Writers of Headlines love to trumpet things like ‘Scientists Say’ or ‘Researchers Say’, as if Scientists or Researchers were one coherent blob that can ‘say’ a thing (I think we are meant to take this as shorthand as some scientists say, but the shorthand is irksome).
Let’s turn the tables: Headlines are talking about AI agents doing your work. See for example America Isn’t Ready for What AI Will Do to Jobs
Headlines are talking about AI agents doing your job, but not often asking the harder question: if agents handle the production part of your job (making ‘stuff’), what’s left? And is what’s left actually the valuable part, or the overhead?
Here is a suggestion I would like to make about what might happen soon(ish): every knowledge worker’s job is about to split into two distinct functions that have traditionally been bundled together. Most people can’t tell them apart. One function is what one produces. The other is what makes the production worth producing. AI absorbs the first. The second is what you actually sell.
The “solo founder running a company staffed by AI agents” vision is gaining traction. Sam Altman talks about it. Andreessen Horowitz funds it. But the vision has a gap: it assumes the hard part is building the agents. The hard part is actually understanding what the boss of those agents does all day, and why it’s valuable.
A Job Is Two Things Pretending to Be One Thing
A knowledge worker’s salary has always compensated two different functions bundled into a single role. The production function, e.g. creating deliverables, writing reports, building analyses, writing code. And the principal function, which involves deciding what to produce, for whom, to what standard, and why.
In a traditional job, you never had to distinguish between these because your employer bought both in a single package called “your salary.” The production function was visible. Here’s the report. The principal function was invisible. The judgment that made it the right report.
David Maister understood this split decades ago. In his model of professional services firms, every firm is a pyramid of three roles: finders, minders, and grinders. The grinders do the work. The minders manage the work. The finders bring in the work and ensure it’s worth doing.
Here’s the thing about Maister’s model that matters now. The partner’s value in a professional services firm isn’t in billing hours. Roughly 60% of partner profit comes from the margin on other people’s work. The partner who does no billable work but brings in the right clients and ensures quality is more valuable than the associate who bills 2,000 hours.
AI agents are the ultimate leverage play: infinite grinders, and seemingly every day now increasingly capable minders. What remains is the finder function. The principal.
The production-principal split reveals that this has always been true of knowledge work generally. We just couldn’t see it because both functions lived in the same person.
What the Putting-Out System Teaches Us
In pre industrial England, there was something called the putting-out system. From roughly the 15th through 19th centuries, merchant-capitalists distributed raw materials to home workers, such as weavers, spinners and dyers, who processed them with their own tools and returned finished goods. The merchant then gathered the outputs and sold them in distant markets. Small businesses of a sort that the workers ran out of their own homes.
The putting-out system eventually gave way to factories, and the reasons are instructive. Quality was inconsistent across dispersed workers. Coordination became a bottleneck, for example it took four to eight spinners to keep one weaver supplied. And when machinery came along that could make these tasks less labor intensive the machinery was large and it simply couldn’t fit in a cottage. That, combined with the coordination bottleneck, centralized work in factories.
The putting-out system didn’t fail because cottage workers were bad at weaving. It failed because the merchant-capitalist’s coordination function became a bottleneck. Managing dispersed workers required an apparatus that could only work at that time as a factory. The principal function—knowing what to make, ensuring quality, coordinating supply chains—proved harder than the production itself.
If one were to become the boss of an agent micro-business, you inherit the merchant-capitalist’s problem. The production function scales easily. The principal function does not, and the problem remains: Who Does That Work?
The Five Faces of the Principal Function
So what exactly is the principal function? It’s not one thing. When you decompose what remains after production is delegated, you find five distinct capacities.
1. The Commissioning Function: Deciding What to Make
Before any agent produces anything, someone decides what should be produced, to what standard, for whom, and why now. This is inherently circular—you can’t write a specification for “figure out what specifications to write.”
Aghion and Tirole drew a useful distinction between formal authority (the right to decide) and real authority (effective control over decisions, determined by who has the information). In a traditional job, your manager often exercised the commissioning function for you. In the micro-business model, you inherit it in full.
This is the meta-task. It cannot be productized because it’s the task that defines all other tasks.
2. The Contextual Function: Maintaining the State of Awareness
This isn’t a deliverable. It isn’t even an activity. It’s a state—ongoing readiness to interpret signals correctly.
Michael Polanyi called it tacit knowledge: not merely knowledge that hasn’t been articulated but knowledge that cannot be fully articulated. Reading between lines in a meeting. Understanding what a client actually needs versus what they asked for. Sensing political shifts in an organization.
The contextual function is what makes the production function useful. Without it, agents produce technically competent outputs that miss the point. This was exactly the putting-out system’s quality problem—dispersed workers couldn’t see the end product or understand how their piece fit. The merchant who understood the final customer was the one who could specify quality.
3. The Translation Function: Bridging Between Domains
Converting between the language of one domain and the language of another. Any competent agent can turn a technical document into plain English. Only someone embedded in context can turn it into the right plain English—the version that lands with this specific audience at this specific moment.
Maister’s framework applies here too. He distinguished between Brains work (novel, creative), Grey Hair work (judgment from experience), and Procedure work (standardized and systematic). Procedure translations are fully productizable. Brains translations are genuinely creative acts. The premium is in the latter, and the latter is where tacit knowledge—your personal jagged frontier—makes all the difference.
4. The Accountability Function: Owning the Outcome
When your agent micro-business delivers a flawed analysis and someone makes a bad decision based on it, who takes the call?
Accountability can’t be productized because it’s not a deliverable. It’s a risk transfer. Think of it like a general contractor. The general contractor charges a margin above subcontractor costs. That margin compensates for coordination and for being the person who gets the call when the plumber’s work floods the house.
This may become the primary basis for compensation in knowledge work: not the hours your agents worked, but the guarantee that output meets the standard. We’ve explored this dynamic before in that the verification burden is the gap between producing something and standing behind it. And as judgment depreciates without practice, the ability to stand behind AI output becomes a scarcer, not more abundant, skill.
5. The Relational Function: Being the Continuous Thread
You’re the person colleagues know, trust, and reach when they have a question that doesn’t fit a project brief. Tanya Reilly called this “glue work.” In the production-principal split, it becomes a structural principle.
Relationships depend on duration. A relationship is not a transaction. It accumulates over time through shared struggle and demonstrated reliability. Agents can’t have relationships—no continuity, no memory of shared experience, no reputation to protect.
The general contractor flywheel applies: context builds trust, trust generates commissions, commissions produce deliverables, successful deliverables deepen context. Each turn of the wheel strengthens the principal function. Each turn is something an agent cannot replicate.
Why This Changes How You Think About Your Career
Here’s the uncomfortable implication. Not all knowledge workers have a meaningful principal function. Some jobs really are predominantly production. When those tasks can be handed to agents, the split reveals there’s little principal function left over.
This isn’t new. It’s newly visible. Maister’s framework already recognized that Procedure-type work requires the lowest individual judgment and supports the highest leverage ratios. In the spirit of Tasks Make Stuff, AI generates outputs, but it doesn’t solve the coordination, judgment, and relationship problems that surround those outputs. The split doesn’t create the problem. It reveals it.
The career strategy follows directly: cultivate your principal function deliberately. Build context. Develop judgment. Invest in relationships. Learn to commission well. Because the production function is exactly what AI agents are designed to absorb. The principal function is what they can’t.
For employers, the question reshapes too. Not “which jobs can we automate?” but “which roles have a principal function worth paying for independently of their production function?” That question reshapes org design more fundamentally than any technology deployment.
The Question Behind the Question
The discourse about AI agents replacing knowledge workers is asking the wrong question. It’s not “can an agent do my job?” It’s “which part of my job is the job?”
To understand this in practice,try this exercise. List everything you did last week. Separate the list into two columns: things an agent could have produced if properly briefed, and the briefing itself—plus the judgment calls, relationship maintenance, and accountability that surrounded the production.
Look at the second column. That’s your principal function. That’s what you’re actually selling.
America Isn’t Ready for What AI Will Do to Jobs
Managing the Professional Service Firm
The origins of the putting-out or domestic system of industrial production in England