How to run a one-person agency with AI agents trained on your craft

An agency owner can move the agency’s repeatable work to AI agents one piece at a time. The owner keeps the judgment, reviews what goes to clients, prices the result instead of the hours, and hires people only where a person adds the most.

This guide shows how to do that in seven steps. It uses the Of One method’s eight starting questions as the practical way in. The full method is published at of1.ai.

What the evidence supports

Two controlled studies show what one skilled person with AI can do, and where it goes wrong.

The first is a randomized trial with 776 professionals at Procter & Gamble. People were assigned to work alone or in pairs, with or without AI. The people working alone with AI matched the performance of the pairs working without it.

The second is a preregistered experiment with 758 Boston Consulting Group consultants. On tasks inside GPT-4’s abilities, consultants using it completed 12.2% more tasks and finished 25.1% faster, with better quality. On a task chosen to fall outside those abilities, the consultants using AI were about 19 percentage points less likely to get it right.

Both results point to the same plan. Give the agents work they are good at. Keep a skilled person checking the output. Move the work over in small pieces so you can see where it breaks.

There is also a money reason to stay small. In Promethean Research’s survey of digital agencies, studios with fewer than 10 full-time employees averaged a 19% after-tax net margin in 2025. Agencies with 50 or more full-time employees averaged 8%. Marketing agency profit margins explains why.

Step 1: Map each service

Start with what you sell. For each service, write down every step from the client’s brief to the final delivery. Next to each step, write who does it today and how long it takes.

Here is a hypothetical example for a monthly paid social retainer. It is an illustration, and the hours are made-up inputs.

Step	Who does it now	Hours a month	Needs the owner’s judgment?
Read the client’s monthly brief and results	Owner	1	Yes
Research competitors and trends	Junior strategist	4	Partly
Write ad copy variations	Copywriter	6	Partly
Resize and adapt creative	Designer	5	No
Build the monthly report	Account manager	4	No
Write the client update email	Account manager	1	Partly
Approve spend changes	Owner	1	Yes

The map shows two things. It shows which steps need your judgment. It also shows which steps repeat every month in nearly the same form. The steps that repeat are the first candidates for agents.

Step 2: Pick what to hand off first

Score each step from 0 to 2 on five questions. Hand off the highest scores first.

How often does it happen? Weekly or more scores 2.
Does it follow clear rules? A template, style guide or checklist scores 2.
Is it easy to check? If you can tell in a minute whether it is right, score 2.
Can a person review it before the client sees it? A draft that waits for approval scores 2.
Is it inside what AI does well? Drafting, summarizing, sorting and reformatting score 2. New strategy with little precedent scores 0.

In the example above, the monthly report and the creative resizing would score highest. The spend decision would score lowest. It moves money, it is hard to undo, and it depends on your judgment about the client.

Break ties by choosing the task that frees the most hours for your best people.

Step 3: Ask the eight starting questions

Before you build anything, answer the Of One method’s eight starting questions for the first task. The method page gives a plain warning: “Every question you skip becomes a guess treated as a fact.”

Here are the questions, with hypothetical answers for the monthly report.

#	Question	Hypothetical answer for the monthly report
1	Decider: who makes the decision?	The owner
2	Goal: what outcome do they want?	Every client gets an accurate report by the third business day of the month
3	Facts: what is true right now?	Each report takes about 4 hours and goes out around day 7. The data comes from two ad platforms.
4	Limits: what are the hard limits?	Client data goes only into approved tools. No report reaches a client without a person’s review.
5	Levers: what can they actually change?	The report template, the data the agent can read, and who reviews
6	Causes: what will shift the outcome?	If the agent drafts from the template and the platform exports, review time replaces build time
7	Proof: how will they know it worked?	Days to send, minutes of review per report, and errors caught at review
8	Timing: when must they decide?	After three monthly cycles, decide to keep, fix or stop

A question you cannot answer is your first gap. Fill it before you build.

The method then asks you to lay out a map of what you know. Two parts of that map matter most here. A kill test is the result that would prove a claim wrong. A human gate is a move that a named person must approve.

Step 4: Train the agents on your craft

A general AI tool has never seen your work. Your clients pay for how you work, so the agents need to learn it.

Give each agent four kinds of material:

Your briefs. Real past briefs, with the questions you asked and the answers you got.
Your past work. Finished pieces you were proud of, with a short note on why each one worked.
Your edits. The before and after of drafts you corrected. Edits show the agent your standards on real work.
Your rules. The style guide, the brand voice for each client, and the things you never do.

Keep collecting edits after launch. Each time you or a reviewer fix a draft, save the before and after. Add them to the agent’s material each month. Marketer of One is built so agents trained on the owner’s craft get better at marketing every month, instead of educating a team that forgets.

Step 5: Keep a review step

Every agent needs a named person who checks its work. In a one-person agency, that person is usually you. As you grow, it can be an ally you trust.

Set three rules for each task:

A draft waits for approval. Nothing reaches a client until the reviewer says yes.
Anything hard to undo stays with a person. Contracts, public posts, ad spend and anything a client pays for need a human gate.
A kill test sends it back. Decide in advance what result means the agent is not ready. For example, a hypothetical rule might be that if review still takes more than half the old build time after three cycles, the task goes back to a person while you fix the training.

Track your review minutes. When they fall and the error count stays low, the agent is learning. When they rise, look at the training material first.

Step 6: Price the result instead of the hours

If you bill by the hour, every hour an agent saves comes off your invoice. The client gets the whole saving.

In SoDA’s survey of 251 agency leaders in late 2025, 39% said clients already expect lower prices because of AI.

Pricing the result protects the margin. Sell a monthly report, a campaign or a set of deliverables at a fixed price. You can also price a retainer by what the client gets each month. When the agents cut the time, the saving becomes margin or room for another client.

Promethean Research makes a similar point. It warns that agencies that treat AI as a cheaper way to produce the same deliverables may see the savings competed away.

Step 7: Hire allies where a person adds the most

Once the repeatable work runs through agents, look at what is left. What is left is often strategy, client relationships, creative judgment and the review step itself.

Hire for those. An ally is one of the very best people the market offers, not an employee hired to fill hours. Pay for talent instead of hours. Add a person when a specific gap costs you clients or quality.

When you do hire, give the new person the same map, the same review rules and the same training material. They start from your way of working on day one.

What 15 to 20 clients takes

Marketer of One is built so one person can run 15 to 20 custom clients at a time. That is Utlyze’s design goal for the system. It is not a measured result, and the real number for your agency depends on your services, how repeatable the work is and how well the agents are trained.

Getting close to that goal needs three things in place. Most services must be mapped and mostly handed off. The review step must be fast and reliable. The pricing must reward results so that more clients add margin instead of more hours.

Your first month

Use this checklist to start.

Map every service from brief to delivery, with who does each step and how long it takes.
Score each step on the five questions and pick one task to hand off first.
Answer the eight starting questions for that task.
Collect ten past briefs, ten finished pieces and ten sets of your edits for training.
Name the reviewer and write the kill test.
Run the agent in draft mode for one full cycle and log review minutes and errors.
Rewrite the price for that service around the result.
Pick the second task.

You can walk one decision through the eight questions by hand at of1.ai/ask. If you want help with the first tasks, Marketer of One offers Consult at $400 an hour with a 5-hour minimum, and Build from $15,000 a month with a dedicated team on call 24/7.