AI FinOps Starts with Enterprise Architecture

Treat tokens as an enterprise resource

A critical mindset shift is recognising that tokens are not merely a technical metric. They are an enterprise resource.

Just as organisations govern compute, storage and network capacity, they must establish controls around AI consumption. Without visibility and accountability, costs can rise quickly while making it increasingly difficult to determine whether AI investments are delivering expected returns.

Token consumption grows through enterprisewide copilot adoption, retrieval systems that return excessive context, applications that pass large volumes of information to models, and agentic workflows that generate multiple AI interactions. Individually, these activities may appear insignificant. At enterprise scale, they can become a substantial operating expense.

Enterprise Architecture (EA) is uniquely positioned to address this challenge. EA teams already define standards, governance policies and technology guardrails. Applying those same disciplines to AI creates a foundation for sustainable adoption, cost transparency and business-value measurement.

The hidden cost of excessive context

One of the largest drivers of unnecessary AI spending is excessive context.

Organisations frequently send entire documents when a summary would suffice, include lengthy conversation histories when only recent interactions matter, or retrieve significantly more content than a use case requires. The result is predictable: higher costs without a corresponding improvement in business outcomes.

Consider a customer-service copilot that retrieves entire knowledge-base articles when only a few relevant passages are needed. Multiplied across thousands of interactions, this architectural decision can significantly increase costs while providing little additional value.

A simple principle can help: provide only the minimum context necessary to produce an effective result.

Achieving this requires more than prompt optimisation. It demands architectural standards, governance controls and measurement capabilities that ensure efficiency is built into solutions from the start.

Match the model to the workload

Another common source of unnecessary spend is using advanced reasoning models for routine tasks.

Many business processes — including ticket classification, meeting summarisation, document extraction and content categorisation — can be handled effectively by smaller, lower-cost models. Using premium models where they are not needed increases costs without improving business outcomes.

Mature AI environments increasingly route each request to the most appropriate model based on task complexity, required performance and cost.  Organisations should apply the same discipline used in cloud computing: align resources to workload requirements. This approach improves efficiency while preserving access to advanced capabilities when they are truly necessary.

Agentic AI requires stronger governance

Agentic AI is becoming a priority because of its potential to automate complex workflows and accelerate decision-making. However, it also introduces new consumption-management challenges.

Unlike traditional applications, agents can take actions, evaluate outcomes and iterate repeatedly. Multiple agents may interact with each other, access tools, retrieve information and generate intermediate outputs before completing a task. Each step increases token consumption and operational cost.

As organisations expand their AI governance programmes, agentic AI introduces additional requirements beyond cost management. Organisations must establish clear guardrails, including execution boundaries, token budgets, monitoring requirements, approval thresholds, defined ownership models, and controls governing agent identities, privileges and access to enterprise resources.

The challenge is less technical than organisational. Whether responsibility resides with Enterprise Architecture, a dedicated AI governance council, a security function or a FinOps team with an expanded mandate, ownership must be explicit rather than assumed. Effective governance requires clear accountability for consumption, security, agent oversight and business outcomes.

AI governance begins with architecture

When AI costs rise, many organisations focus on prompt engineering. While prompt optimisation can improve efficiency, it rarely addresses the root cause.

The larger issue is fragmentation. Different teams often make independent decisions about model selection, retrieval architecture, context management, orchestration and monitoring. The result is duplication, inefficiency and limited visibility into enterprisewide AI consumption.

Enterprise Architecture has solved similar problems before. Whether governing cloud platforms, API ecosystems or service-oriented architectures, EA has historically helped organisations establish standards that improve scalability, consistency and cost control.

AI requires the same architectural discipline.

A mature AI architecture should include shared governance capabilities such as centralised orchestration, policy enforcement, context management, observability and consumption management. Rather than recreating these capabilities across business units, organisations should manage them as enterprise platforms.

AI FinOps provides the visibility needed to understand AI consumption and costs, but visibility is only valuable when it drives action. Without governance, accountability and architectural discipline, organisations can measure AI spending yet still struggle to control it, optimise it and tie it to business outcomes.

Enterprise Architecture provides the structure that allows AI FinOps to succeed. Together, EA and FinOps can establish standardised metrics, cost-allocation models, consumption accountability and performance benchmarks tied directly to business outcomes – just as cloud FinOps only became sustainable once governance and architectural discipline caught up with adoption.

This mirrors the evolution of cloud FinOps. Sustainable optimisation required governance, accountability and architectural discipline. AI will be no different.

What technology and business leaders should do now

To scale AI responsibly and maximise business value, organisations should:

• Establish standards for token consumption and reporting.
• Require model-selection rationale during architecture reviews.
• Define token budgets for AI applications and agentic workflows.
• Implement context-management and retrieval-optimisation standards.
• Evaluate emerging AI governance and observability platforms that provide auditable visibility into AI activity, helping leaders translate insights into action, accountability and measurable business value.
• Create visibility into AI consumption across business units and products.
• Expand architecture review board (ARB) criteria to include AI efficiency, scalability and cost governance.
• Align Enterprise Architecture and FinOps teams around shared business-value objectives.

These actions help organisations move beyond experimentation and create a repeatable framework for AI value realisation.

Final thoughts

The conversation around AI ROI is shifting. The question is no longer whether organisations should invest in AI, but how effectively they can govern, optimise and scale those investments.

The organisations that realise the greatest value from AI will not be those with the largest budgets or the most advanced models. They will be the ones that connect AI investments to a clear value creation strategy and establish the architectural foundations, governance frameworks and operational disciplines needed to deliver measurable business outcomes.

Cloud taught the enterprise this lesson once: innovation without governance erodes the business case. With AI, the meter runs faster. Enterprise Architecture and AI FinOps together turn consumption data into insight, insight into accountability and accountability into business value – before the bill arrives, not after.

Learn about Protiviti’s AI services: https://www.protiviti.com/hk-en/artificial-intelligence-services.

Don’t miss the entire AI FinOps conversation

This article is a continuation of Protiviti’s AI FinOps Insights Series. Explore related perspectives:

• Part 2: The Hidden AI Cost Problem: Lack of Visibility
• Part 3: coming soon
• Part 4: coming soon