Per-token, per-inference, agentic outcome billing, and the margin chaos of AI-native pricing.
$2 per conversation. No seats, no tokens — Salesforce is betting its $35B business on outcome-adjacent pricing for AI agents. Everyone is watching to see if it works.
Per-token pricing was a developer convenience that became a default business model. It creates adversarial optimization and means nothing to enterprise buyers. The replacements are coming.
Every AI company says they charge for results. Nobody agrees on what a result is. Contracts are going to get very interesting.
There's no single answer — the right model depends on cost structure, customer type, and margin tolerance. The three main approaches are per-token (simple but exposes margin), per-task/credit (abstracts cost), and outcome-based (highest value capture but hardest to define).
The cost spread between the cheapest and most expensive AI models can be 62x or more. If you charge a flat rate, customers using expensive models destroy your margin while cheap-model users subsidize them. This makes flat-rate AI pricing fundamentally unsustainable.
Flat-rate AI subscriptions were loss leaders designed to acquire users. As AI usage scales, the variable inference cost makes flat pricing unprofitable. Every major AI company is adding usage-based components — credits, tokens, or metered API calls.
Outcome-based pricing charges for results — a resolved support ticket, a completed task, a successful code deployment — rather than inputs like tokens or API calls. Salesforce's Agentforce at $2/conversation is the highest-profile example. The challenge is defining and measuring outcomes.
AI agents compress seat counts (one agent replaces multiple human seats), generate unpredictable compute costs, and blur the line between user and machine. Per-seat pricing breaks when agents do the work. Most companies are shifting to per-task or consumption models for agent-heavy workflows.