How do you validate an AI startup idea before building?

Validating an AI startup idea focuses on verifying three core pillars before writing code: the existence of a painful "boring" problem, the technical feasibility of current models to solve it, and your unique distribution advantage.

1. Validate the Problem (The "Boring" Work Test)

The strongest signal for an AI startup is a problem that is currently solved by human labor in "boring" or repetitive ways.
* Shadow the "Boring" Jobs: One effective validation trick is to search job boards like Indeed for roles like "remote analyst" or "clerk" and shadow people in those positions. If someone’s full-time job is manually refreshing a page to find government contracts, that is a prime candidate for automation Y Combinator — How To Get AI Startup Ideas @ 27:28.
* Identify Outsourced Tasks: Look for any task currently outsourced to low-wage countries; this is a massive signal that a startup can be built to automate that workflow Y Combinator — How To Get AI Startup Ideas @ 28:55.
* Fall in Love with the Problem: Focus on the outcome, not the tech. Validating that you can "help people avoid traffic jams" is more compelling than pitching "AI-crowdsourced navigation" Lenny's Podcast — A founder’s guide to crisis management @ 1:04:18.

2. Validate the Tech (The "Concrete" Idea & Eval Test)

AI startups often fail by being too vague. Validation requires moving from high-level concepts to "concrete" ideas that can be falsified quickly.
* Define a "Concrete" Idea: Instead of saying "AI for healthcare," validate a specific hypothesis like "AI software to let patients book MRI slots online." A concrete idea allows you to build and either prove or disprove the value immediately Y Combinator — Andrew Ng: Building Faster with AI @ 03:03.
* Build the "Eval" First: Before building the product, define what "good" looks like. Create "Evals" (evaluations) to measure if the model can actually handle the task at a high enough accuracy (e.g., 99%+) to be useful for an enterprise 20VC with Harry Stebbings — Mercor CEO & Co-Founder, Brendan Foody @ 33:41.
* Human-in-the-Loop: For "AI-native services" (like insurance or law), the human often is the interface initially. You validate by providing the final outcome to the customer manually, using AI to help the human scale their work nonlinearly Y Combinator — How to Build an AI-Native Services Company @ 03:03.

3. Validate Demand & Distribution

Before building, you must ensure you have a way to reach users without over-relying on paid ads.
* The Figma Test: Create a wireframe in Figma to visualize the app. Show this to strangers in online communities (Discord, Reddit, LinkedIn) to get "brutally honest feedback" before building any functionality Brett Malinowski — Start A SaaS in 5 Simple Steps @ 02:45.
* Pre-determined Marketing: A key part of validation is establishing a marketing strategy before building. Ask if there is a niche or influencer audience you can tap into immediately; if not, the product will be ten times harder to launch Brett Malinowski — The 17-Year-Old Who Built a $1.12M/Month Ai App @ 15:18.
* The "Sam Altman Test": Ask yourself: "As the models get better, does your service get stronger or does the model itself commoditize you?" You want to build in areas where model improvements make your specific service more valuable, rather than obsolete Y Combinator — How to Build an AI-Native Services Company @ 03:03.

— Sources: 9 videos across 4 creators