In 2024, 31% of FinOps practitioners were managing AI spend. In 2025, 63%. In 2026, 98%.

That trajectory tells you everything you need to know about how quickly AI has moved from an experimental budget line to a core operating cost. The FinOps Foundation's State of FinOps 2026 report, drawing on survey data from FinOps practitioners across global enterprises, confirms it explicitly: AI cost management is the number one most-requested capability across the entire survey. Commercial tooling has not yet delivered it at scale.

73% of enterprise agentic AI implementations went over budget in 2026. Some by more than 2.4 times their original estimates. 72% of organisations experienced an unexpected AI cost spike or surprise bill in the last 12 months.

These are not the numbers of a cost category under control. They are the numbers of a cost category that most enterprise finance functions were not built to manage.

AI FinOps is the discipline that closes this gap. Here is what it is and what it requires.

What AI FinOps Actually Is

FinOps, short for financial operations, is the practice of managing cloud and technology costs through shared accountability between finance, engineering, and business teams. It emerged from the shift to cloud computing, where the move from capital expenditure hardware to consumption-based cloud services created a new cost management challenge.

AI FinOps applies the same principles to AI workloads. But AI costs are more complex than cloud costs in ways that matter for how they are managed.

Cloud cost is primarily infrastructure: compute, storage, networking. The cost drivers are relatively well-understood and the optimisation levers are relatively well-established.

AI cost spans multiple layers simultaneously. Model usage, measured in tokens consumed. Infrastructure, the GPU compute required to run the models. Data costs for training, fine-tuning, and retrieval. Software licences for AI platforms and tooling. Orchestration costs for the frameworks that coordinate agent workflows. And governance costs for the monitoring, safety, and compliance infrastructure that regulated enterprises require.

Each of these layers has different cost dynamics, different attribution challenges, and different optimisation approaches. The FinOps Framework's 2026 update explicitly adds AI as a distinct technology category because managing AI spend requires different approaches from managing cloud infrastructure spend.

AI FinOps is the practice that connects these layers into a managed discipline. It establishes visibility into where AI cost is being generated, accountability for who is generating it, governance structures that keep costs predictable, and value measurement that connects cost to business outcome.

Why Traditional Finance Processes Break Down for AI

Enterprise finance functions were built to manage costs that are either fixed or change slowly. Software licences renew annually. Hardware depreciation follows a schedule. Headcount costs are predictable within a planning cycle.

AI costs break all of these assumptions.

AI costs are consumption-based. They scale with every query, every agent step, and every workflow. A new feature that doubles usage doubles cost without any new procurement decision. A poorly designed prompt that generates excessive token consumption in refinement loops generates cost at a rate that no budget review will catch until the invoice arrives.

AI costs are distributed across multiple providers. A large enterprise may be paying for OpenAI API access, Anthropic Claude Enterprise, Azure OpenAI services, AWS Bedrock, Google Vertex AI, and GPU instances from multiple cloud providers simultaneously. Normalising this into a coherent cost picture requires aggregation work that most existing financial systems cannot do automatically.

AI costs are generated by engineers, not by procurement. In traditional enterprise IT, significant costs are incurred through procurement processes that create visibility. AI costs are incurred by engineers writing code that makes API calls. The engineer making the model selection decision and designing the prompt architecture is generating the majority of the cost impact, but in most organisations they are not working within a budget constraint that makes them feel the financial consequences of those decisions.

The structural disconnect is precisely described in FinOps reporting. Engineering and platform teams hold 35% of the influence over cost-driving decisions including model selection, prompt architecture, and retry logic. But finance accountability for those same decisions sits at 27% of the organisation. When the teams making cost-generating decisions do not feel the budgetary consequences, governance collapses.

The Specific Cost Dynamics That AI FinOps Has to Manage

Token costs and consumption patterns.

Every interaction with an AI model consumes tokens. Input tokens, the prompt and context sent to the model, and output tokens, the response generated. Both are priced. Context length is the primary driver of per-interaction cost, as described in the AI inference cost piece. AI FinOps tracks token consumption by team, by application, by model, and by use case to identify where consumption is high relative to the value generated.

GPU infrastructure and utilisation.

For enterprises running models on their own or managed infrastructure, GPU utilisation is a significant cost centre. GPUs are expensive to provision and expensive to leave idle. AI FinOps monitors GPU utilisation across workloads, identifies underutilised capacity, and informs decisions about when to use cloud GPU instances versus owned infrastructure.

Agentic workflow costs.

AI agents generate token costs across every step of every workflow, including the refinement loops that account for 60% of agentic AI spend. A poorly governed agent workflow can generate costs that are orders of magnitude higher than the value it delivers. AI FinOps tracks cost per completed workflow task rather than total spend, which reveals whether individual workflows are economically viable.

Multi-provider sprawl.

20% of large enterprises are now spending more than $1 million per month on AI costs. That spend is typically distributed across multiple providers with different billing models, different units of measurement, and different invoice structures. AI FinOps normalises this into a unified attribution model that makes cross-provider spending visible and comparable.

Total cost of ownership beyond the model.

AI total cost of ownership includes infrastructure, labour for deployment and maintenance, data preparation and storage, governance and compliance tooling, and the cost of model deprecation and replacement cycles. Most enterprise AI business cases calculate licence costs and ignore the rest. AI FinOps tracks total cost of ownership as the basis for investment decisions.

What AI FinOps Looks Like in Practice

The FinOps Foundation's 2026 framework describes three stages of AI FinOps maturity.

• Crawl: visibility.

The starting point is understanding where AI cost is being generated. Which teams, which applications, which models, which use cases. This requires instrumentation that captures token consumption, GPU utilisation, and infrastructure cost and attributes it to the business function generating it. Most enterprises at this stage are discovering that AI spend is significantly more distributed than they assumed, and that the largest single spend items are often from uses that did not go through any formal procurement or approval process.

• Walk: accountability.

Once costs are visible and attributed, accountability structures can be established. Business units are assigned responsibility for the AI costs their teams generate. Token budgets are set per application and per team. Engineers see cost estimates for their model and prompt choices before they deploy to production. Developers are issued personal token budgets as managed resources, the same way travel budgets work. Unusual consumption patterns trigger alerts rather than being discovered at month-end invoice review.

• Run: optimisation and value measurement.

At this stage, AI FinOps is actively optimising the cost structure. Model routing directs queries to the most cost-effective model that can complete the task. Caching reduces redundant inference calls. Context length management reduces token consumption per query. And cost is measured against business value, not just against the budget, with cost per completed workflow task, cost per resolved customer query, and cost per AI-assisted decision as the metrics that connect spend to outcome.

Only 8% of enterprise FinOps teams report to a CFO. 78% report to a CTO or CIO. This organisational placement reflects the reality that AI cost optimisation requires technical decisions, not just financial ones. The governance framework needs to bridge both.

What Finance Teams Specifically Need to Understand

Enterprise finance teams do not need to understand the technical details of how AI models work. They do need to understand five things that are different about AI costs from the technology costs they have managed before.

• AI costs scale with usage in ways that cannot be capped by procurement decisions:

A new software licence has a fixed cost. A new AI use case has a variable cost that scales with adoption. The finance team needs forecasting models that account for usage growth, not just the initial procurement commitment.

• The teams generating AI costs are not the teams accountable for AI budgets:

Bridging this gap requires chargeback models that make business units accountable for the AI costs their teams generate, not just for the AI budget allocated to them centrally.

• AI ROI needs to be measured at the use case level, not the programme level:

A programme-level ROI that averages across all AI use cases obscures which specific uses are generating value and which are consuming budget without return. Finance teams need use case level economics.

• Model deprecation is a cost event that needs to be planned for:

AI models are replaced more frequently than enterprise software. When a model is deprecated, the applications built on it need to be migrated. This is a cost that rarely appears in AI business cases and consistently surprises finance teams.

• Governance tooling is a cost centre, not a cost reduction:

The monitoring, audit, and compliance infrastructure required for enterprise AI adds to the total cost of AI deployment. It is not optional for enterprises with regulatory obligations. It needs to be in the financial model from the start.

Vishleshan AI's forward deployed engineering (FDE) approach builds AI systems with cost governance designed in from the start. Model routing, token budgets, inference cost attribution, and total cost of ownership modelling are part of every engagement architecture. The goal is AI that is economically sustainable at production scale, not just impressive at pilot scale. Book a Consultation