FinOps In The Age Of AI: Governing Intelligence At Scale

This voice experience is generated by AI. Learn more.This voice experience is generated by AI. Learn more.Sudipto Dasgupta is Global Head of Data, AI, and Automation Platforms at Aon.

gettyAgentic AI and large language models (LLMs) are having a profound impact on enterprises by automating repetitive tasks, reimagining workflows and enabling entirely new business capabilities. From customer service automation, content creation and research assistance to software development and document summarization, organizations are realizing measurable productivity gains from AI-powered solutions. Adoption of generative AI (GenAI) is also steadily accelerating in regulated industries such as insurance, healthcare and financial services.

Industry analysts project explosive growth in AI adoption. Morgan Stanley estimates that GenAI revenue could grow more than twentyfold, from approximately $45 billion in 2024 to over $1 trillion by 2028 as organizations scale AI-enabled business processes and customer experiences.

Despite these benefits, AI introduces a new class of financial and operational challenges. Unlike traditional software licenses or cloud infrastructure, AI costs are highly variable and influenced by multiple dynamic factors. If not governed effectively, organizations can experience unexpected budget overruns.​

As AI adoption scales, enterprises need to evolve FinOps practices beyond cloud infrastructure and develop an AI FinOps framework that provides visibility, accountability, forecasting, optimization and governance across AI consumption.​

These dynamics create several distinct cost challenges that traditional FinOps practices are not designed to address.

Traditional software costs are largely predictable and license-based. AI consumption is driven by tokens, making cost directly proportional to usage. Token consumption is influenced by multiple factors, such as prompt size and complexity, context window length, number of agent interactions, output length and more. A poorly designed AI application can significantly increase token consumption.​

Enterprises increasingly operate in a multi-model environment that includes OpenAI, Anthropic, Google Gemini, Databricks, AWS Bedrock, Azure AI and open-source models. Each provider offers different capabilities, performance characteristics, latency profiles and pricing structures. The most capable model is not always the most economical.

Traditional technology platforms typically evolve over multiyear cycles, allowing organizations to establish predictable cost, capacity and investment plans. AI ecosystems operate fundamentally differently. New foundation models are introduced frequently, existing models are continuously updated and pricing structures often change as providers optimize performance, scale and market positioning.

As model capabilities improve, business teams naturally seek to adopt newer versions to gain higher accuracy, better reasoning, lower latency, additional modalities or enhanced agentic capabilities. These upgrades can significantly alter consumption patterns, model utilization and overall operating costs. At the same time, competition among providers may reduce the cost of some models while premium capabilities command higher prices. The result is that assumptions used for budgeting and forecasting can become outdated in a matter of months or even weeks. ​

Modern coding assistants and autonomous software agents can execute tasks that previously required multiple engineers. While productivity increases dramatically, consumption often grows even faster. Without proper controls, productivity gains can unintentionally drive disproportionate cost growth.

Also, without validation of generated code, the possibility of future technology debt and security vulnerabilities increases. Given the productivity gain and pace of code generation, manual code review becomes extremely difficult.​

Given the shift away from traditional cost governance models, enterprises require a centralized AI control plane that serves as the governance and operational layer for all AI interactions. The AI control plane should provide:

• Token usage observability to provide comprehensive visibility into AI consumption patterns by monitoring token usage across models, applications, teams and business units. It should enable organizations to track prompts, completions, latency and key cost drivers, while delivering real-time insights into consumption trends.

• AI unit economics to provide a business-centric view of AI investments by measuring the cost of individual conversations, transactions, workflows, claims, policies or other business outcomes. It should link AI consumption to value realization, enabling organizations to understand the economic impact of AI initiatives and assess return on investment.

• Model governance to establish the policies, controls and decision frameworks required to manage AI models throughout their life cycle. It should ensure that model selection, routing and usage thresholds are aligned with organizational objectives while balancing cost, performance, risk and regulatory requirements.

• Automated code governance to establish controls and validation processes to ensure that AI-generated code adheres to organizational standards for security, architecture, compliance and maintainability.

• Continuous optimization to improve the efficiency, effectiveness and economic performance of AI systems by continuously refining prompts, context windows, memory utilization, agent interactions and model selection strategies.​​

Building an AI control plane should begin as a cross-functional effort involving technology, finance, security, risk, architecture and business leaders, with clear accountability for how AI is consumed and governed across the enterprise. Rather than attempting to implement every control at once, prioritize the areas of greatest exposure and highest value, establishing visibility into consumption and cost first, then progressively strengthening model governance, security, compliance and value measurement.

As AI adoption expands from individual use cases to agents and enterprise-wide workflows, the control plane should evolve with it, introducing greater automation and more granular controls where scale and complexity demand them. The objective is not to constrain innovation, but to create the operational discipline that allows you to scale AI confidently, economically and responsibly.​

This ultimately shifts AI governance from a discussion about models and tokens to one focused on economics, outcomes and enterprise-scale operational control.

The FinOps question for AI isn’t “how much are we spending on tokens?” It’s “what is that spend buying us?” Enterprises that build the control plane to answer that question now will be more likely to scale AI purposefully, balancing innovation, cost, risk and measurable business outcomes.​​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?