AI Costs Are Moving Into Operating Budgets

The latest AI cost-management evidence points to a significant change in how businesses manage technology spending. In September 2026 research cited by IDC, 90.5% of surveyed enterprises had exceeded their 2026 token budgets by August, while only 9% could proactively review AI costs at least weekly. Earlier, the FinOps Foundation reported that 98% of surveyed practitioners were managing AI spend in 2026, up from 63% in 2025. These findings suggest that AI cost management is becoming a recurring operational responsibility rather than a one-time technology investment.

The shift has a simple explanation. Many AI products are now priced partly around how much intelligence a business consumes, not just how many software licenses it purchases.

That changes budgeting, architecture decisions, and ultimately the economics of automation.

CTA: Model your AI costs. Estimate the cost of each AI-supported workflow before expanding it across the organization.

AI Cost Management Has Become a Mainstream Priority

Traditional enterprise software budgeting is comparatively predictable when a business purchases fixed annual subscriptions.

A company might purchase 100 user licenses, pay an implementation fee, and estimate its annual expenses using a contractual rate. Actual costs can still vary with usage, storage, support, or additional features, but the baseline is relatively clear.

AI introduces another dimension.

An AI-powered support assistant may generate a different number of requests every day. Each request can involve different input volumes, output lengths, context retrieval, and additional model calls.

The cost depends on what the system actually does.

The FinOps Foundation’s February 2026 State of FinOps survey included 1,192 respondents. It reported that the share managing AI spending rose from 63% in 2025 to 98% in 2026. The results reflect the growing scope of FinOps teams, not proof that every enterprise has mature AI cost controls.

FinOps respondents managing AI spend

The important change is not simply growing AI spending. It is the need to attribute that spending to specific products, teams, and business outcomes.

Why AI Costs Behave Differently From Traditional Software

AI spending is not exclusively an operating expense. Some development and implementation costs may qualify for capitalization depending on applicable accounting rules, while AI subscriptions, inference, infrastructure, and support can create recurring expenses.

The distinction in this article is primarily about budget management, not a universal change in accounting classification.

Consider three spending categories.

Cost typeTraditional software exampleAI example
FixedAnnual licensesReserved GPU capacity or fixed AI subscriptions
VariableUsage-based cloud storageTokens, requests, tool calls, inference
HybridPlatform subscription with overagesAI platform fee plus model consumption

AI can combine all three.

A company may pay for the application, the model inference, the retrieval database, and the infrastructure used to operate the workflow.

The variable components become particularly important when usage grows.

An internal assistant that handles 1,000 requests per month may be inexpensive during a pilot. Deploying the same assistant across multiple departments can change the financial equation even if the underlying technology remains unchanged.

That is why AI architecture needs an operating-cost model alongside its technical design.

The Real Cost of an AI-Powered Business Workflow

The difference becomes clear when we examine the economics of model selection.

As of November 2026, OpenAI’s published standard text pricing includes GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output tokens, and GPT-6.1 Sol at $2 per million input tokens and $10 per million output tokens. These rates exclude other possible charges and are subject to change.

Assume an organization processes 100,000 AI-supported tasks each month.

For this illustration, every task uses 2,000 input tokens and generates 500 output tokens in one model call.

Illustrative monthly direct inference cost

GPT-6 Luna GPT-6.1 Sol

$45 $900

Per 100,000 tasks Per 100,000 tasks

Calculated from published November 2026 standard token rates. Hypothetical equal token volumes; not a quality-equivalent benchmark. Excludes caching, tools, retries, infrastructure, and other charges.

This is a 20-fold difference in direct inference spending.

But price alone does not establish value. The less expensive model may handle structured extraction or simple classification effectively while a more capable model may be necessary for complex reasoning.

The correct metric is therefore not simply cost per token.

It is cost per successfully completed business task at the required quality level.

The Hidden Costs of AI Workflows

Token pricing captures only part of the operating expense.

An AI workflow may also require document retrieval, embeddings, vector storage, monitoring, tool execution, security controls, and human review.

Retries and multi-step reasoning can multiply model calls. A task expected to need one model request might use three when the system retrieves evidence, generates a response, and checks its output.

Using the same illustrative token volumes, three sequential calls per task would increase the direct inference portion from $45 to $135 on Luna or from $900 to $2,700 on Sol.

Actual costs depend on the token consumption of each call.

CTA: Model your AI costs. Include workflow frequency, model calls, infrastructure, retries, and human review in the estimate.

Three Second-Order Effects of Usage-Based AI

1. Better Models May Produce Worse Business Economics

The most capable model is not always the most economically appropriate model.

Consider an AI system that classifies customer support tickets.

If a smaller model meets the required accuracy and reliability threshold, using a premium model for every ticket may increase expenses without creating proportional business value.

A better architecture may route straightforward tasks to a lower-cost model while escalating ambiguous cases to a more capable one.

However, routing also introduces engineering complexity. Teams must evaluate escalation errors, latency, monitoring requirements, and quality regressions.

The objective is not to minimize model spending at any cost.

It is to meet the required service level at the lowest sustainable total cost.

2. Successful AI Adoption Can Create a Margin Problem

Usage-based spending creates an unusual challenge.

An AI feature can become popular and financially less attractive at the same time.

Suppose a business charges customers $10 per month for an AI-enabled feature. If heavy users consume $12 in model and infrastructure costs, the feature has negative gross contribution before considering other expenses.

Increasing adoption may worsen the problem unless the company changes its pricing, usage limits, model selection, or workflow architecture.

This is a hypothetical illustration, but the business principle matters.

AI products need unit economics tied to user behavior rather than average API pricing alone.

Companies should examine spending by customer, workflow, request type, and successful outcome.

3. Cost Controls Become a Product Design Responsibility

AI spending cannot remain solely the responsibility of finance teams reviewing monthly invoices.

Developers influence costs whenever they choose a model, expand a prompt, add retrieval steps, increase context windows, or introduce another agent.

Architectural decisions therefore become financial decisions.

The FinOps Foundation’s FOCUS specification provides a standardized approach to normalizing technology billing data. Version 1.4, ratified in June 2026, expanded support for financial reconciliation and commitment analysis.

This can help financial and engineering teams establish a common reporting foundation.

However, external billing normalization does not replace application-level telemetry. A company still needs its own mapping between infrastructure consumption, workflow execution, and business outcomes.

How to Build an AI Operating Budget

A practical AI budget should begin with a measurable unit of work.

That might be one resolved support ticket, one processed invoice, one generated report, or one completed customer interaction.

Estimate the workload using five inputs:

  • Volume: Number of tasks per month.
  • Consumption: Average input and output tokens per model call.
  • Execution: Average model calls, retries, and tool operations per task.
  • Supporting costs: Infrastructure, databases, monitoring, and human review.
  • Business outcome: Percentage of tasks completed successfully without unacceptable errors.

From there, construct a baseline and a growth scenario.

For example, a workflow expecting 100,000 tasks per month should also be evaluated at 200,000 and 500,000 tasks.

Model changes, higher user adoption, and increased request complexity can affect the forecast independently.

A simplified budgeting formula is:

The next step is operational ownership.

Each production workflow should have a designated business owner, technical owner, spending allocation, and acceptable quality threshold.

Teams should track cost per completed task, budget variance, token consumption, and the proportion of work requiring additional model calls.

Monitoring also needs enforcement mechanisms.

Google Cloud, for example, distinguishes ordinary budget alerts from supported spend-cap budgets. Alerts-only budgets do not stop spending automatically, while eligible spend caps can pause covered services, subject to documented limitations.

For critical AI applications, businesses should combine cost controls with graceful degradation, fallback behavior, and customer-impact planning.

A cost limit that unexpectedly stops an essential workflow is itself an operational risk.

So What for Business Leaders?

The transition toward usage-based AI changes how companies should evaluate technology investments.

A pilot budget answers whether the business can afford to experiment.

An operating budget answers whether it can afford to scale.

Those are different questions.

Business leaders should establish a repeatable process for forecasting, observing, allocating, and reviewing AI consumption.

Frequently Asked Questions

1. What are AI operating costs?

AI operating costs are the recurring expenses required to run and maintain AI systems. These can include model inference, API calls, cloud computing, data retrieval, monitoring, subscriptions, security operations, and human review.

2. Why are AI costs becoming harder to predict?

Many AI services use consumption-based pricing. Expenses can change with the number of users, requests, input and output tokens, workflow complexity, and automated tool calls. These variables make fixed annual estimates less reliable.

3. How do you calculate the operating cost of an AI application?

Estimate monthly task volume, model calls per task, input and output tokens, and applicable model prices. Add infrastructure, tools, storage, monitoring, support, and human review expenses. Divide total recurring cost by completed tasks to understand unit economics.

4. What is the difference between AI CapEx and OpEx?

Capital expenditure (CapEx) generally relates to investments in assets or qualifying development expenditure. Operating expenditure (OpEx) covers recurring costs of running a business. AI investments can involve both, depending on contracts, accounting standards, and the nature of the spending.

5. How can businesses reduce AI inference costs?

Businesses can evaluate smaller models, use prompt caching where appropriate, reduce unnecessary context, batch eligible workloads, introduce model routing, and limit redundant calls. Any optimization should be tested against output quality and business requirements.

6. What is AI FinOps?

AI FinOps applies financial accountability and operational cost-management practices to AI workloads. It connects AI consumption to teams, applications, cost centers, and business outcomes, helping organizations forecast spending, allocate costs, and improve efficiency.

Leave a Reply