AI spending is surging and CIOs need better cost controls

AI contracts are becoming a cost-control tool

AI spending will continue to rise as companies move from pilots to wider deployment. The key constraint is not access to AI. It is the ability to control consumption and connect that consumption to business value. CIOs should therefore treat vendor contracts as part of their AI cost architecture.

Token pricing is one place to start. Tokens are the units that AI providers use to measure the text or other content processed and generated by a model. Enterprises can negotiate pricing that defines the relationship between input and output token costs. They can also set hard consumption limits inside AI workflows. When a workload reaches its limit, the system can automatically reduce usage or require management approval before allowing more spending.

Cheparthi, whose position and company are not identified in the provided text, argues that enterprises should move toward outcome- or value-based pricing when possible. As she states, “Traditional consumption models create unpredictability and often misalign with actual business value.” This distinction matters. A consumption model rewards higher usage regardless of the result. A value-based model seeks to connect the supplier’s economics more closely to the outcome the enterprise receives.

Contract terms can also address unused capacity. Cheparthi recommends negotiating provisions that allow unused tokens from one period to roll into another. Without such terms, enterprises risk paying for capacity they did not consume and then purchasing more capacity in the next contract period.

CIOs should not assume that every workload suits value-based pricing. Outcomes can be difficult to define, attribute, and measure. Vendors may also demand a premium when they take on more commercial risk. The contract must therefore specify the outcome, how it will be measured, which party controls the relevant variables, and what happens when targets are missed.

The practical goal is predictable economics. Enterprises need clear unit prices, enforceable usage limits, and escalation rules before AI demand expands across the company. These controls do not require reducing useful AI adoption. They make spending deliberate and give executives a stronger basis for deciding which AI deployments deserve more capital.

AI architecture determines how much each task costs

Model choice has a direct effect on AI economics. Enterprises do not need their most capable and expensive model for every request. CIOs can reduce costs by routing each workload to the lowest-cost model that can meet its requirements.

Routine requests can run on lower-cost or open-source large language models (LLMs). More complex work can use premium subscription models where higher performance justifies the additional expense. This requires clear routing rules based on factors such as task complexity, accuracy requirements, latency, security, and cost. The objective is not to minimize spending on every request. It is to avoid paying premium rates where they produce no meaningful additional value.

Cost visibility is equally important. Cheparthi, whose position and company are not identified in the provided text, recommends tracking cost per token and setting reduction targets for redundant prompts and excessive consumption. She also argues that token usage should become a standard operational metric across AI workflows. These measures give CIOs a clearer view of which applications, departments, and use cases generate costs.

Token counts alone, however, are not enough to measure efficiency. A cheaper model can become more expensive if it produces weak results that require repeated prompts, additional processing, or human correction. Executives should therefore connect consumption metrics with measures such as cost per completed task, accuracy, response time, and business outcome. The relevant question is how much it costs to achieve an acceptable result.

Multivendor architectures introduce their own costs. Supporting several models can increase integration work, testing requirements, security reviews, and operational complexity. Open-source models can also require infrastructure, engineering, monitoring, and maintenance that are not reflected in token prices. CIOs should compare total operating cost rather than model prices in isolation.

The Gartner report identifies AI architecture, how an organization executes its AI strategy, as an important opportunity for cost savings. The excerpt does not provide quantitative findings from that report. The practical conclusion is still clear: architecture decisions determine where AI spending occurs. Enterprises that measure consumption, route workloads deliberately, and reserve premium models for tasks that need them have stronger control over AI economics as adoption grows.

AI budgets need hard limits at the point of use

Enterprise-wide AI budgets provide oversight, but they do not show where consumption originates. CIOs need controls at the department, application, and workflow level. This makes each business unit accountable for its own usage and gives management a clearer view of which AI investments justify further spending.

Cheparthi recommends segmented budgets with strict financial boundaries for each department or application. When usage reaches a predefined limit, the system can automatically throttle access or trigger an escalation process. Additional consumption then requires approval rather than continuing without constraint.

Automation matters because AI consumption can scale quickly. A widely used application can generate large volumes of model requests, while poorly designed workflows can repeatedly call a model without producing proportional value. Waiting for a monthly cost review identifies the problem after the expense has occurred. Automated limits allow enterprises to intervene when consumption reaches an agreed threshold.

The limits should not be identical across the company. A customer-facing AI service may require more capacity than an internal experiment, while a business-critical application may need different escalation rules from an optional productivity tool. Executives should set budgets according to expected business value, workload requirements, and operational importance. They should also define who can approve exceptions and under what conditions.

Hard limits create another risk: controls that are too rigid can interrupt valuable or critical workloads. CIOs should therefore distinguish between alerts, soft thresholds, throttling, and complete suspension. High-priority services may need an escalation path that preserves continuity while still requiring management review of unexpected spending.

Cheparthi argues that enterprises should establish these controls from the start to encourage disciplined usage. The larger management objective is accountability. Department-level budgets make AI costs visible to the teams generating them, while automated enforcement turns spending policies into operational controls. As AI deployment expands, that combination gives executives a more precise way to decide where consumption should grow and where it should stop.

AI governance can expose costs hidden inside vendor bundles

AI costs do not always appear as direct model usage. Vendors increasingly include AI features in broader software packages and charge an AI premium at renewal. These bundled premiums are an often-overlooked source of higher recurring costs. CIOs need governance that examines what the company is buying, whether employees use it, and whether the added capability produces measurable value.

Contract review is therefore part of AI governance. Enterprises should include people with AI expertise in procurement and renewal decisions. These teams can assess whether a vendor’s AI functionality is necessary, whether equivalent capabilities already exist elsewhere in the technology portfolio, and whether pricing reflects actual use. This can also help companies identify duplicate AI capabilities across multiple software subscriptions.

The review should focus on outcomes, not feature availability. A business may pay for AI functions embedded across hundreds or thousands of software licenses even when only a small group needs them. Executives should require evidence of adoption and measurable results before approving broader deployment or renewal. Where vendors allow it, companies can also seek contract terms that separate optional AI functionality from the core product.

Cheparthi argues that stronger governance can prevent “capital leakage” and force more rigorous outcome measurement. She also recommends aligning departmental performance goals with AI initiatives that deliver measurable value. This shifts accountability closer to the business teams requesting and using the technology.

There is an important limit to this approach. Governance can itself become expensive and slow if every AI purchase requires a complex review. The process should therefore reflect financial and operational risk. Large contracts, significant AI premiums, sensitive data use, and overlapping capabilities warrant deeper scrutiny. Lower-cost and lower-risk deployments can follow a simpler approval process while remaining subject to basic usage and cost monitoring.

The objective is not to restrict AI adoption. It is to ensure that renewal spending follows demonstrated value. A strong governance process gives executives a consolidated view of AI commitments, usage, duplication, and outcomes. That information improves contract negotiations and helps capital flow toward AI initiatives that can justify their cost.

Key takeaways for leaders

• Make AI contracts a cost-control tool: Negotiate fixed token economics, usage limits, rollover provisions, and value-based pricing where outcomes can be measured. The goal is predictable spending tied to business value.
• Route workloads based on cost and requirements: Use lower-cost or open-source models for routine work and reserve premium models for tasks that require them. Track cost per token alongside task quality and outcomes to measure true efficiency.
• Enforce budgets at the point of use: Set spending limits by department, application, and workflow, with automated alerts, throttling, and approval rules. Give critical workloads clear escalation paths so cost controls do not disrupt essential operations.
• Make AI governance part of procurement: Review bundled AI premiums, duplicate capabilities, adoption, and measurable results before renewals. Add AI expertise to contract reviews so spending follows demonstrated value rather than feature availability.