Enterprise software procurement was designed around relatively predictable units: users, licenses and subscriptions. Agentic AI is disrupting that assumption. A single employee prompt can trigger tool calls, model switches, retries, background APIs and multi-agent tasks—turning an apparently simple interaction into a chain of potentially billable events.
The rise of agentic AI is creating a new headache for enterprise technology leaders: the software bill may no longer be predictable from the number of users—or even the number of prompts.
Traditional SaaS procurement works because buyers can generally estimate consumption. Ten thousand employees might require 10,000 seats. A fixed subscription can then be compared with expected usage and budgeted over a year.
AI agents operate differently.
A single instruction can initiate a sequence of actions. An agent may call external tools, query multiple systems, retry failed operations, invoke a more capable model, run recursive reasoning loops or hand work to another agent. Each step can potentially generate consumption and, depending on the vendor’s pricing model, another charge.
That creates a procurement problem that extends beyond negotiating a lower per-token or per-user price.
Info-Tech Research Group’s new blueprint, “Negotiate Safe AI Contracts to Prevent Bill Shock,” argues that enterprises need to treat AI contracts as dynamic consumption agreements rather than conventional software licenses.
The research identifies a central weakness in many emerging AI agreements: critical billing definitions may not be fully captured in the signed contract. Instead, details can live in vendor documentation, usage policies or pricing pages that are subject to change.
For CIOs, CTOs and procurement teams, that distinction matters. If the definition of a billable event can change after deployment, the organization’s original financial model may become obsolete without any change in the underlying AI workflow.
John Donovan, principal research director at Info-Tech Research Group, argues that enterprises need financial controls and contractual protections in place before production deployment.
“Clear definitions, consumption caps, audit rights, pricing-change protections, and dispute mechanisms must be negotiated before signing,” Donovan said in the supplied announcement.
The issue becomes particularly important as companies move from generative AI assistants toward autonomous or semi-autonomous agents.
An LLM chatbot typically responds to an interaction. An agent can pursue a goal across several steps. Consider an employee asking an AI procurement agent to identify an overdue supplier invoice, compare the contract terms, check the purchasing system, contact another service through an API and prepare a recommendation.
From the user’s perspective, that is one request. From the underlying architecture, it could involve dozens of model and software operations.
That gap between human-visible activity and machine-visible consumption is at the heart of the emerging AI procurement challenge.
The New Economics of AI Software
The problem is not limited to one vendor or one pricing strategy.
Microsoft is incorporating Copilot and AI agents throughout its enterprise ecosystem. Salesforce is pushing Agentforce into customer and business workflows. Google and Amazon are expanding cloud platforms for building and deploying AI agents, while NVIDIA supplies much of the underlying compute infrastructure.
Across these ecosystems, pricing can involve combinations of seats, tokens, compute, API calls, transactions, agent actions or other consumption measures.
That makes comparing vendors more complicated than comparing conventional SaaS subscriptions.
An enterprise might find that Vendor A has the lower headline price but produces higher effective costs because its architecture generates more tool calls or routes complex requests to more expensive models. Vendor B could have a higher base fee but more predictable consumption controls.
The result is a procurement metric that increasingly needs to move beyond cost per user toward cost per completed workflow.
That shift could also change how finance, procurement, IT and AI governance teams work together.
Info-Tech’s framework identifies four stages: understanding vendor billing mechanics; assessing contractual and architectural exposure; designing financial guardrails; and maintaining ongoing market intelligence as models, meters and pricing tiers change.
The recommended controls include spending thresholds, throttling mechanisms and kill switches. Contract negotiations can also address audit rights, dispute procedures, billing definitions and protections against unilateral pricing changes.
These measures are familiar in cloud cost management, but agentic AI introduces a more complicated variable: the system itself can decide how many operations are necessary to complete a task.
Why Architecture Now Matters to Procurement
That means procurement teams cannot evaluate an AI contract in isolation from the technology architecture behind it.
An enterprise deploying an AI agent should understand what happens after a user submits a request. Which model is called? Can the system escalate automatically to a more expensive model? How many retries are permitted? Can agents call other agents? Are tool calls charged separately? What happens when an external API fails?
These questions affect the financial exposure of the deployment.
They also highlight a growing intersection between FinOps, AI governance and enterprise architecture. The people negotiating the contract may not be the same people designing the agent workflow, yet both decisions can materially affect the final invoice.
Info-Tech’s contract risk workbook attempts to bridge that gap by scoring 22 controls across six risk domains and identifying 19 systemic risk patterns, according to the supplied announcement. The blueprint also includes a governance RACI, dispute playbook and tools for modeling 12-month spending.
The methodology reflects a broader reality of enterprise AI adoption: governance is increasingly becoming an economic discipline as well as a security or compliance function.
From AI Experiment to Production Contract
This matters because enterprises are moving AI beyond isolated experiments.
McKinsey’s 2025 State of AI research found that 88% of surveyed organizations regularly used AI in at least one business function, while most had yet to scale AI across the organization. The research also found that 23% were scaling an agentic AI system somewhere in the enterprise, with another 39% experimenting with AI agents.
The transition to production makes pricing risk more consequential.
A pilot with a few hundred users can absorb unexpected costs. An AI agent embedded across customer service, finance, software development or procurement can create a much larger and more persistent consumption footprint.
The challenge for enterprise buyers is therefore not simply negotiating today’s price.
They need to understand how the price behaves when the AI becomes successful.
That could mean negotiating volume tiers, hard spending limits, notification thresholds, transparent meter definitions and sufficient time to contest invoices. It could also mean designing architectures that limit unnecessary model calls and provide visibility into the cost of individual workflows.
For vendors, the change presents its own challenge. More complex pricing may better reflect the real cost of providing agentic capabilities, but opaque meters can make customers reluctant to deploy agents deeply.
The companies that make consumption transparent and predictable may ultimately have an advantage over platforms that leave buyers guessing.
Agentic AI is promising to turn software from a passive tool into an active participant in business processes. But as enterprises discover, autonomous software also introduces autonomous consumption.
The next generation of AI procurement will have to account for both.
Market Landscape
Agentic AI is moving enterprise software toward consumption models that combine SaaS licensing, model inference, API usage, compute and workflow execution.
The major technology ecosystems are approaching the market from different directions:
- Microsoft: Integrating Copilot and agents across productivity, business applications and Azure AI.
- Salesforce: Embedding autonomous agents into CRM and enterprise workflows through Agentforce.
- Google: Combining Gemini models, cloud infrastructure and agent-development capabilities.
- Amazon Web Services: Providing foundation models, agent tooling and cloud infrastructure through AWS.
- NVIDIA: Supplying the accelerated computing infrastructure required to train and run increasingly sophisticated AI systems.
For enterprise buyers, the competitive question is increasingly shifting from Which AI platform has the best model? to Which platform provides the best combination of capability, governance and predictable economics?
That makes contract design part of AI architecture.
Top Insights
- Agentic AI can turn one employee request into dozens of billable operations, making conventional per-seat procurement inadequate for autonomous enterprise workflows.
- Info-Tech recommends clearer billing definitions, spending caps, audit rights and pricing protections before organizations deploy production AI agents at scale.
- Microsoft, Salesforce, Google and AWS are expanding agentic AI across enterprise software, increasing the importance of transparent consumption models for technology buyers.
- AI procurement increasingly requires collaboration between CIOs, procurement, finance, FinOps and engineering teams because architecture directly influences consumption and costs.
- As enterprises scale AI agents, vendors offering predictable pricing, detailed usage visibility and strong governance controls could gain an important competitive advantage.
