AI Guides › Playbooks
By Nigel Guy · 7 min read
Most people test an AI income idea by building it. They set up the shop, the channel or the newsletter, spend a fortnight on it, and only then find out nobody was going to pay. Building feels like progress, which is why it fails so quietly. A better default is to put the idea through a few cheap checks first and only then give it a real week.
The rule: no AI income idea gets a week of your time until it has passed four gates in order, and each gate has a pass mark you wrote down before you started.
This guide is a method, not a progress report. It does not contain anyone's earnings, because none are needed to make the method work, and none are invented here.
The order matters. Each gate is cheaper than the next, and each one kills a different kind of bad idea. Use a single page with the four gates as rows and write the pass mark in before you test.
| Gate | The question | Pass mark (write yours first) | What it kills |
|---|---|---|---|
| 1. Buyer | Can I name a specific person or business who would pay for this? | Three named people or firms, and one of them has said what they pay for the problem today | Ideas that only exist as "people would love this" |
| 2. Hours | After checking and fixing the AI's output, what is my real hourly rate? | A number above your minimum, counted with every minute of review and revision | Ideas that look fast and are actually slow to make safe |
| 3. Rules | Am I allowed to sell this, and what happens to the money? | Platform terms, the AI provider's usage policy, consumer rules and tax each checked once | Ideas that work until an account is closed or a letter arrives |
| 4. Repeat | Can I make the second sale without a new trick? | A second customer, or the same customer again, using the same process | One-off wins and luck |
Start with a person, not a product. Write three names: people or small businesses you could message today. If you cannot, the idea is a hobby, which is fine, but do not count it as income.
Then ask one of them a plain question: "What do you currently pay, or spend hours on, to solve this?" A real answer is a price or an amount of time. A polite "that sounds useful" is not a pass.
You can use an AI model to sharpen the question list, but it cannot supply the answers. Here is a prompt for the preparation step.
You are a sceptical small-business adviser helping me test an idea before I build anything.
Context:
- My idea: [IDEA_IN_ONE_SENTENCE]
- Who I think would pay: [TARGET_BUYER, e.g. "independent estate agents in my town"]
- What I can actually do or know that the AI alone cannot: [MY_EDGE]
Goal: give me 8 short questions I can ask three real prospects to find out whether they already pay money or time to solve this problem, and what a fair price would look like to them.
Rules:
- Questions must ask about past behaviour and current spend, not whether they "would" buy.
- No leading questions and no pitching.
- If [IDEA_IN_ONE_SENTENCE], [TARGET_BUYER] or [MY_EDGE] is vague or missing, ask me for it before answering. Do not guess.
Output: a numbered list of 8 questions, then two lines on which answers would count as a pass and which as a fail.
Before replying, check that no question could be answered with a polite yes.
Fill in the three bracketed fields.
AI makes the first draft fast. The cost is in checking, correcting and delivering it. Do one paid-quality piece of work for a real or imagined customer and time it from start to finish, including reading every line of the output.
Count these separately: time prompting, time checking facts, time reformatting, time communicating with the buyer. Divide the price you would charge by the total. If the answer is below what you would accept for any other work, the idea fails this gate even if the AI part took five minutes.
Tool costs belong here too. Subscription prices change often, so check the current £ price at checkout and include it in the maths. A paid plan only counts as a cost of this idea if you would not otherwise have it.
Do this once, before the first sale, not after it.
A first sale proves little because you can often get one from goodwill. The repeat gate asks whether the same process works again without extra effort or a favour. Run it a second time with a different buyer, or offer the first buyer a second job. Write down what you did differently. If the answer is "I chased them personally for a week", the process is not repeatable yet, and that is useful to know.
Prefer ideas where your own knowledge is the scarce part and the AI saves you drafting time. These tend to survive Gate 2.
Be warier of anything where the product is generic content, because the buyer can make the same thing themselves in a minute.
Imagine a former lettings coordinator who thinks of offering AI-assisted tenancy-listing write-ups for small agents. At Gate 1 she names three agents and asks what they do now. Two write listings themselves in the evening; one pays a freelancer. That is a pass: there is a real spend of time and one real spend of money.
At Gate 2 she times a listing from photos and notes. Drafting is quick, but checking every claim against the notes, fixing tone and sending it takes noticeably longer than she hoped. She sets her price from the total time, not the drafting time, and sees the idea only just clears her minimum. She decides to trial it as a monthly bundle instead of per listing.
Gate 3 turns up a rule she would have missed: listings must be accurate, so every line she cannot verify has to be removed, not polished. At Gate 4 she repeats it for a second agent. Nothing about it needed a favour, so she gives it a proper week. None of these figures are real; the point is the order of the decisions.