One-person agency marketing: letting AI agents handle the routine work
One-person agency marketing is the model where one person manages many advertising clients by having AI agents handle structured, repetitive tasks each week. Most of an advertising agency's time doesn't go to ideas but to reporting, account reviews, and spending progress tracking. According to Yarchi (X, 18/10/2026), these three tasks are exactly what an AI agent does best. This post guides you on how to hand them over to an agent like Claude Code from Anthropic, while you keep only the decision-making part. For example, a weekly report that takes all day can become a draft waiting for you to read at 8 AM Friday.
This post is based on the process "How to build a one-person marketing agency with AI" by Yarchi (@undefinedKi), rewritten in Vietnamese by OPCAIOS, adjusted with examples for the Vietnamese context, and focused on building it yourself with readily available tools. The original post included sponsored content for a paid AI product; that promotional part is not included here.
What stages make up an agency loop?
An agency loop consists of five stages that every client goes through, repeating over and over. According to Yarchi (2026), they are:
- Closing clients: learning about potential clients, preparing a pitch.
- Auditing: finding what's broken in the advertising account.
- Operations: keeping spending on pace, every day.
- Reporting: showing clients what happened and the next step.
- Growth: turning what works into a brief for the next campaign.
Closing clients and growth need analysis and people. The three middle stages—auditing, operations, reporting—are mostly about gathering data, checking it against rules, and writing it into documents. These three stages take up nearly all of the work week, so this is where an agent helps the most. The principle throughout: the agent collects, checks, and drafts. You decide.
Five layers of a system that works
Miss one layer and the system breaks.
1. Access: what can the agent read?
Access is the list of data sources the agent is allowed to read: advertising accounts (Meta, Google, TikTok), Google Analytics, storefronts (Shopify, Haravan, Sapo), and Google Sheets. Start with read-only access, since the agent doesn't need to write anything yet. Next, the simplest way is one MCP connector per platform. For example, a read-only connector for a Meta Ads account is enough for the agent to pull daily spending without being able to change the budget. See the post learning MCP for free with Hugging Face if you're not familiar with MCP yet.
2. Agent: who does the work
You need an agent that can work with files and code, not just chat—like Claude Code or Codex running in a project folder:
agency/
├── clients/
│ └── khach-a/
│ ├── rules.md # goals, limits, when to report
│ └── reports/ # everything the agent produces
├── skills/
│ ├── kiem-tra/SKILL.md
│ ├── nhip-chi-tieu/SKILL.md
│ ├── bao-cao-tuan/SKILL.md
│ └── brief-sang-tao/SKILL.md
└── log/runs.csv # one line per run
3. Memory: what does the agent know about each client?
The rules.md file for each client has the most impact on the quality of the whole system. Every time you manually fix an agent's output, add a line here so it doesn't repeat next time. For example:
Target metric: CPA, target 150,000đ
Budget: 100 million/month, Google + Meta
Never: turn on paused campaigns, touch brand campaigns
Always: note the source and date range next to every number
Report right away: if pacing is off by 10%, any change over 5 million
Report voice: simple language, numbers first
(The numbers above are just examples.)
4. Clock: when does work start?
Automatic schedule (cron, scheduled task) so the agent starts by itself without you having to remind it: 7 AM every day, 8 AM Friday, first Monday of the month.
5. Review gate: what needs you
End every skill with one line: post the draft and wait for approval, never apply changes on your own. This is the layer that lets you sleep well.
Start with one client and one process. Only add the next process once the first one runs smoothly without you needing to step in.
One template for every process
Everything in the agency follows the same seven-step structure. Once you understand it, building a new process takes just a few minutes:
- Trigger: on schedule, when there's a new client, when there's a message.
- Get data: new numbers from connected accounts.
- Check: against the rules in the client's
rules.md. - Make output: a file (spreadsheet, PDF), not an answer in a chat box.
- Review: you agree or you don't.
- Send: to the client or into the ad account.
- Log: one line, so the next run knows what the last one did.
Write this template once as a skill, then copy it for each process below:
---
name: [process-name]
description: [creates what]. Use when [trigger condition].
---
1. Read clients/[client]/rules.md before everything else.
2. Get [data] in [time range] from every connected account.
3. If one source fails, stop and say which one. Never guess a number to fill in.
4. Check the numbers against the rules: targets, limits, when to alert.
5. Make [output]. Each number comes with source and time range.
6. Post the draft and a 4-line summary. Wait for approval.
Never send on your own or change anything on your own.
7. Write one line in log/runs.csv: date, client, what was done, result.
Step 3 is the most important. A process that silently fills in guessed numbers where data is missing is worse than having no process at all.
Close the client: how do you research a prospect before the pitch?
Researching before the pitch is the agent gathering public information about a prospect into a sourced summary. Most pitches fail because of poor prep, and with an agent, prepping each client takes just a few minutes. The process has four parts:
- Trigger: new prospect comes in.
- Input: their website, ads they're running, two or three competitors.
- Output: a sourced summary and a slide outline for the pitch.
- You decide: whether to pitch or not, and which point to lead with.
Public data is enough for most of this: Meta's Ad Library and Google's Ad Transparency Center show every ad a business is running. Here is a sample prompt:
Research [company] before the pitch.
1. What they sell, to whom, at what price.
2. Ads running on Meta and Google: channels, angles, formats,
how long the longest-running ad has been live.
3. Two competitors: where they position differently.
4. Three gaps we can address, each with proof and a link.
5. Outline for a 6-slide pitch built around those three gaps.
Mark each claim as "verified" or "guess". The outline cannot use guesses.
The last line is the most important one: a pitch based on a wrong guess about the client's own business will end very quickly.
Audit: how do you find what's broken in an ad account?
An audit is the agent comparing real numbers against four categories to find places where money is wasted or measurement is wrong. Run it when you sign a new client, then once a month after. According to Yarchi (2026), a good audit is also the most persuasive argument when pitching. The process has four parts: trigger when there's a new client then monthly; input is the ad account, analytics, the store, and search keyword reports; output is a spreadsheet with a health score and one fix for each area; you decide what to fix this week.
Four categories cover most of the ad spend:
- Spend pace: is each channel spending at the right pace for the month.
- Wasted keywords: keywords that cost money but don't bring orders.
- Measurement gaps: conversions counted twice or not counted at all.
- Ad fatigue: the best ads are losing effectiveness.
Audit [client] in [time range].
For each area (spend pace, wasted keywords, measurement, ad fatigue):
status: Fix now (this week), Review (this month), or OK
score from 0 to 100, main finding, estimated monthly impact
proof: which report, which number
one fix suggestion
Add overall health score and total money that could be saved each month.
Export spreadsheet: overview tab, proof tab.
Suggest only. Don't change anything in the account.
## Operations: How do I monitor spend pace every day?
Spend pace is the ratio between what I've spent and what I planned to spend by today, and the agent checks this ratio every morning for every client and every channel. The process has four parts: it triggers at 7 AM every day; inputs are spending by channel, monthly budget, and targets in `rules.md`; outputs are a spend pace table with a proposal for each client that's off track; you review or reject each proposal.
The logic has four steps:
1. **Pace = amount spent ÷ amount planned by today.**
2. Being off by ±10% is on pace; outside that range is overspending or underspending.
3. Each off-track case: one proposal for a change, with reason and numbers.
4. If nothing is off track, no message is sent, so you only hear about clients who need you.
Example: budget of 30 million đồng per month, today is day 10 of a 30-day month, so the planned amount by today is 10 million đồng. If I've already spent 12 million đồng, the pace is 120%, off by 20%, and the agent proposes a cut.
7:00 AM every day, for each client in clients/:
get spending from the start of the month by channel
pace = spending ÷ (budget × current day ÷ days in month)
flag all channels off by more than ±10%
for each flagged channel, propose one budget change: from, to, reason, expected pace
Only send flagged cases. Don't change anything until I review it.
**Next upgrade: build an internal tool once.** Ask the agent to build a small web app the whole team opens every morning, with a client list, pace indicators, daily spend charts, and approval buttons. According to Yarchi (2026), with a coding agent this is an afternoon's work, running on a laptop or a cheap VPS.
## Reporting: what does a weekly report to the client include?
A weekly report is a 3-page document that the agent writes from advertising platform and shop data, for you to review before sending to the client. The report is what the client actually sees from your work, and many agencies spend a whole day making one. The process runs at 8 AM Friday, pulls this week's and last week's numbers, exports a 3-page PDF with a 4-line summary sent to the client's chat group like Zalo or Slack.
Each report answers three questions:
1. **What happened:** a headline in plain language, key metrics compared to last week.
2. **Why:** daily trends and the ads behind them.
3. **What's next:** actions, marking what needs client approval.
Rule to avoid arguments: platforms like Meta or Google often record more conversions than the actual orders the shop really has, so **take revenue and order count from the shop** and cite the source for every other number.
8:00 AM Friday every week, make a weekly report for [client], [Monday to Sunday of last week].
Page 1: one-sentence headline, spend, ROAS, CPA compared to last week.
Page 2: charts of spend and ROAS by day, top 3 ads with results.
Page 3: what changed, why, next steps. Flag what needs client approval.
Revenue and order count from the shop. Every other number clearly sourced.
Export branded PDF and a 4-line summary. Post as draft for me to review.
## Growth: the creative loop
Clients stay when results keep improving. Results improve when each new ad is built on what's already worked.
- **Trigger:** once a month.
- **Input:** all running ads and results from the last 30 days.
- **Output:** a creative brief based on common patterns in winning ads.
- **You decide:** which test to run.
The loop has five steps: get results for all ads → find what the top 20% have in common that the bottom 50% don't (format, opening line, length, angle, talent style) → write a brief → test each variable against the current best ad → if the new ad beats the control by 10% or more, increase budget; if not, pause and log why.
Once a month, for [client]:
Get all ads from the last 30 days with spend, ROAS, and CPA.
Find what the top 20% have in common that the bottom 50% don't:
format, opening line, length, angle, talent style.
Write the next brief:
hypothesis with expected results, 3 opening lines, scenes to shoot,
format, and one test against the current best ad.
Post as draft. I pick which test to run.
What a Week of a One-Person Agency Looks Like
A week of a one-person agency is a working week where five processes have already been scheduled, so you only need to review and decide. According to Yarchi (2026), the template schedule consists of four milestones. Each morning, the expense rhythm table for all clients is ready before you turn on your machine. For each new potential client, the research is completed before the first call. On the first Monday of the month, the account check and creative brief are already written. Finally, by 8 AM Friday, all weekly reports are written and waiting for you to read.
Your work narrows down to three things: reviewing changes, talking with clients, and deciding the next step. For example, with 5 clients, you read 5 draft reports on Friday morning instead of writing each one yourself. The work still gets done, it's just that you're no longer the one who has to start it.
Tasks You Never Fully Delegate to an Agent
Two deciding questions determine who keeps a task: can it be undone, and does it require judgment?
| Does Not Require Judgment | Requires Judgment | |
|---|---|---|
| Can Be Undone | Gather data, write drafts, check, research — agent does it | Sales angle, choose tests — agent proposes options, you choose |
| Cannot Be Undone | Change budget, send report — agent proposes, you review | Service pricing, contracts, accept or drop clients — only you |
Gradually increase autonomy for the agent:
- Week 1: only write drafts.
- Weeks 2 to 4: take action after you review.
- From month 2 onward: make small changes on your own within the limits noted in
rules.md.
Anything that cannot be undone must always go through a review step.
Where Do You Start Building a One-Person Agency?
Start with one client and one process that takes up the most time in your week, usually the weekly report. First, build the process following the seven-step framework. Next, run it in parallel with the old way for a few weeks to compare results. Finally, only add the next process once the first one runs smoothly.
If you're working alone and want to know what to automate first, read more What Should I Automate First If I'm Working Alone? and the AI system I use to run my business solo. Need someone to help you build the right processes for your work, check out OPCAIOS's SOLO program.
Read more: Breaking Down the Operating Process of a Micro-Agency AI; From "Selling Labor" to "Selling Results".
Source: Yarchi — How to build a one-person marketing agency with AI (X, 18/10/2026), rewritten and adapted by OPCAIOS.
Translated from Vietnamese with AI. Read the original.
Want to apply this in your business?
Book a free 30-minute call (Vietnam time, in English or Vietnamese).