Enterprise AI platform contracts — with OpenAI, Microsoft Azure OpenAI, AWS Bedrock, Google Vertex AI, and similar services — are a new category of enterprise agreement that combines elements of cloud committed spend contracts, software licences, and data processing agreements. Most procurement teams approaching their first major AI platform agreement are working without established playbooks and without market pricing benchmarks. This guide provides the framework. Part of our comprehensive AI and GenAI Software Procurement Negotiation Guide.
Before engaging in commercial negotiations with any AI platform vendor, structure your preparation around five pillars. These pillars address the unique dimensions of AI contracting that distinguish it from traditional enterprise software negotiations.
The foundation of AI contract negotiation is an accurate consumption model. Unlike per-seat software licences, AI token consumption varies with use case, model choice, context window size, and usage patterns. Without a validated consumption model, you cannot negotiate the right commitment level — and either over-commit (wasting money on unused capacity) or under-commit (losing volume discounts and facing overage pricing).
Build your consumption model by running a production-representative POC for 30 days, measuring actual token consumption across all planned use cases. Account for context window overhead — the background context fed with every API call — which often doubles or triples naive per-query token estimates. Document input:output token ratios for each use case, as output tokens typically cost 3–5x more than input tokens.
AI platform pricing is highly competitive and has declined significantly year-over-year. Before negotiating with any primary AI vendor, benchmark their pricing against at least two alternatives. The credibility of your competitive benchmark depends on the alternatives being genuinely comparable — benchmark GPT-4o against Claude Sonnet and Gemini Pro, not against smaller models that do not meet your performance requirements.
Obtain actual vendor quotes rather than published pricing. Published pricing is the maximum — enterprise volume pricing begins at meaningful annual spend levels ($250K+ for most providers, $500K+ for the largest). The gap between published and negotiated pricing for enterprise volumes is typically 20–35%.
Establish your non-negotiable data requirements before engaging vendors. These include: training prohibition (absolute, unconditional), data residency (EU/US/specific region), retention period (0–30 days for enterprise), DPA compliance (GDPR, sector-specific), and sub-processor restrictions. Vendors who cannot meet your non-negotiable data requirements should be disqualified from consideration regardless of price — the compliance and reputational risk of incorrect AI data handling outweighs any commercial benefit.
Map your AI use cases to reliability requirements. Customer-facing AI applications require 99.9%+ uptime SLAs with financial remedies. Internal productivity tools may be serviceable at 99.5%. Development and experimental use cases may accept best-effort reliability. Define your requirements by use case tier and negotiate platform-level SLAs that meet the highest-tier requirement — or negotiate separate service tiers for different use case categories.
Define your exit strategy before signing. If you decide to switch AI platforms in 18 months, what will it cost? Which assets are portable (your data, your application code, your prompts) and which are not (fine-tuned model weights, vendor-proprietary embeddings)? Negotiate exit provisions that minimise lock-in from the beginning — it is far easier to negotiate portability rights pre-signature than post-commitment.
OpenAI Enterprise is OpenAI's dedicated enterprise offering, providing higher rate limits, enterprise data protections, and priority access to OpenAI's latest models. OpenAI negotiates custom enterprise agreements for customers committing to meaningful annual spend.
Want independent help negotiating better terms? We rank the top advisory firms across 14 vendor categories — free matching, no commitment.
| Term | OpenAI Standard Enterprise | Negotiable? |
|---|---|---|
| Data training prohibition | Yes — included as standard for enterprise | Standard; confirm language is categorical |
| Token pricing | Published list price or small discount | YES — volume discounts from ~$500K annual |
| Rate limits (TPM) | Higher than consumer; tiered by plan | YES — enterprise rate limit customisation |
| Model version access | Current models; deprecation per OpenAI policy | PARTIAL — minimum notice period negotiable |
| Data residency (EU) | EU data residency option available | YES — requires specific DPA and configuration |
| SLA uptime | 99.9% monthly uptime target | PARTIAL — financial remedies limited |
| IP indemnification | Copyright Shield — enterprise included | Review scope and limitations carefully |
| Contract term | Annual | YES — multi-year for larger discounts |
When negotiating OpenAI Enterprise, focus your commercial energy on three areas. First, token pricing: negotiate volume tier pricing with clear breakpoints based on your consumption model. Request pricing for each model tier you plan to use (frontier models vs. mini/fast models) and negotiate blended effective rates that reflect your expected usage mix. Second, model continuity: OpenAI has deprecated models on short notice. Negotiate a 12-month minimum availability period from any deprecation announcement, with access to the previous model version during migration. Third, rate limits: confirm that your negotiated rate limits are contractually guaranteed rather than "targets." Best-effort rate limits that Cisco can adjust unilaterally provide no SLA protection.
Azure OpenAI Service provides access to OpenAI's models within Microsoft's Azure infrastructure, covered by Microsoft's enterprise terms. For enterprises with existing Microsoft EA or Azure MACC, Azure OpenAI is often the preferred commercial pathway because AI spend counts against existing commitments.
Azure OpenAI offers two pricing models that serve different use cases. Pay-as-you-go (on-demand) provides token-based billing at published rates with no commitment. Provisioned Throughput Units (PTUs) provide reserved model capacity measured in tokens per minute of guaranteed throughput — priced at a higher rate per token but providing a predictable monthly cost and guaranteed availability.
PTUs are appropriate for production applications with predictable throughput requirements. On-demand is appropriate for development, testing, and variable workloads. Most enterprises end up with a mix: PTU reservations for core production workloads and on-demand capacity for overflow and development.
Azure OpenAI spend counts against your Microsoft Azure Committed Spend (MACC) agreement. This has two implications: Azure OpenAI spend helps you meet your MACC minimum commitment (potentially avoiding shortfall penalties), and MACC discount tiers may apply to your Azure OpenAI consumption. If you have a MACC with meaningful remaining balance, routing AI spend through Azure OpenAI rather than OpenAI direct can provide an effective 15–25% cost reduction through commitment drawdown. Model this explicitly in your procurement analysis.
AWS Bedrock provides access to multiple AI foundation models — including Anthropic Claude, Amazon Titan, Meta Llama, Mistral, Cohere, and others — through a single AWS API. For enterprises with AWS EDP commitments, Bedrock spend counts against their Enterprise Discount Program, making it commercially attractive for AWS-centric organisations.
Get the IT Negotiation Playbook — free
Used by 4,200+ IT directors and procurement leads. Oracle, Microsoft, SAP, Cloud — all covered.
Bedrock's multi-model architecture is both its commercial strength and its negotiating complexity. You may be consuming Claude Sonnet through Bedrock (paying Anthropic's model pricing) alongside Amazon Titan (priced by Amazon directly). The pricing structure, discounts, and terms may differ by underlying model provider. Review the pricing for each model you plan to use on Bedrock separately — do not assume that your EDP discount applies uniformly across all models.
Amazon Bedrock's provisioned throughput option functions similarly to Azure's PTUs — reserved capacity at a predictable monthly rate rather than per-token consumption. This is particularly valuable for latency-sensitive applications where unpredictable shared-capacity response times are unacceptable.
If you are a heavy Anthropic Claude user, you have a choice: access Claude through AWS Bedrock (where it counts against your EDP) or negotiate directly with Anthropic. Direct Anthropic negotiation may provide better model-specific pricing, earlier access to new Claude versions, and more flexible enterprise terms. However, direct Anthropic spend does not reduce your AWS EDP commitment. Model both options, net of EDP commitment benefit, before deciding on your commercial pathway.
Google Vertex AI provides access to Google's Gemini model family along with a range of pre-trained and custom model capabilities. For enterprises with Google Cloud committed use agreements, Vertex AI spend counts against their Google Cloud commitment.
Google Cloud's Committed Use Discounts (CUDs) provide 20–55% discounts for committed resource usage. Vertex AI consumption can qualify for CUD discount tiers when planned as part of a broader Google Cloud commitment negotiation. If you are renewing or establishing a Google Cloud CUD, explicitly include Vertex AI in the scope of your commitment modelling — it is one of the fastest-growing Google Cloud service categories and often underweighted in initial CUD planning.
Google offers AI capabilities through two commercial pathways: Gemini for Workspace (per-seat add-on to Google Workspace) and Vertex AI (API/developer platform). These serve different use cases and are priced differently. Gemini for Workspace is appropriate for end-user AI productivity (writing assistance, email summarisation, meeting summaries). Vertex AI is appropriate for application development, custom model integration, and programmatic AI workflows. Ensure you are purchasing the right commercial pathway for your use case — conflating the two leads to either overpaying for Workspace add-ons when you need API access, or under-licensing when end-users need integrated AI productivity tools.
Design your AI architecture to be model-agnostic from the outset — using abstraction layers (LangChain, LlamaIndex, or custom routing logic) that allow you to route requests to different models based on performance, cost, and availability. This is not just a technical best practice; it is a commercial lever. When you can demonstrate that your architecture allows you to shift a meaningful portion of workload to a competitor model within days rather than months, every AI vendor's negotiating position weakens. Bring this architecture to commercial discussions.
AI model generations turn over every 12–18 months. A commitment made to GPT-4 level models in early 2024 needs to accommodate GPT-5 class models in 2025 without requiring a new commercial negotiation. Negotiate for commitment flexibility — the ability to apply your committed spend to any current and future model in the provider's portfolio at the committed pricing level. This protects your investment from model deprecation and allows you to benefit from model improvements without renegotiation.
AI API consumption anomalies — runaway loops, prompt injection attacks, unexpected context window growth — can generate thousands of dollars in unexpected charges in hours. Require vendors to provide real-time usage alerts (via API webhook, email, or SMS) when consumption crosses 80% and 100% of your monthly budget, plus the ability to configure hard spend caps that automatically throttle or suspend consumption at a defined threshold. This is a reasonable, achievable operational requirement; vendors who cannot provide it represent an unacceptable budget risk.
Open-weight AI models (Meta Llama 3, Mistral, Qwen) provide a meaningful alternative to proprietary API models for many enterprise use cases. Organisations willing to invest in infrastructure management can run Llama 3 70B or similar models on their own compute at a fraction of the per-token cost of frontier API models for equivalent performance on many tasks. Use open-weight model deployment cost estimates as a floor pricing reference in API model negotiations — not as a genuine alternative for all use cases, but as a credible anchor that limits proprietary model pricing power.
Some AI vendors will request that you commit to their entire model portfolio (e.g., "any OpenAI model") rather than to specific model tiers. Portfolio commitments provide flexibility but reduce pricing leverage — you cannot benchmark one model against another if your commitment is portfolio-wide. Negotiate model-tier-specific pricing (e.g., separate rate cards for GPT-4o, GPT-4o-mini, and o1 class models) so that each tier can be benchmarked independently and you have visibility into the cost allocation across your use case portfolio.
Enterprise AI providers use compute infrastructure from cloud providers and potentially other third-party services as sub-processors. Under GDPR and equivalent regulations, you have the right to be informed about sub-processors handling your personal data and the right to object to new sub-processors. Include a contractual obligation for the AI vendor to: (a) provide a current list of sub-processors on request, (b) notify you at least 30 days before adding a new sub-processor, and (c) provide an audit right mechanism (SOC 2 attestation review at minimum, site visit for regulated sector organisations) that allows you to verify security practices without conducting expensive manual audits for every sub-processor.
Avoid prepaying large annual AI commitments upfront without corresponding contractual protections. If an AI vendor requires upfront annual payment, negotiate credit carryforward rights (unused credits roll to the next period), credit remedies for SLA failures, and pro-rata refund rights for material contract changes (such as model deprecations or data handling policy changes). Quarterly billing with annual commitment pricing is achievable for significant spend levels and provides payment flexibility without sacrificing volume discounts.
AI vendors derive commercial value from understanding enterprise use case patterns, even without using your data for training. Your willingness to participate in customer advisory programmes, provide anonymised benchmark feedback, or serve as a public reference customer has real value to AI vendors — value that should be reflected in commercial terms. Explore whether your organisation can offer specific engagement commitments (case study, advisory board participation, conference presentation) in exchange for commercial improvements. This is most effective with newer AI vendors building their enterprise reference customer portfolio.
AI vendors that achieved SOC 2 Type II or ISO 27001 at launch may not maintain those certifications through rapid product changes. Require contractual commitments to maintain specific security certifications throughout the agreement term, with notice obligations and remediation rights if certifications lapse. For regulated sector organisations, specify the exact certifications required (FedRAMP, HIPAA BAA, Cyber Essentials Plus, ISO 27701 for privacy) as a contractual condition of the agreement, not just a checkbox at procurement time.
AI market dynamics change rapidly. Include commercial review triggers in your AI platform contracts that allow renegotiation if specific conditions occur: if the provider reduces pricing for equivalent services by more than 20% for comparable customers (MFC trigger), if a competitor offers equivalent performance at materially lower cost as confirmed by independent benchmark, or if a major model version change materially affects your application performance. These triggers give you a contractual basis for renegotiation mid-term rather than waiting for renewal.
Approaching your first enterprise AI platform contract negotiation?
Our recommended firms have specialist AI contracting expertise and current pricing benchmarks across all major platforms.
Before signing any enterprise AI platform agreement, verify the following provisions are addressed. This is a summary checklist — for a complete 40-point checklist, download our AI Procurement Checklist white paper.
For more AI procurement topics, explore the full AI and GenAI Software Procurement Negotiation Guide. For AI-specific SLA requirements, see our AI SLA Requirements Guide. For data rights model language, see Data Privacy Clauses in AI Vendor Contracts.
Connect with a specialist AI procurement advisor who has current pricing benchmarks and enterprise contract experience across OpenAI, Azure, AWS, and Google platforms.