Research Note · AI · Procurement

Enterprise AI procurement: contracts, pricing and leverage.

Enterprise AI is the most commercially consequential category of software spend that procurement teams are least prepared to negotiate. This note maps the 2026 buying landscape end to end — build vs buy, pricing models, the clauses vendors omit, vendor selection, negotiation leverage, and the governance obligations that now belong in every contract.

By James Hill-WoodUpdated Sep 202312 min readAI research cluster
Bottom line

The AI agreement most enterprises signed for a pilot is now governing production — on pilot-grade terms. Three blended pricing models obscure true cost; six protective clauses are routinely absent; and lock-in is already live. The highest-value move is an advisor-led competitive process: two credible alternatives plus committed volume capture 20–45% discounts and the clauses vendors never volunteer.

01 Key findings

  1. The pilot contract is running production. Foundation-model API terms are written for a few-thousand-dollar trial, not a multi-million-dollar deployment. Most buyers never renegotiated the agreement before scaling.

  2. Six clauses are non-negotiable — and routinely missing. IP ownership of outputs, data-use restriction, model stability, SLA, audit rights and exit/portability are absent from most click-through and standard enterprise agreements.

  3. Three pricing structures, one blended bill. Token, seat and consumption pricing combine within single agreements. Output-token underestimation and unadopted seats drive the largest cost overruns.

  4. Lock-in is not a future risk. Proprietary APIs, captured fine-tunes and embedded workflows compound switching cost. Architect for provider optionality and negotiate extraction rights before you commit.

  5. Competition is the lever. AI vendors in a growth phase respond to commercial pressure. Two credible alternatives plus a committed volume consistently capture 20–45% off list — buyers who negotiate solo pay list-adjacent.

02 The 2026 AI procurement landscape

The enterprise AI market has consolidated into three commercial tiers, each with distinct procurement dynamics. Knowing which tier you are negotiating in determines your bargaining power and your risk exposure.

Foundation model providers — OpenAI, Anthropic, Google DeepMind, and to a lesser extent Meta and Mistral — sell direct API access under service agreements with minimal enterprise-grade terms. The agreement signed for a low-cost pilot is frequently the same one governing a production deployment. Cloud-integrated AI services — AWS Bedrock, Azure OpenAI Service, Google Vertex AI — bundle third-party models with cloud infrastructure, so buyers with existing enterprise agreements can often fold AI terms into more favourable structures. Embedded AI — Microsoft Copilot, Salesforce Einstein, ServiceNow Now Assist, SAP Joule, Workday AI — is sold as seat-based add-ons to existing licences, where procurement leverage is highest because AI can be bundled into broader renewals.

Insider intelligence

The fastest-growing source of unbudgeted AI spend is not new AI contracts — it is existing SaaS vendors activating AI features on your installed user base and billing them as contract additions or true-ups. Review your SaaS agreements now for AI feature-activation rights and auto-billing provisions.

03 Build vs buy

Before any pricing conversation, resolve the deployment posture. Buy — vendor APIs and SaaS — wins on speed to value, managed scaling and access to frontier models, at the cost of dependency, per-unit economics that rarely fall, and the contractual exposures catalogued in this note. Build — self-hosting open-weight models on owned or rented infrastructure — wins on data control, unit economics at sustained high volume, and independence from vendor pricing evolution, at the cost of MLOps burden, capability lag behind the frontier, and real engineering headcount.

For most enterprises in 2026 the answer is neither pure: a bought frontier model for reasoning-heavy, low-volume work, and a built or open-weight model for high-volume, latency-sensitive, data-sensitive workloads. The decisive variable is sustained token volume — below it, buy; well above it, the per-unit economics of build begin to dominate. Model both net of committed-use discounts and internal loaded cost before deciding.

04 AI pricing models

Enterprise AI products use three primary pricing structures, often combined within a single agreement. Each contains commercial traps that vendors rely on buyers to miss.

StructurePrimary vendorsList basisEnterprise discountCore trap
Token-basedOpenAI, Anthropic, Cohere$0.15–$75 / M tokens20–40% committed volumeOutput tokens cost 3–5× input
Seat-basedCopilot ($30/user/mo), Einstein, Now Assist, JouleFlat fee per user / month15–30% at 500+ seatsPaying for unadopted seats
Consumption-basedAWS Bedrock, Vertex AI, Azure AICompute, calls, processed units20–45% committed-useOpaque stacked charges

A 2× error in output-token volume produces a 3–4× error in actual cost; 5,000 Copilot seats at 30% adoption means 3,500 seats generating no value. Insist on vendor-provided usage modelling and adoption-indexed payment before committing. Discount ceilings differ sharply by tier:

Specialised AI
50% off
Cloud AI
45% off
Foundation APIs
40% off
Embedded AI
35% off
Note

Ceilings reflect the strongest instruments — specialised vendors (Cohere, Mistral, AI21) discount hardest under a credible competitive-alternative threat; embedded AI discounts most when bundled into an existing EA renewal. The list winner is rarely the negotiated winner; model every tier net of committed credits.

05 Non-negotiable contract clauses

Standard enterprise AI agreements — click-through API terms or negotiated contracts alike — routinely omit or weakly address six commercial protections. These are baseline requirements for any AI deployment at scale, not optional extras.

ClauseWhat to demandCommon vendor defaultExposure if omitted
IP ownershipCustomer owns all outputs; no "derivative works" carve-outAssigns outputs but reserves derivativesVendor claims rights to your work product
Data usageNo training or fine-tuning on your data without per-use consentBroad "product improvement" rightsCompetitive-intelligence leakage
Model stability90-day change notice; access to prior versionsDeprecation on short noticeProduction workflows break silently
SLA99.95% uptime; P50/P99 latency; real credits99.5–99.9%, service credits onlyDownstream disruption uncompensated
Audit rightsInspect usage, billing and access logsNone grantedUndetectable billing errors
Exit & portabilityExport data, weights, prompts; 90–180-day transitionNoneLock-in enforced at renewal

Our dedicated AI contract clauses guide provides clause-level language you can drop into a redline.

06 IP ownership and data rights

IP ownership in enterprise AI is more complex than vendors represent, and the commercial stakes are higher than in traditional software licensing. When your systems process millions of interactions through a vendor's model, every one potentially contributes to the vendor's understanding of your industry, customers and processes — a competitive-intelligence risk, not merely a question of who owns a single output.

Four risk areas require specific contractual treatment: data residency (where data is processed relative to GDPR and the EU AI Act); model training rights (explicit prohibition on using your data without consent); competitive intelligence (barring the vendor from using insights from your usage to inform their roadmap or competitors); and regulatory disclosure (the vendor's duty to disclose how your data was used during an investigation or data-subject access request).

From the advisory desk

The elastic-rights trap: we have reviewed agreements in financial services, healthcare and manufacturing where the vendor reserved broad rights to use customer data for "product improvement," "safety monitoring" and "service quality" — categories elastic enough to cover almost any use. The business users who signed had not consulted legal or procurement. This is the AI equivalent of the classic Oracle licensing trap: consequence-laden provisions buried in terms nobody read.

07 Vendor-by-vendor considerations

Procurement posture differs by tier. Match the vendor's commercial framework to your existing estate and workload profile before you shortlist.

OpenAI & Anthropic
Direct foundation APIs
Best for: teams that need frontier capability and can build their own commercial guardrails.
Strengths
  • Enterprise tiers add data isolation, zero retention, output IP
  • OpenAI Enterprise from ~$250k annual commitment
  • Most responsive to competitive pressure
Limitations
  • Default terms are pilot-grade
  • Historically weaker SLAs
  • Token-cost modelling is error-prone
Microsoft Copilot
Embedded / seat-based
Best for: Microsoft-heavy estates renewing an EA or MCA agreement.
Strengths
  • Negotiable inside EA/MCA renewals
  • Leverages existing discount frameworks
  • Bundling leverage across the estate
Limitations
  • Requires M365 E3/E5 — model combined cost
  • Adoption risk on committed seats
  • Complex discount stacking obscures net price
AWS & Google Cloud AI
Cloud-integrated / consumption
Best for: buyers already committed to AWS or Google Cloud infrastructure.
Strengths
  • Folds into EDP / committed-use frameworks
  • Multi-model platforms (Bedrock, Vertex AI)
  • Benchmarks against existing cloud spend
Limitations
  • Deliberately complex, stacked pricing
  • Standard cloud teams under-equipped to evaluate
  • Compute, storage and egress add-ons

For benchmark rates and tactics see our OpenAI Enterprise pricing guide, Microsoft Copilot enterprise guide and Gemini Enterprise licensing guide; the broader Microsoft EA guide covers the full commercial framework AWS and Azure AI negotiations sit within.

08 Lock-in and governance

Vendor lock-in in 2026 is an active commercial reality, operating through three mechanisms. Proprietary API dependency: each vendor's parameters and function-calling conventions differ, so code written tightly against one API requires significant rewriting to migrate — mitigate with an abstraction layer and standardised-endpoint compatibility. Fine-tuned model capture: a model fine-tuned on your data often exists only in the vendor's infrastructure — retain datasets in portable formats and negotiate weight-extraction rights. Workflow integration depth: AI embedded in customer service, knowledge bases and code pipelines compounds switching cost — make deliberate architectural choices that preserve optionality. Our AI vendor lock-in guide details mitigation for each.

EU AI Act

The EU AI Act (in force from 2024, phasing through 2027) classifies AI systems by risk tier and places obligations on deployers as well as providers. High-risk uses — employment, credit scoring, critical infrastructure, certain healthcare — require conformity assessments, human oversight and audit logs. Your compliance depends on vendor cooperation: build documentation, audit-cooperation, change-notification and conformity obligations into the contract. Non-compliance exposes enterprises to fines of up to 3% of global turnover.

09 Procurement framework

Four considerations drive AI vendor selection and RFP scoring. Weight them to your situation before shortlisting.

Factor 01

Existing enterprise agreements

Fold AI into EDP, EA or MCA renewals for structural discount and bundling leverage. Standalone AI contracts forgo it.

Factor 02

Workload & pricing fit

Match token, seat or consumption pricing to your actual usage profile — and validate it with vendor usage modelling before you commit.

Factor 03

Competitive tension

Never engage a preferred vendor without two comparable proposals. Credible alternatives are worth 20–45 points of discount.

Factor 04

Regulatory exposure

Weight EU AI Act risk tier, data residency and audit obligations into selection — they are contract requirements, not afterthoughts.

10 Our recommendation

Direct foundation APIs
When you need the frontier

Reasoning-heavy, differentiated workloads. Force enterprise data isolation, output IP and 90-day model-stability terms before you scale a single line past the pilot.

Cloud-integrated AI
When already on a hyperscaler

You run material AWS, Azure or GCP spend. Negotiate AI consumption inside your EDP or committed-use framework, and model the stacked per-call, compute and egress charges net.

Embedded AI
When Microsoft/SaaS-heavy

You are renewing an EA with material M365 or SaaS seats. Bundle AI into the renewal and index payment to actual adoption, not committed seat counts.

11 Running the process

The single highest-value process choice for enterprise AI buyers:

Advisor-led competitive process Recommended

Run two or more vendors concurrently, benchmark terms against real deals, and sequence commitment to quarter- and year-end pressure. Captures 20–45% discounts and the six protective clauses in one motion.

Direct single-vendor signing Weaker

Accepting the pilot-grade agreement at production scale. Surrenders competitive leverage and typically omits IP, data-use and exit protections — the exposures that surface at renewal.

Run an advisor-led AI procurement process

Our AI practice deploys former OpenAI, Microsoft, Google and AWS executives to benchmark pricing, harden clauses and negotiate on your behalf.

Request AI advisory →

The Licensing Edge

Weekly AI and licensing intelligence for enterprise IT leaders. 3,000+ subscribers.