Building AI Startups in 2026: Three Categories, One Worth Betting On

Executive overview

Most founders frame AI as the problem customers want solved, when it is only ever part of the solution space — the real problem is the same one customers had before AI existed. Jason Cohen (WP Engine, SmartBear) cuts the AI product landscape into three categories and argues only one is viable for bootstrapped founders today: AI for experts, where imperfect AI output can be corrected by the user. The other two — incumbent bolt-ons and AI for non-experts — fail for structural reasons unrelated to market size. A further filter applies: the improvement must be 3–10x measurable, not 20–30%, or it won't change buyer behaviour.

AI is the solution space, never the problem space — build what people already wanted, made newly possible.

AI framing: problem vs solution space

• Customers want more leads, better code, faster writing — they have never wanted "AI"
• Positioning around AI as the value creates a category with no demand
• Correct frame: what did customers already want that AI now makes possible for the first time
• Budget bias toward AI is real but only tells you to include it, not what to build
• Restaurant phone-tree example: the product is automated multilingual ordering, voice AI is the mechanism

Three categories and why only one works

• Category 1 — incumbent bolt-ons: Notion AI, Sheets AI; happening everywhere, largely useless, not a startup opportunity
• Category 2 — AI for experts: sell to professionals who can catch and fix AI errors; market is narrower but the unreliability problem is managed
• Category 3 — AI for non-experts (muggles): market is 100–1000x larger but AI getting to 70–80% and stopping is a full blocker for someone who cannot debug the remaining gap
• For a B2B problem-solution founder, Category 2 is the only one where current AI reliability is not a deal-breaker
• Vibes-coded SaaS at 80% completion is not a SaaS company

The 3–10x threshold test

• WP Engine won a commodity hosting market by being 4x faster, not 30% faster — customers could feel it without measuring
• A 20% AI productivity gain is too weak to change buying behaviour or justify switching costs
• Going from one article a week to two articles a day is a threshold crossing; writing slightly faster is not
• AI coding is 10–100x in narrow contexts (unfamiliar library, first-time setup) and slower than manual in large codebases — context determines everything
• Target a domain where AI is genuinely 3x+ on value delivered, not just cost saved; cost savings are a weaker pitch

Moats and competition

• Every reasonably sized market will be flooded with AI products using the same underlying models
• Tech is not a moat when everyone accesses the same APIs and writes prompts
• Only viable differentiation: a sharply defined narrow customer and a product built obsessively for that customer
• Breadth of target market reduces product quality for any specific customer; narrowness is the only path to "absolutely amazing"
• Success is a Venn diagram of many constraints with a centre that may not exist — finding it requires real customer contact, not pre-launch theorising