individual plus supply chain" model is unique to China, allowing solo founders to tap into dense manufacturing ecosystems .
Government Backing & Support
Local governments are building full-stack support systems to attract OPC talent :
· Subsidized Computing Power: Shanghai's Pudong offers OPCs free computing power worth up to 300,000 yuan ($43,876) .
· Dedicated Hubs: Specialized parks like Beijing's Zhongguancun AI North Latitude Hub provide shared workspaces, business registration services, and investor matchmaking .
⚠️ Crucial Legal & Financial Risks
This model is not without pitfalls; the "limited liability" shield can be pierced:
· Asset Separation: The sole shareholder must prove personal and company assets are separate. If not, they bear unlimited joint liability for company debts .
· Strict Accounting: Maintain separate bank accounts, keep meticulous records, and complete a mandatory annual audit. Avoid using personal accounts for business .
OPC vs. Sole Proprietorship (个体工商户)
One-Person Company (OPC) Sole Proprietorship
Legal Status Corporate legal entity Natural person (no legal entity)
Liability Limited liability (to investment amount) Unlimited liability (with all personal assets)
Bottom Line: An OPC is a powerful, low-risk vehicle for AI-savvy founders. However, the administrative and legal discipline required is high—formal accounting and strict financial separation are mandatory to maintain liability protection .
If you are considering this path, would you like to know more about the specific registration process or capital requirements?
It was very recently in an article and I just glanced through it because it caught my eye. Your post made me think of it. It was just a couple days ago. It might have been in Caixin. I'll see if I can find it again.
If you offered me $40k in free computing power I'd register an OPC also!
Thanks for this -- I think this also speaks to the ongoing evolution of CCP's business and economic policies. Bankruptcy laws, LLCs and financing long played a role in Western capitalism. China now has a long track record of remixing successful Western business approaches for their own economic development while carefully controlling political expression.
Observing companies' behavior becomes especially intriguing because China has many copycats, intense competition, and often only one national champion.
I often times couldn't help but think about the managerial principles and leadership that emerge, regardless of location, and how they evolve into a national champion, whether in the US, Europe, China, or Southeast Asia. To me, this reveals a universal truth. Thanks again for reading and keep pushing our thinking along the way.
不客气 It's amazing to watch. In the last couple decades, the business development here is breathtaking. The integration of local government and business is profound. Things go from concept to operational in a few months.
Thanks for the additional reference. In Europe, one‑person companies are mostly discussed as a concept and a management practice. I had no idea they are so widely operationalized in China, where they are almost the norm. The speed of adoption there is remarkable.
I'm in Wuhan and briefly talked about it with a friend, then I saw the piece...I'm pretty sure it was in Caixin....wherever it was, and just now your piece reminded me of it. it really struck me. The tech/supply chain/mfg/distribution stack here is so insanely robust, it makes total sense. One person can be a corporate level enterprise.
It‘s important not to mix up everything. A software workflow is not a business workflow and both are not governance workflows. Not everything needs the same speed. And not every tech or business workflow can be treated the same. It’s a difference if you build something new, if you integrate something, if you process ad hoc tasks or if you maintain something. You need a meaningful cadence for workflows that are connected.
Regarding shipping software: long before AI there was a concept called DevOps. It‘s worth to research that, because how it was implemented in some companies misses the concept entirely. It was long before AI possible to ship to production the same day you developed the code. With a proper DevOps Setup.
Regarding business workflows: the root cause for problems lies ALWAYS in the set up of the teams. The right mix of roles and accountability solves 80% of all problems.
Governance can be slowlier. The root cause is that topics are treated as governance topics when they are actually daily business. Root cause is wrong setup of roles and accountability.
The big art is to set up several teams and leadership levels in a way that it all runs smoothly.
This is what I do for a living, besides other topics.
It starts with the mindset and it is a never ending work.
The goal is to become a learning organization with a cadence of learning that is not rushed and is sustainable.
Before we confront AI opportunities and challenges, DevOps serves as a playbook along the way. Because of the page limit, I couldn’t discuss The Phoenix Project, but I love how you concluded.
The key is to build a learning organization. Deploying an agent doesn’t justify shortcutting learning for human controllers when setting goals and reengineering processes end‑to‑end across silos.
Very insightful. Reminds of how Elon Musks thinks and operates.
Yes, glad that you picked it up.
And for those who might be curious, in this earlier piece (further down in the middle part of the essay) I compare the SpaceX approach versus NASA when it comes to space exploration.
https://howardyu.substack.com/p/the-reactor-is-the-easy-part
What your whole piece implies thus is the companies are themselves a bottleneck for results they are time clocks are set for human speeds for ai work utilization first they need to be redesigned grounds up in a special new way. The hard part is it only means one thing reduce clog and manpower for decision making implying first downside redesign then redeploy a company structure designed around ai speeds, implying 24 hrs working, faster decision making, rapido finance approvals and finance remodeling, rapid testing and rapid collecting data and recaliberating.
Very true, but I also think that for a traditional company, their competition is not a tech giant's clock speed. Rather, it is their peer-group competitor. I like to think that if an operating company is willing to invest in people to think systematically, it could stay slightly ahead and then it is winning. Thank you for pushing our thinking collectively.
My two cents: I have been thinking about this topic for some time now in the context of my own organization. What follows is where my thinking is today. It may well change as I learn more, so apologies in advance for the long comment.
This is a thought-provoking piece. I agree with the core warning: simply adding AI to existing processes will mostly speed up local activity without necessarily improving company-level outcomes. Faster code, memos, tickets, and decks do not automatically create customer value.
As the saying goes: “Every system is perfectly designed to get the results it gets.” If an organization is slow, bloated, over-governed, under-accountable, and full of low-value work, AI will not fix that on its own. The current system is producing those outcomes. The system itself has to be redesigned.
But I do not think the starting point should be “how do we redesign workflows?” I think it should be more basic: why does the organization exist, what customer and societal value is it accountable for, what risks must it control, and what operating model should sit around that?
For many established companies, especially in regulated industries, the “company of one” or ultra-lean AI-native firm is a useful provocation, but not a complete model. Banks, insurers, healthcare companies, manufacturers, telecoms, utilities, airlines, and public companies cannot simply optimize for speed or headcount reduction. They also have obligations around auditability, regulatory reporting, cybersecurity, safety, privacy, segregation of duties, model risk, operational resilience, board oversight, and customer protection.
So, to me, the real redesign unit is not the workflow. It is the organization as an accountability system.
Before deciding where AI agents go, leaders should probably ask some first-principles questions:
Why does this organization exist?
Which customers are we choosing to serve, and which are we not?
What do those customers actually value, based on behavior rather than slogans?
What regulatory and governance obligations are non-negotiable?
What risks are we willing to take, and what risks must we never take?
What capabilities must remain inside the firm?
What can be automated fully?
What should be automated with a human in the loop?
What should remain human-led, with AI support?
What should remain deliberately slow or human-only?
Who is accountable when an AI-enabled process fails?
What evidence must exist for auditors, regulators, customers, and the board?
This matters because workflow optimization can easily become cost-cutting in disguise. Many organizations may well have more labor than they would need in a better-designed operating model. But that does not mean the people are useless. More often than not, the system has created unnecessary work: approvals, handoffs, reports, reconciliations, committees, controls, exceptions, meetings, and manual bridges between broken systems.
In other words, many companies are overstaffed relative to the value they deliver, but understaffed relative to the complexity they have created.
That distinction matters. If leaders cut people before removing the work, they create brittleness, burnout, quality problems, and eventually rehiring. If they remove the work first, the organization can become smaller, clearer, and more capable.
The Nokia example is powerful because Nokia’s problem was not just slow software. It was slow learning and a broken organizational adaptation. The 48-hour compile time was a symptom of a deeper operating model issue. In the same way, today’s token budgets and AI copilots are symptoms, not strategy. They may increase individual throughput while the company remains constrained by governance, testing, approvals, legacy data, compliance interpretation, and decision rights.
The question should not be, “How do we make every department faster?” It should be, “What prevents validated customer value from flowing safely through the institution?”
The word “safely” matters. In many industries, some friction is not waste. Some friction protects fairness, safety, privacy, customer rights, explainability, resilience, and accountability. The point is not to eliminate all human review. The point is to distinguish between friction that protects value and friction that protects bureaucracy.
A practical AI transformation might classify work into four categories:
1. Fully automated — repeatable, rules-based, low-discretion tasks with embedded controls and audit trails.
2. Automated with human review — AI does much of the work, but humans review exceptions, approve consequential decisions, and remain accountable.
3. Human-led, AI-supported — judgment-heavy areas such as strategy, regulatory interpretation, crisis response, ethical tradeoffs, high-impact customer decisions, and board matters.
4. Human-only or deliberately constrained — areas where legal, reputational, safety, or dignity considerations require strong human authority.
This requires more than process mapping. It means redesigning decision rights, accountability, data architecture, controls, escalation paths, assurance, and governance.
The data layer is especially important. Many firms cannot responsibly automate because their data is fragmented, poorly governed, inconsistently defined, or not auditable. If sales, finance, risk, and operations all define “customer” or “exposure” differently, AI will not create a learning organization. It will create faster confusion.
There is also a political economy problem inside the firm. Every efficiency program quietly raises the question: who captures the gains?
Customers may get better service. Shareholders may get higher margins. Executives may get bonuses. AI vendors may capture more spend. Employees may get redeployed or eliminated. If the distribution of gains is not addressed honestly, employees will rationally resist. The people who know most about how to simplify the organization are often the same people whose roles are threatened by simplification.
So AI transformation needs a credible internal bargain. If employees help remove low-value work, are they rewarded, redeployed, reskilled, given mobility, or cut? If managers shrink bureaucracy, are they promoted for impact or punished for having fewer people? If all gains flow upward, complexity will be defended because complexity protects jobs, budgets, and status.
This is also why “customer value” alone is not enough. Companies say they are customer-centric, but most operate through internal proxies: departmental metrics, quarterly targets, risk avoidance, executive narratives, and political incentives. Support optimizes handle time. Sales optimizes bookings. Finance optimizes budget control. Engineering optimizes velocity. Legal optimizes risk reduction. Each metric may be rational locally, while the customer experience worsens overall.
A deeper design question is: what set of metrics prevents local optimization? A serious operating model should balance customer outcomes, quality, risk, cost, speed, learning, employee capability, resilience, and regulatory compliance. Otherwise, AI helps each function game its own scoreboard faster.
My takeaway would be this:
Do not design AI around the existing org chart. Design the organization around accountable value creation under explicit constraints. Then decide which work belongs to humans, which belongs to machines, and which must be governed jointly.
For regulated incumbents, the future is probably not “no one works here.” It is more likely to mean fewer manual processors, fewer coordinators, fewer reporting layers, more embedded controls, more data stewards, more model-risk capability, more customer outcome owners, stronger assurance, clearer accountability, and smaller teams with much greater leverage.
That is not just workflow redesign. It is a new theory of the firm in the age of AI.
The hard question for incumbents is:
If we were founding this organization today, with current AI capabilities and current regulatory obligations, what would we build, and what inherited work, roles, rules, and structures would we refuse to recreate?
The biggest question is: how many companies will actually be willing to reinvent themselves before they are forced to?
Nokia is a useful reminder here. Reinvention is easy to praise after the fact, but incredibly hard to do from the inside. It means challenging existing power structures, incentives, operating models, product assumptions, talent models, and even the identity of the company.
Most organizations will probably do the safer thing first: add AI to the current system, automate pieces of work, reduce some costs, and declare progress. Far fewer will ask the harder question: if we were building this company from scratch today, with current AI capabilities and current regulatory obligations, what would we refuse to recreate?
That is where the real transformation is. Not in adopting tools, but in having the courage to redesign the system that the tools are being placed into.
I think that's the reason we have human agency. Leadership is not merely to preside over the inevitable, but to change the course of history itself. Thank you for sharing this greater aspiration.
Every organization I have worked in has lacked sufficient capacity to pursue everything important. There are always deferred initiatives, underexplored customer needs, data problems no one has time to fix, products that could be improved, and moonshots that stay on the shelf because they are too expensive or time-consuming.
If AI increases a team’s intelligence and productivity, the first question should not be, “How many people can we remove?” It should be, “What new value can we create with this leverage?” That value should be considered in a broader order: society first, then customers, employees, and stakeholders. The real opportunity is not just to make the current business more efficient, but to build something useful, responsible, and durable, something that can outlast the individuals and leaders who create it.
As the saying goes, “The best way to predict the future is to create it.” AI gives organizations a chance to do exactly that by not optimizing yesterday’s business, but creating a better version of tomorrow’s.
That requires moving away from short-term thinking. Human resources should not be viewed only as a cost to reduce, but as a source of capability, imagination, judgment, and institutional knowledge that can help transform the industry in which the organization operates.
When an employee leaves, even if they are not a superstar, they often take tacit knowledge with them: process understanding, operational context, informal networks, and trusted relationships. Of course, persistent underperformance has to be addressed, but leaders should not assume AI can simply replace what walks out the door. In its current form, AI can assist, document, and accelerate work, but it cannot fully replace lived context, judgment, and human trust.
AI should shift the conversation from pure efficiency to possibility: new products, services, business models, customer experiences, controls, or operating models that were previously out of reach.
That does not mean the talent mix stays the same. If the nature of work changes, skills need to change too. Some roles may shrink or disappear, while new ones emerge. Leaders may need fewer people doing repetitive coordination or manual processing, and more people who can define problems, work with data, manage risk, design better customer experiences, govern AI-enabled processes, and pursue new opportunities.
So, I think that the goal should not be simple workforce reduction. It should be workforce reallocation and reinvention.
Here is an example of why one or a few people or “no one works” here is not possible in a regulated world and many industries anytime soon: https://www.linkedin.com/posts/linkashwani_ai-nyt-activity-7445917307769364480-Q0kI
And if we focus too much on the cost cutting, this is where we will end up:
Ford had to rehire 350 engineers after its AI got vehicle quality wrong
The automaker's VP of hardware engineering says Ford mistakenly believed AI could replace experienced workers, a blunder it fixed just in time to top JD Power's quality ranking for the first time in 16 years
More here:
https://thenextweb.com/news/ford-rehired-350-engineers-ai-quality-jd-power
Thank you so much for giving so much thought to the piece, and for your critical views. I especially like the point that the one-person company concept is often a provocation rather than a direct label in many highly regulated industries. In that regard, the goal is not to become a one-person company, but relative to other bloated peer groups, you are one inch ahead in satisfying all the different non-negotiables.
What I really appreciate in your comment is that you highlight the multiple dimensions people need to think about systematically. It is not a single dimension, such as just thinking about the customer. Other companies cannot do that.
Thank you again for sharing your work in progress and your current thinking. This is how we get the truth out as the landscape continues to evolve. I am sure some of your input will serve as inspiration for my next piece in the future.
I spent some time looking at model routers and concluded that the cost differential between a subscription and API based usage was still too much. And the economics of context caching dramatically raises the cost of model switching for long-running work.
I can imagine applications backed by a model router where each session is small and well-bounded. And I see a lot of potential there for subagents and agent teams. But for long-running high-complexity sessions the cost differential remains huge.
I know, thank you for giving me additional thoughts as the situation evolves so rapidly. Have to keep thinking about and staying close to the phenomena.
Thanks. Amazing insight. It looks a lot like a typical lean project. Instead of improving processing speed or operation time (value-adding time), concentrate on eliminating non-value-added activities. Do the right thing before doing the thing right. The second insight is the bottleneck one, which again is often found in manufacturing activities. It does not make sense to improve a machine's OEE if the next machine is a bottleneck in the process. The transition to an AI agent world may well deliver a future where non-value-adding activities have been minimized or reduced to almost zero.
For that to happen, the concept of value stream would be important to emphasize and implement.
It is lean project all over again, which I find comforting because there is a dominant narrative saying AI is changing everything and all the old rules do not apply. As I go through more and more of the deeper thinking, we actually know a lot about how to cope with the upcoming challenge.
There are the lean startup methodologies, and there is the re-engineering playbook. It is just that we need to apply them much more broadly in context, not just living within a small group or within software engineering. It is the same philosophy, applied deeply across the entire organization or enterprise.
Thank you so much for the spot-on observation. I'm glad that piece has been helpful.
Thanks for the great historical account. I wasn't aware of the details behind the Nokia story, which goes alongside The Mythical Man-Month as a cautionary and instructive tale.
I buy the overall concept -- in chemistry this is analogous to the rate-limiting step of reaction kinetics. For an AI-centric business, though, I'm afraid we still lack transparency on cost to benefit. I'm a research scientist using a $20/mo Anthropic subscription, and Claude Code tells me I used an equivalent of $180 in tokens yesterday to accomplish some pretty routine code organization (much of which involved cleaning up its own mess). Its worked great for parts of this project *AND* I 100% could not afford to pay full freight here.
This is partly why Cursor is a cautionary tale: they're downstream of the model providers, and now Anthropic is under-cutting them on cost to a similar product by a factor of ~10x. Building a business (or a work process) on something you don't own or control seems like a big liability. And yes,
I can imagine a future where model compute becomes a commodity and prices drop, but that comes with it's own set of challenges.
There's a learning curve to new tools and tech, and I think we're still early days for AI use *optimization*. But we're also facing huge uncertainties regarding the funding and financing of model providers that adds existential risk to an over-reliance on some of these tools.
I like it. Thank you so much for pointing it out, almost from the ground up, with first‑hand observation.
Quite honestly, it feels like the push to deploy AI everywhere is driven by a pricing policy that U.S. Frontier Labs could have changed overnight. The discussion about token expense wouldn’t exist if they hadn’t switched from a monthly seat to token billing. But the broader point you describe is spot on. Thank you again.
And, critically, their switch to token pricing wasn't accidental. It was driven by IPO demands of opening their books, showing revenue to cover the mountains of debt. In this regard, I think Ed Zitron is spot-on in his financial analysis of the so-called frontier labs.
In my limited testing, Anthropic retains a clear lead on Chinese open-source models in terms of the models themselves and the overall usability of the ecosystem. Claude Code is a great product, and Opus is solid (and Fable was blocked by the government -- a whole separate cautionary tale there!). But OpenAI and Anthropic prices are being heavily subsidized by investors, and those costs are not sustainable.
I think this is a sharp reminder of what drives this: price switching, which, for some reason, was less discussed among executives in my classes or in the media. As you said, the blocking of Fable warrants another extensive essay for executives worldwide to reduce reliance on a single service provider.
The token pricing or "correct model for the job" conversation is being addressed by OpenRouter. Your AI disruption article explains this very clearly Howard.
Per the "No One Works Here" thing...I was just reading a short blurb about "1 person companies" in China. I forget the actual number, but it was in the hundreds of thousands...maybe millions...whatever, it was a huge number...of one person companies utilizing AI, mostly DeepSeek, that have been formed in China in the last 12 months.
Really? I haven't even thought about that yet. Thank you for the ground-up observation.
Couldn't find the piece, but a quick DeepSeek query here....
"One-person companies (OPCs) in China have moved from a niche concept to a national strategy, fueled by AI and new regulations . By the end of 2025, nearly 7.32 million new OPCs were registered—a 42.3% year-on-year increase .
What Defines an OPC?
· Sole Shareholder, Not Sole Employee: An OPC has only one shareholder (natural or legal person) , but can hire staff and appoint directors .
· Limited Liability: The core advantage. You are only liable up to your capital contribution, protecting personal assets—unlike sole proprietorships which have unlimited liability .
· New Flexibility: Revised Company Law allows a natural person to own multiple OPCs .
Why AI is the Catalyst
AI dramatically lowers startup costs. Entrepreneurs now use AI tools to handle design, operations, and marketing tasks that once required a whole team . This "individual plus supply chain" model is unique to China, allowing solo founders to tap into dense manufacturing ecosystems .
Government Backing & Support
Local governments are building full-stack support systems to attract OPC talent :
· Subsidized Computing Power: Shanghai's Pudong offers OPCs free computing power worth up to 300,000 yuan ($43,876) .
· Dedicated Hubs: Specialized parks like Beijing's Zhongguancun AI North Latitude Hub provide shared workspaces, business registration services, and investor matchmaking .
⚠️ Crucial Legal & Financial Risks
This model is not without pitfalls; the "limited liability" shield can be pierced:
· Asset Separation: The sole shareholder must prove personal and company assets are separate. If not, they bear unlimited joint liability for company debts .
· Strict Accounting: Maintain separate bank accounts, keep meticulous records, and complete a mandatory annual audit. Avoid using personal accounts for business .
OPC vs. Sole Proprietorship (个体工商户)
One-Person Company (OPC) Sole Proprietorship
Legal Status Corporate legal entity Natural person (no legal entity)
Liability Limited liability (to investment amount) Unlimited liability (with all personal assets)
Bottom Line: An OPC is a powerful, low-risk vehicle for AI-savvy founders. However, the administrative and legal discipline required is high—formal accounting and strict financial separation are mandatory to maintain liability protection .
If you are considering this path, would you like to know more about the specific registration process or capital requirements?
It was very recently in an article and I just glanced through it because it caught my eye. Your post made me think of it. It was just a couple days ago. It might have been in Caixin. I'll see if I can find it again.
If you offered me $40k in free computing power I'd register an OPC also!
Thanks for this -- I think this also speaks to the ongoing evolution of CCP's business and economic policies. Bankruptcy laws, LLCs and financing long played a role in Western capitalism. China now has a long track record of remixing successful Western business approaches for their own economic development while carefully controlling political expression.
Observing companies' behavior becomes especially intriguing because China has many copycats, intense competition, and often only one national champion.
I often times couldn't help but think about the managerial principles and leadership that emerge, regardless of location, and how they evolve into a national champion, whether in the US, Europe, China, or Southeast Asia. To me, this reveals a universal truth. Thanks again for reading and keep pushing our thinking along the way.
不客气 It's amazing to watch. In the last couple decades, the business development here is breathtaking. The integration of local government and business is profound. Things go from concept to operational in a few months.
Thanks for the additional reference. In Europe, one‑person companies are mostly discussed as a concept and a management practice. I had no idea they are so widely operationalized in China, where they are almost the norm. The speed of adoption there is remarkable.
I'm in Wuhan and briefly talked about it with a friend, then I saw the piece...I'm pretty sure it was in Caixin....wherever it was, and just now your piece reminded me of it. it really struck me. The tech/supply chain/mfg/distribution stack here is so insanely robust, it makes total sense. One person can be a corporate level enterprise.
It‘s important not to mix up everything. A software workflow is not a business workflow and both are not governance workflows. Not everything needs the same speed. And not every tech or business workflow can be treated the same. It’s a difference if you build something new, if you integrate something, if you process ad hoc tasks or if you maintain something. You need a meaningful cadence for workflows that are connected.
Regarding shipping software: long before AI there was a concept called DevOps. It‘s worth to research that, because how it was implemented in some companies misses the concept entirely. It was long before AI possible to ship to production the same day you developed the code. With a proper DevOps Setup.
Regarding business workflows: the root cause for problems lies ALWAYS in the set up of the teams. The right mix of roles and accountability solves 80% of all problems.
Governance can be slowlier. The root cause is that topics are treated as governance topics when they are actually daily business. Root cause is wrong setup of roles and accountability.
The big art is to set up several teams and leadership levels in a way that it all runs smoothly.
This is what I do for a living, besides other topics.
It starts with the mindset and it is a never ending work.
The goal is to become a learning organization with a cadence of learning that is not rushed and is sustainable.
Great call‑out, as always.
Before we confront AI opportunities and challenges, DevOps serves as a playbook along the way. Because of the page limit, I couldn’t discuss The Phoenix Project, but I love how you concluded.
The key is to build a learning organization. Deploying an agent doesn’t justify shortcutting learning for human controllers when setting goals and reengineering processes end‑to‑end across silos.
Thanks again, Bianca, as always.