AI Guides › Playbooks
By Nigel Guy · 6 min read
Most people who want to run a one-person AI company start with the tools. They pick a model, open a no-code builder, and spend a month on something with no named buyer. It feels like progress because things get built, and it fails because nothing was ever sold.
The rule: pick the model by who pays and what they pay for, then give it a fixed test with a kill threshold before you build anything.
A note on the headline claim that solo companies earning seven figures are growing fast. Search results quote US Census figures and secondary reports, but I could not tie them to a primary source I could check, so this guide does not rely on any number. The durable part is the mechanism: AI lowers the cost of producing software and service work, so one person can serve a narrow market that a payroll could not.
| Model | You sell | Income shape | Main catch |
|---|---|---|---|
| 1. Narrow-niche AI product | Software access, usually a subscription | Slow start, recurring | Support, and a model provider that can change terms or prices |
| 2. Automation agency | Built workflows plus a monthly retainer | Faster start, tied to your hours | Capacity: you are the factory |
| 3. Small product portfolio | Several modest products at once | Spread risk | Attention gets thin and none gets finished |
Fill in one card per idea. If you cannot fill a line in one sentence, you do not have an idea yet, you have a mood.
Narrow beats clever. A tool that drafts one document type for one profession can win because the buyer recognises their own work in it. Start with the card, then sell before building: describe the outcome to ten people in that profession, and take payment or a signed pilot commitment for a rough version. Build the first version as a prompt, a template and a simple form, not a platform.
Costs at time of writing: Claude Pro is listed at $20 a month (about £15 at time of writing, so check the £ price at checkout), and Claude Max starts at $100 a month (about £75). If you build on the API, usage is billed per token and varies by model, so price your product against your measured cost per customer, not a guess.
You sell a finished workflow for a business you can name, such as chasing late invoices or sorting enquiries, plus a monthly fee to keep it running. n8n's cloud Starter plan is listed at €20 a month billed annually (about £17 at time of writing), with 2,500 executions, and the self-hosted community edition is free on GitHub, though you then run and secure it yourself. Check the live page, because plans and limits move.
The usual trap is selling hours in disguise. Fix the scope of each build, write down what it does not do, and charge a retainer for monitoring and fixes so income does not reset to zero each month.
A portfolio only works if each product costs almost nothing to run and each gets a hard test. Cap yourself at two or three live experiments. Review them on the same day each week and apply the kill threshold you wrote. Products that pass get your time; products that fail are closed, not "paused".
Say you spent years as a bookkeeper and know that independent physiotherapy clinics send patients appointment reminders by hand. Your card: buyer is a clinic owner; painful job is reminders and follow-up messages; AI drafts the messages, a human approves them; price is a monthly fee; test is five clinic conversations in week one and three paid pilots by week four; kill threshold is fewer than three. You pick model two first, because it tests demand fastest, and only consider turning the repeated workflow into a product (model one) once three clinics ask for the same thing.
Fill in the square brackets. Run each prompt in a fresh chat.
You are a sceptical advisor to a UK sole trader choosing a first AI-powered business.
Goal: stress-test one idea so I can decide in an hour whether to test it.
My inputs:
- Buyer: [JOB_TITLE_AND_BUSINESS_TYPE]
- Painful job: [TASK_THEY_DO_REPEATEDLY]
- My relevant experience: [YOUR_BACKGROUND]
- Model: [PRODUCT / AGENCY / PORTFOLIO_ITEM]
- Weekly hours I can give: [HOURS]
If any input is missing or vague, ask me before answering. Do not guess.
Steps:
1. Restate the idea in one sentence.
2. List the three assumptions that must be true for someone to pay.
3. For each, name the cheapest way to check it in under a week.
4. Say what AI should do and what a human must still check.
5. Propose a kill threshold with a number and a date.
Format: a table for steps 2-3, then short bullets. No hype, no invented market
statistics; say "unknown" where you cannot know.
Before answering, check that every assumption is testable and that you have
invented no figures.
You are a practical workflow designer. I run a one-person automation service.
Design a first workflow for this client: [CLIENT_BUSINESS_TYPE].
The task to automate: [TASK_DESCRIPTION].
Tools I plan to use: [TOOLS_E.G._N8N_OR_ZAPIER_AND_EMAIL].
Data involved, including any personal data: [DATA_TYPES].
Ask me for anything missing. Then give: (1) numbered workflow steps; (2) where a
human approves before anything is sent; (3) what could fail and how I would
notice; (4) what this workflow explicitly does not do, to put in the scope
document. Output as a numbered list plus a short risks table. Flag any step that
handles personal data. Before answering, check each step has an owner, a trigger
and a failure alert.