How to find AI startup ideas worth working on

Executive overview

Most founders who can't find a startup idea are looking in the wrong place — defaulting to whatever's trending on X or building a hackathon demo. The real path runs in two directions: aggressively inward, mining your own deepest expertise, or aggressively outward, getting inside industries you know nothing about yet.

AI has dramatically raised the value of niche domain knowledge. Problems once too small for software are now big enough to replace full-time employees.

The founders closest to real problems are better positioned than anyone — including investors scrolling feeds from their phones.

Look within: exploit rare expertise

• The best ideas come from being at the "PhD edge" of a domain — where you've seen problems no outsider could even name.
• Founder-market fit matters more than the idea itself; if not you, then who?
• Salient (AI voice agent for auto debt collection) came from a founder who ran Tesla's lease payment operations and saw how manual the process was.
• Diode Computer (AI circuit board copilot) required a co-founder who was both a hardware EE and a software engineer — a combination rare enough to spot a gap no one else could.
• Spur (AI QA agent) came from a Figma engineer who watched teams burn hours maintaining front-end tests.
• David AI found its niche in multimodal speaker-separated data because its founders worked at Scale and saw where Scale wasn't going.

Internships and prior jobs as idea sources

• A large share of YC's billion-dollar companies trace directly to an internship one founder had.
• DataCurve's founder (19, dropped out of Waterloo) pivoted from a generic ChatGPT wrapper to AI data tooling by returning to what she'd built as an intern at Cohere.
• Founders often feel allergic to their own expertise — they've ground through years on something and want to chase something new. The insight they dismiss as "boring" is often the one no one else can articulate as well.
• If you're in college and want startup ideas: do internships at companies on the bleeding edge of something.

Work on things that capture the human imagination

• B2B SaaS ideas manufactured during a pivot are almost always uninspiring and unlikely to sustain long effort.
• Can of Soup (AI-native social network) came from a founder who was pivoting through plausible-but-boring ideas until his old boss at Substack told him: "Work on something that captures the human imagination."
• Happenstance (intelligent people search, replacing LinkedIn's text-based search with LLM + vector search) came from a founder who was genuinely bothered that LinkedIn search was still using plain indices.
• Founders tend to have blinders that prevent ambitious ideas from surfacing — the subconscious filters them out before they're consciously considered.
• Easydubs (universal real-time translator) is an example of an idea that only appears once those blinders are removed.

Get out of the house: embed in real industries

• When internal expertise is exhausted, the answer is to physically enter other worlds.
• Egress Health (AI back office for dentists) came from a founder who spent a day in his mother's dental practice watching insurance pre-authorization get done by hand.
• Family connections provide access; access is often all you need to discover an underserved space no engineer has seen yet.
• Happy Robot (AI logistics coordination for truckers) came from founders with no trucking background who were simply personable enough to get inside the industry.
• Any industry where no good software engineer has ever looked is a candidate — AI means even niche verticals are now worth automating.

Go undercover: get jobs in the industries you want to automate

• One AI billing company founder got an actual job as a medical biller for a New York optometrist office — without disclosing he was building software to automate the role.
• He ran his own AI agent locally on two MacBooks to automate his own job while learning the process from the inside.
• This is legal when using open-source software on your own machine for a remote laptop job.
• The ideal targets are knowledge-work roles that require only a 2–4 week training program — exactly the jobs LLMs are best at automating.
• Alpha trick: search Indeed for "remote analyst", "remote clerk", and similar roles to find obscure automatable jobs.
• Alternatively, shadow a friend who has a boring job for a day — Sweet Spot found their government contracting idea by following a friend whose entire job was refreshing a government procurement page.

Outsourced jobs as a signal

• Any job category outsourced to low-wage countries is a strong signal of a startup to build right now.
• Lilac Labs (AI drive-through order taker) discovered that most US drive-throughs already outsource the order-taker role to BPOs overseas — an obvious automation target.
• Automatt (AI-native RPA) targeted UiPath's market after founders noticed enterprises were paying expensive certified consultants just to make UiPath work — a sign the underlying product was broken.

Hang out with people at the edge

• PrayDB (Postgres-native vector search, replacing Pinecone for many use cases) came entirely from a friend mentioning a sync problem between Postgres and Pinecone.
• Reducto (PDF chunking for RAG applications) found its idea by being inside YC alongside other founders building AI applications, where the chunking problem kept surfacing.
• Living at the edge means constantly using the latest models and tools — you're among the first to notice what's newly possible.
• GigaML (AI customer support, now serving Zepto) entered what looked like a crowded space, but their technical depth meant they could actually deliver what competitors only promised.

Build something — anything — to generate expertise

• Shipping a product makes you an expert in your users and in building products, even if the first idea fails.
• Juicebox / PeopleGPT started as a freelancer marketplace with no differentiation, but the founders built deep familiarity with hiring workflows — which led them to build LLM-powered recruiter search that actually took off.
• The act of building forces real contact with users, which generates ideas that desk research never surfaces.

On competitive spaces and timing

• GigaML avoided a crowded-seeming space (AI customer support) and succeeded because almost no competitor could actually deliver the technical result at enterprise quality.
• Appearing crowded on TechCrunch is not a signal to avoid a space — it's often a signal the space is real.
• AI moves fast enough that new viable ideas are generated every few months; staying in the game is itself a strategy.
• Founders working on cutting-edge AI have unusual morale reserves — the excitement sustains longer pivots than would have been normal five years ago.
• Many of YC's best recent companies took about a year of pivoting to find the right idea; that's now normal.