One-person business

Running a one-person company with AI agents: a 30-day plan

Spend week one putting everything in one place, week two on one agent for follow-ups, week three on one agent for research and drafts, and week four on review and a weekly rhythm you can keep.

A one-person company has one bottleneck: you. Every open loop and every "I will get back to you" sits in the same head. AI agents do not remove the bottleneck, but they can take the re-reading, the chasing and the first drafts off it. This is a 30-day plan to get there without a mess.

A note on intent. If you searched "how to start a small business", this page assumes you already have one. For starting one, the US Small Business Administration's business guide covers planning, registering, funding and licences in plain language; your own country will have an equivalent.

The example setup

The pattern we describe is the one behind Amili. One solo founder runs a job, co-runs a small training studio and builds a product, with AI agents working day and night in Claude Code, Anthropic's tool for running Claude against files, commands and connected tools. We keep it general; the pattern is what transfers: everything the agents need lives in files they can read, each agent has one job and a written brief, nothing is sent, paid or deleted without a human yes, and a short review happens every week. The plan below builds exactly that.

Week 1: capture everything in one place

This week you add no agents. An agent that cannot see your open loops, prices and promises will invent them. Pick one place: a folder of plain text files (what Claude Code reads best), or a project in Claude or ChatGPT for chat use (Claude's help page says projects hold files and instructions that every chat in the project can use). Then fill it.

• Write a one-page "about my business" file: what you sell, to whom, prices, hours, what you do not do.
• Write a "how I write" file: five real messages you sent that sound like you.
• Dump every open loop into one list: promises made, replies owed, decisions pending. Date each one.
• Collect the ten questions people ask you most, with your standard answers.
• Write the floor: no paying, no sending as you, no passwords, nothing irreversible.

End of week check: can a stranger read these files and answer "what is open right now?" correctly? If not, the files are not done.

Week 2: one agent for follow-ups

The first job has the clearest payoff and the lowest risk: keeping the open-loop list true and drafting the follow-ups.

Brief it to read the open-loop list every morning, mark anything overdue, and draft the chase message for each one in your tone. It drafts; you send. In Claude Code this is a CLAUDE.md file with the rules plus a scheduled run; in a chat tool it is a project with the same instructions and a habit of opening it at 9:00.

• Write the follow-up brief: inputs (the list), output (a dated list of drafts), floor (draft only).
• Run it by hand for five days. Compare its "overdue" list to your own memory each day.
• Send at least three of its drafts after editing, and note what you changed.
• Add the repeated corrections to the brief.
• Only then: schedule it, if your tool can.

End of week check: did it miss an open loop you knew about? If yes, the capture is incomplete; go back to week 1 for that item.

Week 3: one agent for research and drafts

The second job is heavier: comparing options, reading long documents, writing first versions. The risk changes from "missed" to "invented", so the checks change too. Give it bounded tasks with a required output shape: "compare these three tools on price and reminders, link every claim to the vendor's page, mark anything you could not verify". Treat every number and link as unverified until you open it.

• Write the research brief: always link sources, always mark "check" on unverified claims, never guess numbers.
• Give it one comparison task and open every link it returns.
• Give it one draft task from your bullets; save your edited version back into the files as a style example.
• Keep the follow-up agent running; do not let the new job crowd out the first.
• Note where it wasted your time, and shorten the brief to prevent it.

End of week check: did you catch at least one invented or unverifiable claim? You almost certainly will; that is the check working.

Week 4: review, then a weekly rhythm

The last week turns a month of experiments into a routine. Without it the agents drift and you stop trusting them.

• Reread every brief. Delete rules that never mattered; add the corrections you made three times.
• Read the whole open-loop list once, by hand. Close what is done, re-date what slipped.
• Decide whether a third job is worth adding (bookkeeping prep, meeting notes) or whether two is enough. Two is often enough.
• Put a 30-minute weekly review in your calendar: briefs, open loops, what to stop.
• Write down what you still do by hand on purpose, so you do not drift into automating it.

The weekly review is also where you check cost: the API bill if your agents are metered, the usage limits if they run on a subscription.

What does this not solve?

Agents do not give you more judgment, more clients or more energy. They do not notice what you never wrote down, and they cannot carry responsibility for a promise or an invoice. The point of the follow-up agent is that you send more of the messages you meant to send, not that someone else sends them.

Amili fits weeks 1 and 2 of this plan rather than all of it: you tell it things as they come up (a sentence, a voice note, a photo), it keeps the open loops and drafts the follow-ups, and it never sends or spends without your yes. It is in private beta from 14 October 2026, by invitation and free, with most integrations still coming, and it does not do the deep research or writing of week 3. See how it works.

Questions people ask

Do I need to know how to code to run AI agents?

No, if you stay with chat tools and projects: the briefs in this plan are plain English. Claude Code gives you more (scheduling, reading your own files, connecting tools) and takes an afternoon to learn. Start with chat; move only if you hit a limit.

What is the difference between an AI assistant and an AI agent?

An assistant answers when you ask. An agent runs a task on its own, often on a schedule, with access to files or tools, and comes back with a result. The follow-up agent in week 2 is the simplest agent: same assistant, fixed brief, runs every morning.

Can the agents talk to my clients directly?

They can draft, and we recommend they stop there. A wrong promise cannot be unsent, and clients of a one-person company are paying for that person. Nothing goes out in your name unread.

What if I fall behind during the 30 days?

Stay on the week you are on. The plan is a ladder, not a calendar. A six-week version with the same order beats a four-week version with gaps.

About this page. Written by Amili, an AI assistant. Sources are linked in the text. Last updated: 2026-10-06.