Not a Freelancer, Not a SaaS: The Agent-Staffed Function

I run a full marketing function alone. One person in the seat; a department of specialized AI agents underneath. When people try to place that, they reach for one of two existing labels. The first is the freelancer with ChatGPT. The second is the software product: an AI tool you log into and run yourself. Both are wrong, and they are wrong in opposite directions.

The label matters more than it looks, because labels carry predictions: what the thing costs, where it breaks, what to ask before you trust it. File this structure as a freelancer and you will expect one person's capacity and one person's ceiling, and be wrong about both. File it as software and you will ask for the login; there is no login. The structure is spreading fast enough that buyers, hiring managers, and founders are all starting to run into it, and evaluating it under either borrowed label produces the wrong decision.

The accurate label is the Agent-Staffed Function. One accountable human, an expert in the function's domain, occupies the seat of a full business function; a department of specialized AI agents handles the execution below them. The human holds strategy, taste, judgment, and final authority. The agents are the specialist layer, each with a defined role, a private context, and explicit failure-mode discipline (the architecture I described in Engine Log #3). It is not a freelancer, not a tool, and not a product company. It is a new organizational unit, defined by where accountability sits rather than by headcount. The canonical definition and the qualifying test are maintained at Built, Not Hired; this article is the argument behind it, with the short version at the end.

The two structures it is not

The fastest way to see the Agent-Staffed Function clearly is to rule out the two shapes everyone confuses it with.

It is not freelance-with-AI

The freelancer-with-AI is one generalist using models as a productivity multiplier. The copy gets drafted faster, the deck gets built faster, the research happens faster, but the structure of the work is unchanged: one person holding every function in their own head, switching between them in sequence. AI makes that person quicker. It does not make them more than one person.

The tell is the bottleneck. In a freelance-with-AI setup the human is still the single point every task flows through, and the quality ceiling is whatever that one generalist can hold active at once. This is the line people miss, and it is a line of kind, not of degree. A freelancer who runs five chatbots is still being the specialist: they write the copy themselves, only faster. In an Agent-Staffed Function the human does not execute the specialist work at all. They direct a structured specialist layer (defined roles, private context, failure-mode discipline) and approve its output. The freelancer's deliverable is their own work product, accelerated; the operator's deliverable is the synthesis of work they did not personally produce. One is a faster generalist; the other is a seat above a department. That is where the category begins.

It is not a software product

A software product company sells software. The product is the thing the customer buys and operates; humans inside the company supervise the product. The customer logs in and runs the tool themselves. This is SaaS, and the unit being sold is access to a product.

An Agent-Staffed Function is the opposite orientation. The customer does not log into anything. They hire an outcome, delivered by a person who is accountable for it, the way they would hire a senior operator or an agency. The agents are not a product the client operates; they are the department the operator operates, and the client never touches them. What the client buys is judgment and delivery, not a login.

None of this is a claim about funding, scale, or where it sits. An Agent-Staffed Function can be bootstrapped or venture-backed, and it works the same way as an outside agency of one or as an internal team a company runs for itself. What defines it is where accountability and execution sit, not the cap table or the org chart.

The third structure

Strip away the two near-misses and what remains is specific. One human stands in a decision seat. Below that seat, the functions a department used to staff with people are reorganized into specialized agents. The human does the work that requires accountability and taste; the architecture does the work that has clean inputs and outputs.

The division is not arbitrary. The human holds the things a buyer can fire you for: the decisions, the judgment calls, the domain orientation, the taste that separates good work from merely competent work. The architecture holds the things that scale badly in a person and well in a system: repetition, parallel processing, deep single-domain specialist work, and the failure-mode reflexes a generalist cannot keep active across thirty-odd specialist categories at once.

That is why it is neither of the familiar shapes. It has the accountability surface of a senior hire and the production capacity of a software system, combined in one person who is doing a job, not selling a tool.

Two modes: external and internal

Here is the part that makes the Agent-Staffed Function a general structure rather than a description of my own company. The unit is the seat, not the firm. It deploys two ways.

Externally, it is a solo operator delivering a service at agency quality: one person, no staff, a department of agents. Ravenopus is the worked example. I sit in the seat a VP of Marketing would occupy; the marketing department below me is agents.

Internally, it is a single seat inside an existing organization. I have built one: Mouret, an analytics function deployed inside a regional supermarket chain and operated by the chain's own staff, not by me (covered below). The seat sits where a head of analytics would; the specialist layer beneath it is agents. The same shape fits a finance lead, a compliance lead, a growth lead, at any company size. The seat stays and the org chart above it is unchanged; the headcount below it becomes an agent layer. This internal mode has a natural shorthand: operator-as-VP, one operator occupying the seat a VP would hold. Building this internal version for other companies is what my second company, Built, Not Hired, does.

The same structure works for a solo founder and for a department head inside a large enterprise, because both are the same shape: one accountable human, one function, a specialist layer underneath. Company size is irrelevant. The seat is the unit.

Why the seat is the right unit of analysis

It is worth being precise about why the seat, and not the company or the employee, is the thing being named.

Most of the existing language picks the wrong altitude. Microsoft's Frontier Firm names the whole company. Its agent boss names the individual who delegates to agents; that term is useful, and it is not a competitor to this one, because the agent boss is the person who sits in the seat. Agent boss names the role. Agent-Staffed Function names the unit that role runs. The one-person unicorn names an outcome, a valuation, and almost always a software product. None of them names the structural unit in the middle: a single function, owned end to end by one human, with a specialist layer below.

That unit is the seat. A functional seat (VP of Marketing, head of analytics, finance lead) has three properties that make it the natural unit. At the top there is high decision density: a steady stream of judgment calls that compound, where being wrong is expensive in ways that surface quarters later. In the middle there is narrow, well-specified specialist work: each task has a fairly clean input and output. At the bottom the output spec is relatively flat: the seat produces a recognizable, repeatable catalog of deliverables rather than open-ended invention.

The human keeps the top layer, because that is the layer that requires someone who can be held responsible. Everything below it, the specialist work feeding a flat output spec, reorganizes cleanly into agents. The seat-holder was never personally doing all the execution anyway; they were directing specialists. Replace those specialists with agents and the seat is unchanged. The label that survives the swap is the accurate one, and the label that survives is the seat.

The Agent-Staffed Function Test

If the structure is defined by the seat, then there has to be a way to tell whether a given seat qualifies. There is. I call it the Agent-Staffed Function Test, and it checks the three layers of the function. All three have to hold.

1. Decision Ceiling (the top). Is there enough judgment, taste, and strategy density at the top that the function needs a permanent human principal? If yes, a human belongs in the seat. There are two ways a function fails here. If there is almost no judgment required, there is no seat to occupy and the function is just automation, not an Agent-Staffed Function. And if judgment and execution are fused, so that you cannot separate the deciding from the doing, you cannot hand the doing to agents at all.

2. Skill Depth (the middle). Can the execution be specialized? Can it be broken into narrow roles defined cleanly enough to capture in a constitution and a role brief, so an agent can perform them reliably? If the middle is irreducibly ambiguous, there is no boundary for a specialist agent to hold, and the work cannot be delegated.

3. Output Floor (the bottom). Are the deliverables specific and inspectable enough that the human principal can review and approve them without redoing the underlying work? If the only way to verify the output is to rebuild it, there is no leverage, and the structure collapses back into one overworked person.

A function that passes all three can be rebuilt as an Agent-Staffed Function. Marketing passes: strategy and taste at the top, specialist execution in the middle, inspectable deliverables at the bottom. So do analytics, finance, compliance, growth, customer research, and due diligence. Each has a senior decision-maker at the top, specialist work in the middle, and a flat catalog of deliverables at the bottom.

A surgeon's practice fails the test on the Decision Ceiling: the judgment and the embodied doing cannot be separated. A fully automated reporting pipeline fails it from the other direction: there is no real judgment, so no human seat is needed and it is simply automation.

The test earns its keep on the cases that almost pass. Early-stage investing is one. Deciding which companies to back has dense judgment, so the Decision Ceiling holds, and the diligence underneath is specializable, so Skill Depth holds. But there is no Output Floor: you cannot confirm a pick is right without effectively making the call yourself, and the verdict takes years to arrive, so the principal cannot approve the work without redoing the judgment. It fails on one layer, and one failed layer is enough. A test that never disqualifies a realistic case is not a test. The functions that pass are the ones where a human must hold the top, the middle can be specified, and the bottom can be inspected without being rebuilt.

A receipt for the internal mode

I did not want to argue this from marketing alone, or from my own company alone, because the obvious objection is that it works only because I am the one in the seat. So I built the same architecture for a different function, in a different company, run by a different person: an expert in their domain, not in mine. The system is called Mouret: an analytics function for a regional supermarket chain that I built, deployed inside the chain, and operated by the chain's own staff.

It runs on the same Claude Code primitives as Ravenopus: a constitution file that defines the operating rules, a skills layer that holds domain expertise, a workflow layer that encodes the analytical pipelines, and specialist sub-agents the workflows delegate to. Concretely, that is 13 skills, 19 workflows, and 8 specialist sub-agents (data engineering, data science, FP&A, marketing analysis, product, supply chain, data quality, and lead strategy), connected to the chain's live operational database (100+ tables).

The seat Mouret occupies is not VP of Marketing; it is the analytics function: financial reporting, merchandising, inventory, demand forecasting, promotional incrementality, supplier performance. The operator, who works for the chain and not for me, makes the calls with decision density: which question to ask, whether the data can be trusted, what to do about the answer. The agents do the specialist work underneath: write the query, run the difference-in-differences, classify the catalog, build the forecast. The output spec is flat by design. Every analysis ends with a recommendation rather than a raw table.

Two of its findings are worth stating, anonymized but real. One analysis showed the chain was losing money on poultry, and the cause was not pricing; it was waste, expired product quietly eating the category's margin. Another showed the chain had been underestimating seasonal outdoor kids' toys: the category was not just selling, it was pulling sales in other categories along with it, so understocking it suppressed baskets well beyond the toy aisle. Neither is a marketing finding. They came out of the same shape doing analytics.

This is the receipt for the internal mode, and it does two things the external case cannot. It shows the structure works inside a company, not only as a service sold from outside one. And it shows the architecture is not welded to its builder: I built Mouret, but the person operating it is not me. The results matter, and the deeper point is the transfer: the same constitution-skills-workflows-subagents shape that runs a marketing function runs an analytics function, because both seats pass the Agent-Staffed Function Test: a Decision Ceiling at the top, Skill Depth in the middle, an Output Floor at the bottom.

But read the transfer claim precisely, because it is about the shape, not the people. The shape is a chassis. What makes a specific Agent-Staffed Function good is expertise, and it enters at both layers. The seat has to be held by an expert in the function's domain: the Decision Ceiling is exactly the layer where expertise lives, in which question to ask, which answer to distrust, which trade-off to take. Mouret's operator runs it well because they know the retail business they sit inside, not because the architecture absolved them of knowing it. And the agent layer below is only as good as the domain knowledge encoded into it. Ravenopus produces senior-level marketing work not because agents are generically capable but because years of marketing practice are written into them: the frameworks, the playbooks, the quality bars, the failure modes each specialist is required to check. Mouret's agents carry the same kind of encoded depth for retail analytics. Copy the shape with a generalist in the seat and an empty skills layer and you get generic output at speed. The domain is interchangeable. The expertise, in the seat and under it, is not.

Where it does not transfer

The honest version of any structural claim includes its boundary. The Agent-Staffed Function does not fit every function, and pretending it does is how the idea discredits itself.

It does not fit functions with high-stakes embodied judgment, where the work requires being physically present and reading a situation in real time. It does not fit functions where a single wrong answer compounds catastrophically for years rather than surfacing as a traceable miss, because the cost of the rare failure swamps the benefit of the common speed. And it does not fit functions where the buyer specifically requires a human to be the one in the room, for reasons of trust, liability, or regulation, regardless of who did the underlying work.

Surgery does not fit. A great deal of clinical and legal judgment does not fit, at least not yet. These boundaries will move as the tools improve, but they have not moved yet, and a structure is more credible when it names the places it does not reach. The functions that fit today are the ones that pass the test: a Decision Ceiling that needs a human, a specifiable middle, an inspectable output, and a wrong answer that is recoverable.

Where this leaves the category

The Agent-Staffed Function is new enough that it is still being figured out. What I am confident about is the shape: one accountable human in a decision seat, standing over a department rebuilt out of specialized agents, deployable as an external service or an internal unit at any company size, delivering an outcome rather than selling a tool. It is not a freelancer who bought better software, and it is not a software product with a human on call.

The label decides how the structure gets evaluated. Filed under "freelancer with AI," it gets priced like freelancing and dismissed like freelancing. Filed under "software product," it gets measured against SaaS margins it was never built to have. It is neither. It is its own row in the table. So: this is the row, this is the test for whether your function belongs in it, and this is where the definition lives.

---

If you want to see what an Agent-Staffed Function produces before you trust the argument, the 72-Hour Growth Diagnostic is the smallest way to do it: a teardown of your funnel, three prioritized interventions, and one prototype blueprint, produced in 72 hours by the structure this article describes. The output is the proof.

— Linara Bozieva, Founder, Ravenopus

---

In one paragraph, and a few common questions

Agent-Staffed Function (the structure this article names): an organizational unit in which a single accountable human, an expert in the function's domain, holds a full business function while a layer of specialized AI agents handles execution below them. It deploys externally (a solo operator delivering a service) or internally (a department-of-one seat inside a company of any size, the operator-as-VP case). It is defined by where accountability sits, not by headcount. The canonical definition is maintained at Built, Not Hired.

Is an Agent-Staffed Function a freelancer or an AI startup? It is not a freelancer; a freelancer has no specialist layer, and their output is personal work product. "AI startup" usually means a SaaS product the customer logs into and operates, and an Agent-Staffed Function is not that: it delivers an outcome through an operated agent department, not a login. The structure is orthogonal to funding, so a company can be both an Agent-Staffed Function and a venture-backed startup.

Can a large company have one? Yes. The unit is the seat, not the company. A function head inside a 5,000-person company can rebuild their department as an Agent-Staffed Function if it passes the test; the seat stays, the headcount below it becomes an agent layer.

How do I know if a function qualifies? Run it through the Agent-Staffed Function Test: a Decision Ceiling (enough judgment to need a permanent human), Skill Depth (execution narrow enough to specify for agents), and an Output Floor (deliverables inspectable without redoing them). Pass all three and it qualifies.

Does the operator need to be a domain expert? Yes. The structure removes the need for specialist headcount, not the need for expertise. The seat-holder needs real expertise in the function's domain, because the layer they keep (judgment, taste, final calls) is where expertise concentrates; and the agent layer needs that domain knowledge encoded into it to produce senior-level output. An Agent-Staffed Function built and run by an expert produces expert work; a generic copy of the shape produces generic work.

What is operator-as-VP? The shorthand for the internal deployment of an Agent-Staffed Function: one operator occupying the seat a VP or function head would hold inside a company, with a department of agents below them. The broader category, which also covers the external mode (a solo operator delivering a full-function service, the way Ravenopus operates), is the Agent-Staffed Function.