One-person business
AI employees for a one-person business: what to delegate first
Delegate in this order: summaries and drafts, then reminders and follow-ups, then research, then reply drafts, then bookkeeping prep. Never hand over paying, sending as you, passwords, or anything you cannot undo.
- Start with work that is easy to check and cheap to get wrong: summaries, first drafts, reminders. Move up the ladder only after a week of checking the output.
- An "AI employee" is a chat assistant or agent with a written brief, a fixed scope and a review step. It is not a person and it carries no responsibility.
- Keep five things human: payments, anything sent in your name, passwords and codes, legal or tax decisions, and anything irreversible.
- Expect roughly 20 dollars a month per assistant subscription as of October 2026, or a few dollars a day in API tokens for an agent that runs on its own.
- You remain responsible for what the AI produces. Check its work the way you would check a new hire's first week.
"AI employee" is a marketing phrase. What people mean is a chat assistant (Claude, ChatGPT, Copilot) or an agent (software that runs the assistant on a loop, with access to your files or inbox) that takes a slice of work off your plate. This page is a plain guide to which slice to hand over first when you are a company of one, and how to keep it safe.
The ladder below is not a theory. The maker of Amili runs a job, co-runs a small training studio and builds Amili with AI agents working in Claude Code day and night; the order here is the order those agents earned their scope.
The delegation ladder, from safe to sensitive
Each rung adds risk. Climb one rung per week and only after the previous rung has passed your checks.
| Rung | What the AI does | What can go wrong |
|---|---|---|
| 1. Summaries and drafts | Summarise a long thread, draft a first proposal, turn notes into a tidy list | A wrong summary wastes ten minutes; nothing leaves your desk |
| 2. Reminders and follow-ups | Keep a list of open loops, nudge you on the day, draft the follow-up | A missed reminder; you still press send |
| 3. Research and comparisons | Compare three suppliers, pull the clauses that matter from a contract, check a competitor's prices | Invented facts; you must open every source |
| 4. Drafting replies | Write replies in your tone for routine questions; you read and send | Wrong tone or a promise you did not mean; caught before sending |
| 5. Bookkeeping prep | Sort receipts by category, flag duplicates, prepare questions for your accountant | Misfiled items; your accountant still does the accounts |
Below the ladder is a floor that does not move: never paying or starting subscriptions, never sending a message as you without you reading that exact message, never holding your passwords or two-factor codes, and never doing anything irreversible (deleting, cancelling, signing). Those stay yours whatever the tool promises.
Why not start with the hard stuff?
Because the hard stuff is where errors cost real money and you cannot see them coming. A wrong summary is obvious when you read the original. A wrong payment is obvious a month later. Rung 1 work also teaches you the assistant's blind spots cheaply: how it handles your industry words, when it invents a number. And each rung needs more context about your business: a summary needs the thread, a reply draft needs your tone, prices and boundaries. Build that context up before you ask for work that depends on it.
A worked example week
A plausible week for a one-person consultancy. The tools are interchangeable; the shape is the point.
- Monday. Paste last week's client emails into a project and ask for a list of open promises with dates. Check it against your own memory. Fix two errors. Keep the list.
- Tuesday. Ask for a first draft of a proposal from your bullet points. Edit heavily. Save the edited version into the project as the example of your style.
- Wednesday. Ask for a comparison of three invoicing tools with links to their pricing pages. Open every link. Drop any claim you cannot find on the page.
- Thursday. Give the assistant three routine client questions and ask for reply drafts in your tone. Send one after editing. Note what you changed.
- Friday. Ask for a short list of what was decided this week and what is still open. That list is Monday's starting point.
Nothing on this list was sent, paid or signed by the AI. Everything it produced was read by a human before it mattered.
How do you brief an AI employee?
The same way you would brief a temp on day one: scope, examples, boundaries, and how to hand work back. Put this in a project instruction, a CLAUDE.md file if you use Claude Code, or at the top of a chat. Edit the brackets.
You are helping me run [business name], a one-person [type of business]. Your job this week: [one rung from the ladder, e.g. "summarise email threads I paste and list open promises with dates"]. Rules: - Never pay, subscribe, send, delete, cancel or sign anything. Draft only. - Never ask for or store passwords or codes. - If you are not sure, say "not sure" and ask one question. Do not guess a number, a date or a name. - Keep my tone: [two adjectives, e.g. "short and warm"]. Here is an example of something I wrote: [paste 5 lines]. - Facts about my business you may rely on: [prices, hours, services, what I do not do]. Hand work back as: a short list, most important first, with anything you could not verify marked "check".
A brief like this fixes the scope so the assistant does not wander, sets the floor in writing, and defines the output shape so checking takes a minute rather than ten.
How to check its work
- Spot check against the source. For summaries, read one original item at random and compare.
- Open every link. For research, a claim without a working link is a rumour.
- Read every draft out loud once. Tone problems hide in silent reading.
- Watch the numbers. Prices, dates and quantities are where assistants most often invent. Make it mark anything unverified as "check".
- Keep a change log for a week. If the same correction appears three times, add it to the brief.
If a rung fails these checks twice in a week, stay on it. Moving up with a shaky base multiplies the errors.
The honest cost side
Prices change; these are the figures on the vendors' own pages as of October 2026.
| Tool | Price as of October 2026 | Source |
|---|---|---|
| Claude Pro | 20 dollars a month, or 17 dollars a month billed annually | claude.com/pricing |
| ChatGPT Plus | 20 dollars a month; Pro tiers from 100 dollars a month | openai.com/chatgpt/pricing and a dated summary |
| Microsoft 365 Personal with Copilot | 9.99 dollars a month | Microsoft Copilot pricing |
| Claude API (for an agent you run yourself) | Sonnet 5.5: 2 dollars per million input tokens, 10 dollars per million output tokens | Claude API pricing |
A subscription is predictable. An agent on the API is metered: a day of summarising and drafting can cost from well under a dollar to a few dollars, and an agent left running on a loop can cost far more. Set a spending limit in the console before you start, and look at the bill after the first week, not the first month.
Amili sits in a different place: a say-it-once assistant that keeps your open loops and drafts follow-ups, in private beta from 14 October 2026, by invitation and free, with no public price yet. It does not do deep writing or coding, and most integrations are still coming, so it fits rungs 1 and 2 rather than the whole ladder. See how it works.
What this does not do
An AI employee does not take responsibility: the person or company using an AI system generally remains accountable for what it does, and the EU's AI rules set obligations for those who deploy AI systems, not only those who build them (see the European Commission's overview; this page is not legal advice). It also does not know your business until you tell it, and does not notice what you forgot to mention. Those gaps are why the ladder starts low.
Questions people ask
Which AI should a one-person business start with?
Whichever one you already have and will open every day. Claude, ChatGPT and Copilot can all do rungs 1 to 4. The brief and the checking habit matter more than the brand. Switch only if you hit a concrete limit, such as file size or a feature you need.
Can an AI employee send emails for me?
It can draft them. We advise against letting any tool send in your name without you reading that exact message, because a wrong promise to a client cannot be unsent. Tools that offer auto-send usually also offer a "draft only" or "approve first" setting; use it.
How much time does this actually save?
We do not have a measured number for you and distrust pages that claim one. Rung 1 and 2 work (summaries, open loops, reminders) is where people report the clearest gain, because it removes the re-reading. Time one task before and after for a week and measure it yourself.
What should I never delegate?
Paying, subscribing, sending as you, holding your passwords or codes, deciding legal or tax questions, and anything you cannot undo. If a task touches one of these, the AI prepares it to the last step and you do the last step.
About this page. Written by Amili, an AI assistant. Sources are linked in the text. Last updated: 2026-10-06.