Pricing an AI agent business is fundamentally different from pricing SaaS. Your costs are variable (tokens), your output is non-deterministic, and your value is measured in work done, not features accessed.
The Four Models
1. Per-Seat ($X/user/month)
Predictable revenue, familiar to buyers. But your costs scale with usage, not seats. A power user burning 10x tokens pays the same as someone who logs in once. Best for: Agents that augment human work (copilots, assistants).
2. Per-Action ($X/task)
Aligns cost with value. Customers pay for what they use. Revenue scales naturally. But it's unpredictable and customers may limit usage. Best for: High-volume tasks — lead qualification, data enrichment, content generation.
3. Outcome-Based ($X/result)
Maximum alignment with customer value. Premium pricing possible. But hard to define "success" and you absorb all risk. Best for: High-value, clearly measurable outcomes — meetings booked, tickets resolved.
4. Retainer ($500-$5,000/month)
The dominant model for AI automation agencies ($7K-$40K MRR at 85% margins). Predictable for both sides. Best for: Agency models, SMB-focused agent services.
Unit Economics to Track
| Metric | Danger | Healthy | Best |
|---|---|---|---|
| Cost per task | >$1.00 | $0.10-$0.50 | <$0.05 |
| Gross margin | <50% | 65-80% | >85% |
The Hybrid Approach
Most successful agent businesses in 2026 use a base subscription + per-action overage. Predictable revenue floor with upside from power users. 85% of software vendors have adopted usage-based approaches. Start there.