Capitalizing Token Costs
Imagine. Your company has unleashed the power of agentic AI in the hands of every employee. Everyone’s productivity is growing 10x- 100x; your sales team delivers growth that you could never imagine, and the costs of AI platforms used to achieve these results are incomparably small compared to the benefits. If this is what’s happened to your business, this article is not for you. Most companies that have implemented agentic AI capabilities have not fully captured additional value; however, they have undoubtedly incurred the full incremental costs. How should we account for these incremental AI costs?
There is no separate accounting model for AI token costs. The treatment should be based on the nature of the activity consuming tokens. Token costs may therefore be capitalized when they are directly attributable to qualifying software development activities during the applicable capitalization period, but the same token costs are expensed when they relate to experimentation, maintenance, ordinary business operations, or other noncapitalizable activities
At the end of this article, you will see a list of useful resources on the topic that cover a lot of questions surrounding accounting for token costs. Below, we will focus on one topic: whether it is appropriate to capitalize token costs and, if so, what conditions must be met for capitalization.
Why capitalize AI token costs?
Tokenmaxing behaves like a maximax strategy in a zero-sum game. According to game theory, maximax is not the best strategy to play in a zero-sum game (assuming the effects of information asymmetry are negligible). It can lead to unpredictable losses and uncertain gains. An organization's goal should be to maximize the positive difference between expected gains and losses while also avoiding scenarios where this difference is negative.
As an economist, I have certain notions deeply ingrained in my head: the perception that costs incurred for my business should ultimately increase its profits. This increase in profits can arise either from income created (e.g., incremental contribution margin from sales resulting from additional advertising spending) or because incurred costs are lower than alternative costs (i.e., compliance costs that allow one to avoid fines and penalties for non-compliance).
However, we also have to consider the fact that both alternative costs and incremental income are non-deterministic features of our reality. This is why the success or failure of any single decision is almost never defined by performance. To evaluate performance in a way that maximizes yield on capital, we need to allocate capital based on the respective probabilities of alternative costs and incremental income.
Cost capitalization extends this principle by assessing whether additional contribution (i.e., benefits or cost savings) from particular costs is predictable enough for incorporation into financial forecasts, plans, and scenarios. The contribution is predictable when it is directly attributable to the asset's utility, the asset can be acquired at known costs, both contribution and utility can be measured in finite numbers, and such measured values make the project or enterprise commercially viable.
Vibe-coded solutions created by non-developer members without formal due process cannot produce a predictable economic contribution to the business, as these solutions are more akin to experimental prototypes. However, an internal or external software development team engaged by management may accept this prototype as an input and build a production-grade version. Further, when developing a production-grade version of an application, the software development team may incur AI token costs that are more likely to be perceived as a predictable contribution to the business than the AI token costs consumed during vibe-coding app development.
When considering whether AI token costs are eligible for capitalization, we need to review the underlying economics of AI token costs. Capitalization of AI token costs under the following two strategies might be reasonably viewed as potentially eligible for capitalization:
Cost reduction strategy:
The new capability should either:
expand a single employee’s capacity, allowing them to perform the work of multiple peers, or staff employees, or contractors currently engaged to perform this work; or
remove the need for a human employee whatsoever.
The additional costs of implementation and ongoing operation of the new solution are determined to be lower than the cumulative costs of operating the existing solution over the horizon of 2-5 years;
Profitable growth strategy:
The new capability should increase sales by an amount such that the corresponding contribution margin exceeds the additional costs of implementing and operating the new solution.
When management’s strategy does not allow articulation of specific results and the anticipated AI token costs needed to achieve those results, capitalization is generally not appropriate under accounting standards, underlying economics, or common sense.
When (if ever) is it appropriate to capitalize AI token costs?
Generally, AI token costs consumed by your finance or marketing team to create process automations for internal use would NOT be capitalizable. Even if they create functional prototypes of internal-use software that saves teams a significant amount of time.
However, when costs of certain software development projects are capitalizable, direct AI token costs consumed on such projects would also be capitalizable. The ability to capitalize costs depends on whether the entity tracks AI token costs by activity level or has a reasonable proxy for that tracking (e.g., when an employee is fully engaged in a single development project eligible for cost capitalization during the reporting period). Implementing FinOps best practices, including resource tags, will enable entities to achieve that required level of transparency.
When evaluating whether costs of software development are capitalizable, we first need to determine which accounting guidance applies:
• Subtopic 350-40: Software that is solely for the entity’s internal use (e.g., payroll processing, ERP, inventory stock management, agentic process automation) and software sold to customers exclusively on a hosted basis.
• Subtopic 985-20: Costs to develop software for sale or licensing to customers, including as part of a product or providing a service.
• Subtopic 730-10: Costs to create code solely for a particular project that are experimental in nature (i.e., considered R&D) and have no alternative future use are expensed as incurred.
• Subtopic 340-40: Fulfillment costs guidance applies to software that an entity develops for others under a contractual arrangement.
• Subtopic 350-30: Other costs of internally generated agents/automations generally require expensing as incurred unless they arise from identifiable, contractual-legal rights.
Once an entity identifies the applicable guidance, it applies it to AI software development without specific exceptions or different requirements than for other types of software development.
US GAAP does not specify income statement classification of development costs; therefore, it depends on the facts and circumstances [350-40-25-12, 25-14]. Generally, costs of software developed for use in the entity’s internal functions (accounting, sales, BD) are classified based on their intended functional use. Costs of software developed for hosted access by customers are classified as R&D.
ASU 2025-06
Under ASU 2025-06, all references to project stages in ASC 350-40 have been removed. However, the ASU did not change the guidance in ASC 985-20.
Under the new guidance, software costs should be capitalized once both thresholds have been met:
Authorization threshold: Project is authorized, and funds are committed.
Probable-to-complete threshold: Project completion is probable, and it is expected that the software will be used to perform the function intended (i.e., no significant development uncertainty exists).
Eligible development costs incurred after the above criteria are met and until all substantial testing is completed are capitalized. Costs incurred after all substantial testing is completed are expensed as incurred (unless they are capitalizable upgrade or enhancement costs). [350-40-25-6]
The following characteristics might be useful in the assessment of whether the project is eligible for capitalization:
• The project has documentation such as an approved implementation plan
• The software has novel functions or features with the uncertainty not being resolved through coding and testing.
• The significant performance requirements of the software have not been identified or continue to be substantially revised.
• A working prototype of the software has been created.
• Technological feasibility has been established.
Upgrades and enhancements
Upgrades and enhancements are defined as changes to existing software that result in additional software functionality (i.e., the ability to perform additional tasks), extend the life of externally marketed software, or significantly improve an existing software product’s marketability. Other updates (i.e., those that do not meet the definition of a product enhancement) and activities to correct errors or keep the software updated with current information are considered maintenance.
Identifying upgrades and enhancements often requires judgment to determine whether tokens spent create new functionality or maintain existing functionality without adding new capabilities or tasks the software can perform (i.e., merely allow the software to continue executing its current tasks).
Token costs used to upgrade or enhance software follow the same capitalization requirements as costs of new software development. Token costs used solely to maintain software are expensed as incurred.
Strategic Accounting with Andrei
Want to learn about integrating business strategy into your accounting process? Curious about the latest guidance? Or maybe you just want to get quick wins?
Our newsletter offers real insights from practitioners. If you have a topic you’d like to discuss, book some time with me!
Improve Your Accounting Strategy
Is your accounting strategy under-utilized or non-existent? Schedule some time with me for a complimentary call to learn how you can strengthen your process.
In case you missed it….
Rare Gems
Determining Useful Lives of AI Assets and Accounting Considerations in AI Projects by CBIZ provides a great summary of accounting for development costs at companies that develop or fine-tune AI models.
Unlocking agentic value: a new investment discipline for the agentic era by EY. This article offers great insights into how the AI service cost structure has changed over time and what initial steps AI consumers can take to regain control.
Rethinking AI TCO: Why Cost per Token Is the Only Metric That Matters by Shruti Koparkar. An in-depth analysis of AI cost management by one of the few companies that benefits most from additional AI spending (Nvidia).
Giving AI Billing a Common Shape by FinOps Open Cost and Usage Specification (FOCUS). Structured approach for analyzing AI-related bills, primarily for technology businesses.
Managing AI Value Using FinOps Practice Operations by FinOps Foundation.
AI tokenomics: A CFO’s guide to governing the AI P&L by Deloitte. Advocating the view that AI
Hot Topic: Software data costs - Accounting considerations by KPMG.
Insights by Tokenomics Foundation.
Claude Tokens Compute Cost Code by Wired. A nice read about how companies achieve results that they think are worth all that extra AI token spend.
Where AI Costs Belong on the P&L by OnlyCFO’s. Great post on allocating AI expenses by P&L line items.
How are AI agents spending your tokens by Matty Smith, DigDig, May 5, 2026. The quote worth quoting: “Even when running the same agent on the same task, costs varied by up to 30x.”
How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks by Longju Bai, Zhemin Huang, Xingyao Wang, Jiao Sun, Rada Mihalcea, Erik Brynjolfsson, Sandy Pentland, Jiaxin Pei