The AI Platform Contract Negotiation Playbook

Pricing volatility is the central commercial risk

Token-based AI platform pricing has dropped 70 to 90 percent across major providers since 2023, and the trajectory continues. GPT-4 in 2023 priced at roughly $0.03 per 1K input tokens; equivalent capability in 2026 prices below $0.005. Claude, Gemini, and Bedrock-hosted models have followed similar trajectories. The implication for enterprise contracts is that long-term commits at today's prices are bets against an industry-wide deflationary trend.

The four credible providers compared

Four providers dominate enterprise AI platform contracts. OpenAI Enterprise leads on raw capability for general purpose tasks and developer ecosystem. Anthropic Claude Enterprise leads on safety posture and regulated industry adoption. Google Gemini Enterprise leads on Google Workspace integration. AWS Bedrock leads on multi model flexibility and integration with existing AWS commit. Each has distinct contract templates, distinct data and IP provisions, and distinct commitment structures.

Data rights and the training carve out

The single most important contract provision in any AI platform agreement is the training data carve out: the explicit commitment that customer prompts and outputs will not be used to train provider models. All four credible providers offer the carve out at enterprise tier. The strength of the language varies; the audit and verification mechanisms vary; the data deletion timelines vary. The customer side mistake is to accept the standard language without negotiating the variations that matter.

Model rights, IP indemnity, and output ownership

IP indemnity is now standard for enterprise tier across all four providers, with caps and exceptions varying. Output ownership is uniformly assigned to the customer. The remaining negotiation surface is the indemnity cap (often initially set below contract value), the indemnity exceptions (some carve outs are negotiable), and the indemnity trigger conditions (ongoing model usage versus specific output disputes).

Pre paid commit versus consumption pricing

Most enterprise tiers offer two pricing models: pre paid commit (pay X for Y tokens, with discount), or consumption (pay per token at retail). Pre paid commit looks cheaper per token; in deflationary pricing markets, consumption is often cheaper across a 12 to 24 month horizon. The math depends on the customer's expected usage trajectory and the rate of provider price decline.

Multi provider strategy as commercial leverage

The single most powerful negotiating lever is operating across multiple providers simultaneously. Multi provider architectures are technically achievable through Bedrock (single API to multiple models), through orchestration layers (LangChain, LlamaIndex), or through provider-specific integrations. The architectural complexity is real; the commercial value typically exceeds it. Enterprise customers running active multi provider strategies negotiate 30 to 50 percent better terms with their primary provider than single provider customers do.

SLA, uptime, and model deprecation

AI platform SLAs lag traditional cloud SLAs in maturity. Uptime commitments vary from 99.5 percent to 99.95 percent across providers. Model deprecation policies (advance notice before removing a model) vary from 30 to 365 days. Customers who depend on specific model versions for production workflows must negotiate the deprecation policy explicitly; default policies provide insufficient runway.

Provider counter moves and how to handle them

AI platform provider account teams have a small set of repeatable counter moves: the strategic partnership framing, the early adopter positioning, and the deprecation timeline pressure. None are illegitimate; all are negotiation. The framework includes the standard responses we deploy.