The Models
Five distinct pricing axes are live in the CRM AI market as of April 2026:
- Token-based: OpenAI ($1.25/$10 per M for GPT-5; $0.10/$0.40 per M for GPT-5-nano), Anthropic ($3/$15 per M for Sonnet 4.5; $0.80/$4 for Haiku 4.5), Google Vertex (Gemini 2.5 Pro at comparable rates), AWS Bedrock pass-through.
- Conversation-based: Salesforce Agentforce, list price $2 per conversation (heavily discounted on enterprise commits, often landing $0.50–$1.00).
- Seat-based: Microsoft Copilot for Sales at $30/user/month, Copilot for Service at $50/user/month; HubSpot Service Hub Pro AI at $90/seat add-on.
- Outcome-based: HubSpot Breeze (April 14, 2026) — Customer Agent at $0.50/resolved ticket, Prospecting Agent at $1.00/qualified lead. Sierra and Decagon publish per-resolution rates negotiated per contract.
- Credit-based: ServiceNow Now Assist credits, ~$0.04–$0.12 per credit depending on tier; Workday AI Catalog credits; Zendesk AI suite consumption credits.
Pros and Cons
| Model | Predictability | Risk | Best for |
|---|---|---|---|
| Token | Low | Uncapped | Prototyping, low-volume custom |
| Conversation | High | Over-charge on short turns | Customer-facing chat |
| Seat | High | Pay for unused | Internal tools, SDR/CSR teams |
| Outcome | Medium | Definition disputes | Aligned vendor relationships |
| Credit | Low (opaque mapping) | Vendor-controlled | Bundled with platform |
Conversation pricing penalizes well-designed agents that resolve in two turns; rewards verbose ones. Seat pricing penalizes bench capacity; rewards utilization. Token pricing penalizes context-heavy retrieval; rewards prompt caching discipline. Outcome pricing only works if the outcome definition matches the buyer’s understanding — and the contract should make that explicit.
Total-Cost Modeling
Model worst-case, realistic, and best-case scenarios under your actual usage pattern. A pricing model that looks cheaper per unit can cost more in your specific mix.
Worked example for a 50,000 conversations/month deployment, average 3 turns, 50% resolution:
- Token (Sonnet 4.5, 8K context): ~$8K/month inference + platform.
- Conversation (Agentforce $1 negotiated): $50K/month.
- Outcome (HubSpot $0.50/resolved): $12.5K/month.
- Seat (50 reps at $50): $2.5K/month — but only useful if reps are the constraint.
The same workload swings 6x across pricing models. Model both scenarios with realistic resolution rates.
Cost Considerations
Hidden costs that distort comparison:
- Platform fees (Agentforce platform license $50K+/year before per-conversation).
- Tool-call costs (each MCP call has compute).
- Retraining/fine-tuning fees (Vertex, Bedrock charge per fine-tune).
- Egress costs (data leaving cloud regions).
- Implementation labor (consulting partners $250K–$1.5M for enterprise rollout).
Negotiation Levers
2026 pricing is fresh for most vendors and discounts are real. Volume commits get 20–60% off list. Multi-year lock-ins carry risk in this pricing volatility — most teams favor 12–18 month terms with annual repricing. Negotiate: price-cap clauses, outcome-definition transparency, exit/portability provisions, model upgrade pass-through (avoid being stuck on an old model when the vendor releases a better one).
What to Do This Week
For your largest AI vendor relationship, build the cost model under a 50% volume increase and a 50% volume decrease. The right contract handles both without renegotiation.
