AI Team Build · Part 1 of 4
Hiring was the slow, expensive, hard to reverse answer. Writing job descriptions for AI agents was the other one.
In short
A small marketing agency can run the workload of a much larger shop by building a team of AI agents instead of hiring, if each agent gets one job, one written mandate, and a human who approves its output. Conspicuouz Creative Group did this with two operators. The first attempt failed. The second one held.
Where it started
The first attempt was one long conversation. Everything went into it. Client context, brand rules, a request for a proposal, then a request for a blog post, then a request to fix the blog post because it did not sound like the proposal.
By the fourth week the conversation knew a great deal and remembered almost none of it in the right order. Every new session started with a paste job. The paste job got longer. The output got more generic.
That was the failure. Not the tool. The shape of the work.
A chat is a conversation you think with. It is not a system you can hand work to.
Anthropic draws the same line in its own training material: Chat is for thinking with Claude, Cowork is for delegating to Claude, and Code is for building software with Claude (Claude Academy, Introduction to Claude Cowork). Part 2 of this series covers those three surfaces. Part 1 covers what had to change before any surface mattered.
The workload
Here is the list as it stood. Not the aspirational version, the actual one.
| Function | What it meant week to week |
|---|---|
| Strategy and positioning | Offer design, pricing, scope decisions, KPI definition |
| Creative direction | Brand interpretation, design standards, video standards, final creative approval |
| Content production | Blog copy, landing page copy, email sequences, social captions |
| Video | Scripts, generation, editing, captions, lower thirds |
| Paid media | Briefs, creative variants, budget pacing |
| Technical marketing | Schema, crawler access, site speed, tracking, UTM hygiene |
| Client leadership | Debriefs, approvals, renewals, the hard conversations |
| Sales | Qualification, outreach, proposals, follow up |
| Operations | Asset intake, task setup, reminders, weekly reporting, access checklists |
| Data | Baselines, monthly reads, before and after comparisons |
| Legal and contracts | Change orders, compliance flags, agreement review |
Two people. Every row.
Most weeks the bottom half of the table did not get done. Reporting slipped. Reminders went out late or not at all. The work that produces revenue crowded out the work that protects it.
Anyone who has run a small shop knows the feeling of a Friday where the client work shipped and the business itself got nothing.
The obvious answer
Headcount was the responsible looking move. Post the role, run the interviews, make the offer, onboard for sixty days, then find out in month four whether the fit was right.
Three problems with that, in the order they hurt.
Slow. A hire takes months to reach full output. The workload above was due on Tuesday.
Expensive. Salary is the visible number. The invisible number is the operator time spent managing, training, and correcting a new person while still carrying the original load. For a two person company, the manager is also the producer. Every hour of management is an hour of production lost.
Hard to reverse. A bad hire in a fifteen person company is a problem. A bad hire in a two person company is a crisis. The decision felt like it had to be right the first time, which is exactly the kind of pressure that stalls a decision entirely.
None of this makes hiring wrong. It makes hiring the wrong first move for a company this size, and the rest of the series depends on that distinction.
The turn
The shift was not a new tool. It was a new question. Instead of asking what a chat window could do, the question became: what would this role be, if a person held it, and what would the job description say?
A job description does four things. It names the role. It gives the context the role operates in. It defines the task. It says what a finished piece of work looks like. That structure turned out to be the entire instruction set an AI agent needs. Anthropic's own documentation describes project instructions as a way to define roles or perspectives Claude should adopt and set guidelines for a particular workflow (Claude Help Center, Understanding Claude's personalization features). A role and a workflow. A job description, in other words.
The temptation is to build a single assistant that does everything, because that feels efficient. It is not. An agent with eleven responsibilities produces the average of eleven kinds of output. An agent with one responsibility produces the best version of that one thing, and it can be corrected without breaking the other ten.
Every agent on the team can be described in one sentence that names the role and the output.
Blog Writer: produces publish ready long form posts that get cited by AI answer engines.
Ops Coordinator: chases, files, checks, and reports, and never creates.
If the sentence needs an "and also," the role is two roles.
The instinct is to start with the agent that writes the thing that gets paid for. The agency's first useful agent was the one that chased, filed, and reported, because that was the row on the workload table that never got done, and an undone row in the bottom half of the table is what quietly costs a client.
A writing agent produces output the operators would have produced anyway, just faster. A coordination agent produces output that was not happening at all.
Start where the gap is, not where the glamour is.
One more thing a job description forces: an ending. A person in a role knows what finished looks like because someone told them. An agent does not, unless the description says so. The last section of every instruction the agency writes is a description of the finished piece, in enough detail that two different people would recognize it.
The composite
Ridgeline Plumbing and Air is invented. It is a home services company in a mid-size market with six trucks, one office manager named Dana, and an owner who answers the phone on weekends.
Dana runs scheduling, invoicing, the Google reviews response queue, the monthly newsletter, and the social posts nobody has time for. The newsletter has gone out four times this year.
Dana's workload table is shorter than the agency one above. It is the same shape. Two or three people, a dozen functions, the protective work always losing to the urgent work.
The instinct at Ridgeline is the same instinct the agency had. Hire a marketing coordinator. The math is the same too. A coordinator costs a salary plus the owner's time to manage them, arrives at full speed in three months, and is hard to let go if the fit is wrong.
The alternative Ridgeline will build across the next three parts. One agent that drafts the review responses in the company's voice for Dana to approve. Then one that drafts the newsletter from the month's completed jobs. Then a chain, where the first agent's output feeds the second.
Each one gets a job description. Each one gets a human who says yes before anything goes out.
The honest line
The agency did not replace staff with software. There was no staff to replace. What changed is where the two operators spend their hours. Less producing. More reviewing, correcting, and deciding. Part 4 accounts for that trade in full, including the parts that did not go well.
Suggested shape: what the shift from producing to reviewing actually felt like in week one, stated plainly. This slot clears the Proof floor on this post. Fill it or remove this block before publish.
What this does not do: an AI agent does not know your business until you write it down, and it does not stay correct on its own. Every agent needs a written role and a human approval step. Skip either and the output drifts within weeks.
Do this today, free, under five minutes
Write one sentence for one role you wish you could hand off. Name the role and name the output.
"Review Responder: drafts a reply to every new Google review in our voice for me to approve."
That sentence is the first line of the job description you will build in Part 3.
Common questions
Yes, for production and coordination work, with limits. A two person agency can build separate AI agents for writing, operations, and reporting, each with a written role and a human approval step. Strategy, client relationships, creative judgment, and final approvals stay with people.
Each AI agent gets a single role and a single type of output, described in one sentence. An agent asked to do everything produces average work across everything. An agent with one job can be corrected without breaking the others.
A chat is a conversation. It does not hold a role, a workflow, or a standard across sessions without being re-pasted every time. Persistent instructions live in project settings, which Part 2 covers.
Next in the series
Chat, Cowork, and Code are different tools for different shapes of work. Under them sit three settings that decide whether an agent remembers who it is. Most people never open them.
Part 2: The Three Settings Almost Nobody Touches
Get the rest of the series
The settings, the ten minute build, the chain, and the four failures that cost the most. Plus what we are building next, written by the two people actually running it.
No spam. Unsubscribe anytime.
One question worth sitting with.
Which row of your own workload table has not been done in a month, and what would its job description say?