What an AI startup idea actually is, and the two kinds
The term means two different things: one is gated by a model capability and dies when the vendor ships it, the other is a margin bet. How to tell them apart.
- AI Startup Idea
- An AI startup idea is a specific claim that a named set of buyers will pay for a workflow that a recent model capability made either newly possible or newly profitable.
Where the phrase comes from
"Artificial intelligence" was coined for the 1956 Dartmouth Summer Research Project proposal, written by McCarthy, Minsky, Rochester and Shannon. Both halves are older than the search volume. "Startup idea" became a term of art through Paul Graham's November 2012 essay, which split ideas into organic ones — you noticed a gap because you already live in the domain — and made-up ones, which come from sitting down to brainstorm. Steve Blank's January 2010 definition supplies the other half: a startup is a temporary organisation searching for a repeatable, scalable business model. The idea is hypothesis one in that search. Nothing more.
The compound is recent, and it is a query rather than a concept. People type "AI startup ideas" and get theme lists. Y Combinator's Requests for Startups page is the best-known version, and it is honest about what it is: areas the partners want to fund. A theme is not an idea. "AI for legal" is a theme. "Submit freight damage claims into carrier portals for 3PLs who currently pay a clerk to do it" is an idea, because it names a buyer and a workflow.
The term means two different things. Label them.
I see these collapsed constantly, and they fail in opposite directions.
Type A — capability-gated. The business was not buildable before a specific model capability existed. You can name the API and the month. It dies when the model vendor, or the system of record, ships the thing you were charging for.
Type B — margin-gated. The business was always buildable, with people. A model changes the cost of delivery, not the possibility. It dies when the unit cost at list token prices exceeds what the buyer will pay, which is discovered somewhere around customer 500, not customer five.
Type A fails loudly and fast. Type B fails quietly over a year of decaying gross margin. I have taken money for both — Metadata.io was mostly Type B with a Type A wedge — and given the choice I take the Type B with defensible unit economics every time. A Type A moat is a six-month lead on an API that everyone else is also reading the docs for.
What it is not: an "AI use case"
The neighbouring term is AI use case, and the two get swapped in nearly every deck I read. A use case is a task a model can perform: summarise this transcript, classify this ticket, fill this form. An idea is a use case plus three things a use case never contains — a buyer with a job title and a budget line, a reason the incumbent software vendor has not already shipped it, and a cost per completed unit below the price.
Two adjacent terms need separating too. An AI wrapper is a packaging judgement, not a category of idea; most durable software wraps something. A Request for Startups is a theme list, which is to say a list of places to start looking, not a list of ideas. "Agentic" carries two incompatible claims, which I've pulled apart separately. The same labelling problem shows up in "AI automation".
Worked example: freight claims into portals with no API
File damage and overage claims into carrier web portals on behalf of third-party logistics firms, who today pay clerks to retype data out of a TMS and into a browser.
Screenshot-driven computer control shipped as a public beta API on 22 October 2024. Before that date you built per-portal scrapers that broke on every layout change, which is why nobody sold this. That date is what makes it Type A.
The kill condition is equally datable. If a carrier ships an API, or an integration arrives through something like the Model Context Protocol (introduced 25 November 2024), the automation collapses into a normal integration and the price collapses with it. Computer use is a last resort, and pricing a business on a last resort is a bet with an expiry date.
Then the arithmetic. Assume 40 model turns per claim, 15,000 input tokens and 1,500 output tokens per turn — screenshots are heavy, roughly (width × height) / 750 tokens each, so a 1092×1092 capture is about 1,590 tokens before any conversation history. At Claude Sonnet list prices of $3 per million input and $15 per million output: 600,000 input tokens = $1.80, 60,000 output tokens = $0.90. $2.70 per clean run. Now assume one in three claims needs a retry and one in five escalates to a human reviewer, and the loaded cost lands north of $5. Prompt caching pulls the input side down, since cache reads bill at a fraction of base input price. The question stops being about vision and becomes arithmetic: does a 3PL pay more than $5 per claim, and how many claims do they file a month? Build the margin meter first.
The qualification test
| Test | Pass looks like | What it catches |
|---|---|---|
| Date | You can name the model, API or release month that made it buildable | Type B wearing Type A's clothes |
| Kill | You can state the single event that ends the business | Ideas with no falsifiable failure mode |
| Margin | Cost per completed unit at list prices, including retries, under the price charged | Demos that never survive volume |
| Buyer | A job title, a budget line, and what they pay for this today | An "AI use case" with no purchaser |
| Loop | The work terminates under a step cap enforced outside the model | Agents that bill forever and finish nothing |
Founders skip the last row. Unbounded loops are the most common production failure in agent products, and the fix is not a better prompt — it is a hard step cap and a budget ceiling enforced in the harness, outside the model's control.
An idea that passes the date test and fails the margin test is a feature, and usually someone else's. An idea that passes the margin test and fails the date test is ordinary software, which is a fine thing to build and should be funded as such — at software multiples, by people who know that is what they bought. The ones worth your next two years pass both, and you can say out loud what would kill them. Profitability among AI startups clusters exactly where both of those questions got answered honestly.
Sources
- Paul Graham, "How to Get Startup Ideas" (November 2012) — Origin of 'startup idea' as a term of art; the organic vs. made-up distinction.
- Steve Blank, "What's A Startup? First Principles" (25 January 2010) — A startup as a temporary organisation searching for a repeatable, scalable business model.
- Dartmouth: "Artificial Intelligence" coined at Dartmouth — 1956 Dartmouth Summer Research Project; McCarthy, Minsky, Rochester, Shannon.
- Anthropic: Introducing computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku (22 October 2024) — Date screenshot-driven computer control shipped as a public beta API.
- Anthropic: Introducing the Model Context Protocol (25 November 2024) — Open standard for connecting models to tools and data; relevant to the 'API exists' kill condition.
- Anthropic pricing — Claude Sonnet list price of $3 per million input tokens and $15 per million output tokens.
- Anthropic docs: prompt caching — Cache reads are billed at a fraction of base input price.
- Anthropic docs: vision / image tokens — Image token estimate of (width × height) / 750, used for screenshot cost arithmetic.
- Y Combinator: Requests for Startups — Example of a published theme list, which is not the same thing as an idea.