How AI Tools Make Money

Four real business models. Most AI tools sit in one. Which one decides whether they last.

1. Per-Seat Subscription

Cursor, Copilot, ChatGPT Plus. Predictable revenue, classic SaaS economics. Works because the model cost per seat is bounded by daily-use limits. Most defensible at scale.

2. Per-Token / Per-Use

Anthropic API, OpenAI API. Margins thin at the model layer; volumes are huge. Works for the foundation labs, hard for wrappers.

3. Outcome-Based

Intercom Fin ($0.99/resolution), Sierra (per-resolved-call). Aligns with customer value. Hard to deliver if your unit economics shift with model price changes.

4. AI-Augmented Existing Product

Notion AI add-on, Zendesk AI Agents, Salesforce Einstein. The AI is a feature, not a product. The existing distribution carries it. The most reliable model in 2025.

Who Actually Makes Money

• Foundation labs — Anthropic, OpenAI. Volume covers cost; pricing power is real.
• Distribution-rich incumbents — Microsoft, Google, Salesforce. AI features lift retention and ACV.
• Vertical AI with workflow lock-in — Harvey (legal), Hippocratic AI (healthcare). Defensible.
• Thin wrappers on foundation models — mostly do not. Margins crushed by next price drop.

The Honest Take

If you are building an AI tool and your moat is "better prompts," you do not have a moat. Distribution, data, and workflow integration are the moats. Pick at least one before you launch.