There is no shortage of AI startup ideas. There is a shortage of AI startup ideas that someone has asked for. We went looking for the second kind in the complaints that users of AI products leave about the AI products they already pay for: 561 complaints about AI writing assistants, 489 about AI sales assistants, 417 about AI agents for business operations, and a smaller set about chatbot platforms. Nearly 1,500 in total, clustered into about 55 recurring pain points.

The headline finding is not what you would expect from reading launch announcements.

People rarely complain that the AI is bad

Across the whole set, complaints about model quality ("the output is generic", "the voice intonation is off") exist but are a minority. The dominant complaints are about everything around the model:

  • Feature overload. The single largest AI pain point in our data: "Feature overload makes AI writing tools unusable", 117 complaints across 10 products. Vendors ship a new capability every week and the interface is now a wall of buttons nobody can find.
  • Pricing that punishes success. "Per-seat pricing punishes team growth", 155 complaints across 5 operations platforms. The team that gets value from the tool is the team that gets penalised for adding people to it.
  • Setup that needs an expert. "AI agent setup fails without expert hand-holding", 34 complaints across 7 products. The demo works; the deployment doesn't.
  • Forced UI changes. "Degraded UI after forced software upgrades", 32 complaints across 6 sales tools. Teams trained on last month's interface get a new one with no warning.
  • Vendor support that never answers. Chatbot vendors, specifically, draw the complaint that their own support "response time is just too long". Five products.

The implication for founders is uncomfortable if you were hoping to build a better model, and encouraging if you were not: the gaps are in packaging, onboarding, pricing and operations, which are things a small team can build without a research budget.

Ten AI startup ideas from the data

Make existing AI tools usable

Fix the economics

Make the output ship-ready

Why this works better than the "AI for X" list

"AI for lawyers", "AI for dentists" lists start from a technology looking for a customer. The ideas above start from customers who have already adopted the technology and hit a wall. That changes three things:

  1. Demand is proven. The buyer already pays for an AI tool. You are selling the missing piece, not the concept.
  2. Distribution is obvious. Every one of these has a named community: agency owners, Airtable ops admins, JMeter users, WhatsApp bot builders. You know where they are because the complaints tell you.
  3. The moat is specificity. The incumbent could build any one of these. It won't, because it is busy shipping the sixty-first feature. That is the whole opening.

For more on the counting method behind these, see how we validate ideas with complaints. The AI Writing Assistant, AI Sales Assistant and AI Agents markets are updated as new complaints come in.

Every idea above links to its evidence page — the products the complaints came from, how many people said it, and what would have to be true for the idea to work.

Browse all AI ideas