# What keeps an AI startup hard to copy

A prototype and a defensible business are different things. I asked an AI to lay out a repeatable way to find niches where a business can still defend itself once the underlying model is nearly free.

This is a framework on paper. I have not run it as a pipeline, and the example numbers in the study are only illustrations.

## What I found
- Start without a category. Five agents list problems across industries, constraints and geographies, aiming for 300 to 500 small, concrete problem statements, and one agent scores them on pain alone, blind to any company that already sells a fix.
- Treat regulation as a source of opportunity as well as a barrier. Compliance, audit and privacy rules create demand that cannot be skipped, and the framework brings this in only at a later stage so that it does not bias the first list.
- Rank niches by an index of market attention with five signals: funding speed, buyer pain, ecosystem speed, governance readiness and academic activity. The suggested weights are 30%, 25%, 20%, 15% and 10%, and they are a starting proposal. Whether these choices predict durable businesses remains untested.
- Ground each problem with at least three independent sources, for example a job post, a tender and a forum thread. A problem that shows up in only one place counts as a rumor.
- The framework makes four design choices for every solution: it lives inside the tools people already use, it logs every decision and cost, it is governed with human approval and audit trails, and it connects to the core business systems first.
- Test each idea against a hostile question: does it survive if a top model costs almost nothing? The study argues, in reasoning written by the model it tests, that the moat moves from model quality to governance, data rights and distribution.
- End with three 18-month scenarios per niche, base, bull and bear, each with leading indicators to watch, and choose between bottom-up and top-down sales by stated rules, such as auditability and deal size.

## How it was made
The study is dated 30 January 2026 and links 34 distinct web pages. An AI research agent gathered the sources and drafted it. A swarm of AI agents is also the method the framework itself proposes. I asked the question. Some linked sources are blog posts and professional network posts, so I treat the framework as a way of thinking, not as verified fact. Figures were current at that date.

## Sources
- [European Parliament, EU AI Act: first regulation on artificial intelligence](https://www.europarl.europa.eu/topics/en/article/20230601STO93804/eu-ai-act-first-regulation-on-artificial-intelligence)
- [European Commission, navigating the AI Act](https://digital-strategy.ec.europa.eu/en/faqs/navigating-ai-act)
- [Osborne Clarke, EU AI Act becomes law](https://www.osborneclarke.com/insights/eu-ai-act-becomes-law-1-august-what-are-compliance-deadlines-businesses)
- [Cooley, the proposed digital omnibus on AI (November 2025)](https://www.cooley.com/news/insight/2025/2025-11-24-eu-ai-act-proposed-digital-omnibus-on-ai-will-impact-businesses-ai-compliance-roadmaps)
- [Gartner, 2025 software buying trends](https://www.gartner.com/en/digital-markets/insights/2025-software-buying-trends)
- [World Economic Forum, an AI-first start-up playbook](https://www.weforum.org/stories/business/ai-first-start-up-playbook-a-successful)
- [Qwen3 technical report](https://arxiv.org/pdf/2505.09388)
