Building is cheap now, so the constraint moved. What is scarce is an idea attached to people you can reach — and no list supplies that.
AI made it cheap to build software, which means the constraint moved. Ideas are no longer scarce and neither is execution. What is scarce is an idea attached to people you can actually reach, and no list of AI SaaS ideas can supply that.
Ideas below are worked examples of the test, not a menu to pick from.
This page is deliberately not a list of fifty ideas. Anything on such a list is an idea a thousand other people read the same week, arriving with no distribution and no domain advantage — how to find SaaS ideas covers the method that produces better candidates.
The specific risk in this category: your product is a thin layer over a model, and the model gets better.
If your entire value is a prompt plus a form, the next capability release absorbs it. That has happened repeatedly to summarisers, rewriters and simple extractors, and it will keep happening.
The products that survive add something the model does not have: a workflow, an integration into where the work actually happens, stored context specific to the customer, or accountability for the output being right.
| Vulnerable | More durable |
|---|---|
| Summarise a document | Summarise every case file and file it into the practice system |
| Generate marketing copy | Generate copy from this client's brand rules and past approvals |
| Answer questions about a PDF | Answer questions across a firm's document history with an audit trail |
The right column has a moat that is not the model: data, workflow position, or a specific vertical's rules. Positioning covers naming what you replace, which is the same question in different words.
Most SaaS has near-zero marginal cost. AI products do not, and founders coming from conventional software often price as though they do.
Work out the inference cost of a heavy user, not an average one. A flat monthly price with unlimited use gives your most enthusiastic customers the ability to cost you more than they pay, and those customers are the ones who stay.
This is a value-metric problem before it is a pricing problem — usage-linked units suit AI products better than flat pricing does. The value metric framework covers the four tests, and pricing a vibe-coded SaaS covers the wider case.
An AI product that produces output nobody can verify has a trust problem it cannot market its way out of.
If a customer cannot judge whether the answer is correct, they will not rely on it for anything that matters — which caps what they will pay. The stronger version either produces output that is obviously right or wrong at a glance, or shows its working.
Vertical products have an advantage here, because the customer usually knows their domain well enough to spot an error immediately. That is another argument for building where you have domain knowledge — finding an idea in your industry covers it.
Not recommendations. Illustrations of the test doing its job.
Meeting notes for a specific profession. Passes 1 and 4 if you know the profession. Test 2 depends entirely on whether you integrate into their practice system or just produce text.
Compliance document checking for one regulation. Strong on 2 and 4 — the rules are specific and the output is checkable. Test 1 depends on whether that profession has a place they gather.
Customer support drafting for small e-commerce. Fails 2 unless it learns the specific store's history and policies. Watch test 3 carefully; support volume is spiky.
General AI writing assistant. Fails 1 and 2 outright. Named here because it is the most commonly attempted and the least defensible.
If an idea passes all four tests, the next step is not building. It is ten conversations with people who have the problem — the build is the cheap part now, which makes validation relatively more valuable than it used to be.
An idea that passes is a candidate, not a plan. Score it against the wider criteria in the five tests, then set a price before you launch rather than after — pricing with no customers covers the method.
And if you have already built something with AI tools and are working out what comes next, the gap is usually the business layer rather than the product. What comes after vibe coding covers that transition.
Four tests: you can name a specific buyer rather than a demographic, the idea survives the underlying model improving, the inference cost per user leaves a margin, and the output is verifiable by the customer. Most AI ideas fail the second.
Because the product is a thin layer over a model and the model gets better. If your entire value is a prompt plus a form, the next capability release absorbs it. Durable products add workflow, stored customer context, integration, or accountability for the output.
Usually with a usage-linked value metric rather than flat pricing. Unlike conventional software, AI products have real marginal cost, so a flat monthly price with unlimited use lets your heaviest users cost more than they pay — and those users are the ones who stay.
Rarely. An idea from a list arrives with no distribution, no domain advantage, and a thousand other readers who saw it the same week. The scarce input is an idea attached to people you can reach, which no list can supply.
Ten conversations with people who have the problem, before building. Because AI has made building cheap, validation is relatively more valuable than it used to be — the expensive mistake is now building the right thing for the wrong people.
Describe what you are thinking of building. Marcus tells you which test it fails and whether that is fixable.
Try GhostCoach free →14-day free trial · cancel anytime · 30-day money-back on Lifetime