LLM purchasing for CFOs.
A procurement guide · July 26, 2026
Buying LLM capacity is not like buying traditional software. Usage is variable, pricing changes frequently, and the cheapest list price is rarely the best deal. This checklist helps CFOs buy AI capacity without committing to waste.
1. Match the pricing model to the workload
Before negotiating, understand what you are buying. Real-time customer-facing features need on-demand or reserved throughput. Batch workloads can use batch APIs at steep discounts. Experimental or internal tools may not need an enterprise agreement at all.
2. Do not commit before you know your pattern
Committed use discounts look attractive, but they lock you into a forecast. Most teams overestimate stable usage because growth and model switching are hard to predict. Start with pay-per-token, build three months of data, then negotiate.
3. Require attribution and guardrails
No procurement deal should close without a plan for tagging, dashboards, and spend limits. If you cannot attribute usage to a team or feature, you cannot manage it. If you cannot cap spend, you cannot protect the budget.
4. Keep fallback providers
Single-provider dependency is a pricing and availability risk. The contract should allow, and the architecture should support, fallback to at least one alternative provider. This also gives you leverage in renewal conversations.
5. Check the fine print
Look for: data residency commitments, audit rights, rate-limit terms, overage pricing, and notice periods for price changes. Billing granularity matters too — daily is far better than monthly when you are trying to catch a spike.
CFO checklist
- Workload mapped to pricing model
- Three months of usage data before commitments
- Attribution tags defined and enforced
- Cost caps and alerts configured
- Fallback provider tested
- Data residency and security terms confirmed
- Overage and price-change clauses reviewed