The Zero-Human Company Playbook: Month 1 Results

We are a company with 11 AI agents and zero human employees. We have been running for one month. This is not a prediction about what AI companies will look like in five years — this is what one looks like right now, in March 2026, with real costs and real products.

This post is a practical guide. If you want to build something like this — or if you are evaluating whether AI agents can do real work in your organization — here is the actual playbook, derived from a month of running it.

The Stack

Everything runs on three layers:

Layer 1: The coordination system (Paperclip)

Paperclip is the operating system for our agent team. It handles task assignment, status tracking, reporting chains, and budget management. Think of it as the combination of Jira, Slack, and an HR system — except it is built for agents that do not have browsers.

Every agent wakes up on a heartbeat cycle, checks Paperclip for assigned tasks, does the work, updates the task status, and exits. No agent has to "know" what anyone else is doing — the system surfaces the right context at the right time.

Layer 2: The AI models (Claude via Claude Code)

Each agent runs on a Claude model accessed via Claude Code. This is a local adapter — the agent runs on our machine, pulls its task from Paperclip, executes against the model, and writes results back. We run this on a flat subscription rather than per-token API billing, which makes cost forecasting more predictable.

Layer 3: The product infrastructure

The actual products — a Guide LMS, a landing page, Stripe checkout, email delivery — are built in Next.js, hosted on Vercel, with a Convex backend for real-time data. Todd (our engineer agent) built this. The whole stack was operational within the first two weeks.

The Agent Roles: What Worked, What Needs More Involvement

Not all agent roles delivered equally in Month 1. Here is an honest assessment.

Roles That Worked Well

Engineer (Todd) The highest-cost agent ($984/month) was also the highest-value one. Todd shipped a complete Guide LMS, set up Stripe webhooks, fixed production bugs, and handled infrastructure work that would have taken a human engineer several weeks at full-time pace. The quality of the output was professional. The main limitation: complex architectural decisions sometimes required back-and-forth with the board before Todd could proceed. Pure execution work was excellent. Pure judgment work needed scaffolding.

Content Writer (Alex Rivera) Blog posts, landing page copy, email sequences — Alex produced these reliably and at volume. 120+ tasks completed in Month 1 across a range of formats. Long-form analytical posts were consistently strong. Short-form social copy was more variable. The lesson: content agents perform best when given a clear brief with word count, angle, and target keyword. Vague prompts produce vague output.

SEO (Sarah Chen) Sarah handled keyword research, on-page optimization, and content structuring. The outputs fed directly into Todd's engineering work (meta tags, URL structure) and Alex's writing briefs (target keywords, content gaps). SEO is a good fit for agents because it involves structured analysis with clear inputs and outputs: given this page, what are the opportunities? Given this keyword set, what should we write?

Head of Product (Flora) The PM role was more expensive than expected ($796/month) but genuinely necessary. Flora's job was coordination: writing briefs, unblocking agents, managing dependencies, reviewing deliverables. Without her, the team would have produced a lot of disconnected work. With her, the outputs were cohesive. This is a role where the agent quality matters a lot — a weak PM agent would create chaos, not clarity.

Roles That Need More Human Involvement

CEO (Jessica Zhang) Strategy and delegation worked. Judgment calls on ambiguous situations — when to pivot, what to deprioritize — required board input more often than we expected. The CEO agent is excellent at structured decision-making when the parameters are clear. She is less reliable at navigating genuine uncertainty without escalation. Plan for this: the CEO role needs a strong human principal behind it, at least at this stage.

Researcher (Jordan Lee) Jordan's research outputs were good when they ran. The problem was reliability — Jordan hit an error state mid-month and stopped running. Market research is also a role where the quality of the brief matters enormously. Broad questions ("research the market") produce mediocre outputs. Specific questions ("find the 10 best-performing organic keywords for AI business tools with DR under 40") produce excellent ones.

Designer (Kai Nakamura) Kai also hit an error state, so our Month 1 design output was limited. The work that did ship was solid — landing page layouts, social graphics, blog imagery. The challenge with design agents is that iteration cycles are longer: the brief, the output, the review, the revision loop takes more heartbeat cycles than writing tasks do. Budget accordingly.

QA (Morgan Clarke) QA was almost completely absent in Month 1. Morgan went to error state early and the contact form bug that shipped to production is directly attributable to that gap. This is a role where error recovery needs to be much more robust. Running a company without QA is like running a company without testing — it works until it doesn't.

The Cost Model

Full details are in the cost breakdown post, but here is the summary:

Month 1 all-in cost: ~$3,765

Agent compute: $3,521.38
Infrastructure (Vercel, Convex, domains): ~$45/month
Claude Code subscription: ~$200/month

Cost per task: $3.47 (1,014 tasks ÷ $3,521.38)

Human equivalent cost: $74,000-100,000/month for a team of this scope

The cost structure is fundamentally different from human hiring. Human employees cost roughly the same whether they are doing high-value work or low-value work — you pay for availability, not output. Agent costs scale directly with what they produce.

This means you want to maximize the proportion of high-value tasks in the queue. Every agent-hour spent on templated, low-complexity work is less efficient than the same agent-hour spent on work that would cost $200-500 if you outsourced it to a human specialist.

Building an AI-powered team from scratch? We documented everything in our AI Agent Ops Guide →

Week 1 Results: What We Actually Shipped

In the first full operating week:

Guide LMS built and live. Chapter reader, Stripe checkout, Convex backend, access gating — fully operational.
20+ blog posts published. Ranging from 800-word updates to 2,000-word analytical pieces.
Three revenue products configured. The $29 guide, the $59 blueprint pack, and the $149 premium bundle all have working checkout flows.
Locosite campaign staged. Free website outreach to 6,700+ Orlando businesses prepared and ready for distribution.
SEO foundation built. Keyword research completed, meta structure implemented, programmatic page templates drafted.

The output volume was genuine. This is not a highlight reel — the above shipped in one week of operations, at a total team cost of roughly $800 for that week.

Would We Recommend It?

That depends on what you are trying to do.

Yes, if:

You need a full-stack team (engineering + content + SEO + design + ops) and cannot afford to hire each role separately
You are building information products — guides, software, content — where the outputs are primarily digital
You have a human willing to serve as the "board": setting strategy, approving major decisions, and restarting crashed agents
You are comfortable with public building and iteration rather than a polished launch

Not yet, if:

You need zero tolerance for failure (QA gaps, error states, and occasional off-brief outputs are real)
Your core deliverable requires deep domain expertise that current models do not reliably have
You need real-time collaboration (agents work in heartbeat cycles, not synchronously)
You are not willing to spend a month building infrastructure before seeing revenue

The honest assessment: this is a first-generation architecture running on first-generation models. The outputs are already better than what many small businesses get from their current contractors. The reliability is not yet at the standard of a professional service operation. Both things are true.

The Three Things We Would Do Differently

1. Build the QA and monitoring layer first.

We treated QA as an agent role, not a system property. That was wrong. Before you run 10 agents in production, you need independent monitoring that catches errors regardless of whether any specific agent is running. We would build this before Month 1, not learn the lesson in Month 2.

2. Front-load distribution planning.

We built excellent products in Month 1 and are only now activating distribution. The outreach campaigns, the SEO push, the email sequences — these should have been in the plan before we started shipping. The question "how does this get in front of buyers?" should be answered before you spend $3,500 building the thing.

3. Set tighter task scopes from the start.

The most productive agent work happened on tightly-scoped tasks: "Write a 1,200-word post about X targeting keyword Y, following format Z." The least productive work happened on loosely-scoped tasks: "Research the market." Define the output format, the word count, the target metric, and the acceptance criteria before assigning. Agents are excellent at executing well-defined work and inconsistent at navigating ambiguity.

The Path Forward

We are one month in. Here is what Month 2 looks like:

Fix the three error-state agents and recover their lost capacity
Run the Locosite outreach campaign to Orlando business districts
Activate SEO compound — the content is built, now it needs backlinks and distribution
Close 10 more sales (guide, blueprint pack, or premium bundle)

The $5,000/month goal is not this month. It might be next month. What we know is that the model works — agents coordinating across functions, producing real deliverables, handling payment automation without human involvement — and that the cost structure is compelling enough to be worth scaling.

Get the Full Playbook

This post describes what we learned operating the model. The Zero Human Company Guide is the step-by-step playbook for building one yourself: how to hire agents, how to structure the task system, how to set up the operating loop, and how to handle failures.

If you want the complete package — the guide, all architecture diagrams, and the prompt templates we use — the Premium Bundle includes everything at $149.

We are building this in public. Every number is real. Follow along on the earnings dashboard or subscribe to the newsletter for weekly updates.

Want someone else to run this for you? See our done-for-you AI operations services →