Accounting for AI Development Costs Under US GAAP

Key Takeaways

• US GAAP contains no AI-specific accounting guidance, so entities generally apply existing software cost standards by analogy.
• Internal-use AI software, software sold only as a hosting arrangement when the customer cannot take possession, and implementation cost incurred in cloud computing arrangement that is a service contract (aka CCA or SaaS) generally fall under ASC 350-40.
• AI software sold, leased, or marketed externally generally follows ASC 985-20, with costs expensed until technological feasibility is established.
• For internal-use projects before ASU 2025-06 is adopted, capitalization begins after the preliminary project stage is complete and management commits to funding the project.
• LLM training, validation, and optimization costs may be capitalized during the development phase when they are analogous to coding and meet the applicable capitalization criteria.
• ASU 2025-06 removes project-stage references for internal-use software and introduces a probable-to-complete threshold after significant development uncertainty is resolved.
• Ongoing monitoring and maintenance costs are expensed as incurred unless the work adds new functionality.

AI adoption is accelerating as entities use the technology to improve operations, efficiency, and profitability. As AI becomes more embedded in core business processes, related development costs are becoming increasingly significant.

Many AI applications rely on foundation models, including Large Language Models (LLMs), which are pre-trained on large datasets and can be fine-tuned for specific tasks.

Accounting for AI development costs under US GAAP can be complex because the technology is evolving quickly and often requires systems tailored to specific operational needs. Accurate accounting is important because errors can affect earnings, balance sheet strength, debt covenant compliance, borrowing costs, and stakeholder expectations.

Which Accounting Standard Applies to AI Software?

Because US GAAP does not include AI-specific accounting guidance, entities generally apply existing ASC guidance by analogy. Internal-use AI software, including software sold only as a hosting arrangement, typically falls under ASC 350-40; AI software developed to be sold or marketed externally generally falls under ASC 985-20.

Large-scale AI system changes may also trigger business process reengineering considerations under ASC 720-45. This article focuses on AI software development costs, not broader business process changes.

How Do You Account for Internal-Use AI Software (ASC 350-40)?

AI software projects with the following attributes are guided by ASC 350-40 by analogy:

1. Software to be used internally upon meeting the two required characteristics as per ASC 350-40-15-2A (whereby the software is acquired, internally developed, or modified solely to meet the entity’s internal needs and during the software’s development or modification, no substantive plan exists or is being developed to sold, leased, or marketed externally)
2. Software marketed or sold only as a hosting arrangement by the entity, and the customer is not able to take possession thereof and

• The entity is the customer and obtains access to software in a hosting arrangement, and has the contractual right to take possession of the software at any time during the hosting period without significant penalty and it is feasible for the entity to either run the software on its own hardware or contract with another unrelated party

1. For the latter, if the entity is only able to take possession by incurring significant cost and/or not able to use the software separately without a significant diminution in utility or value, it is considered a service contract and only the implementation cost is capitalizable under ASC 350-40 (which is the most common scenario).

The capitalization of direct internal and external costs starts once the preliminary project phase is completed, and implicit or explicit approval by management with the relevant authority is obtained, including the commitment to fund the project (prior to the adoption of ASU 2025-06).

It is expected that, unlike non-AI software projects, a significant portion of the software-related costs will be incurred during the preliminary project stage, given the longer learning curve associated with advanced technologies, as well as the high-risk development issues that could affect project completion. Costs incurred during the preliminary project phase are expensed as incurred. The gathering and assessment of business requirements, determination of AI model performance and infrastructure requirements, cost comparisons, for example, off-the-shelf vs. self-development, scalability planning, and the like, are indicators that the project is in the preliminary phase. Indications that the preliminary phase is likely complete include an executed contract with a third party to develop the GenAI software, approval of expenditures for internal development, or a commitment to obtain an AI model from a third party. 

LLMs are trained using vast datasets to recognize patterns in text and generate coherent responses. The training process typically involves two main stages: pre-training and fine-tuning. These activities are similar to coding, as LLMs are trained on large code repositories and can perform tasks such as code completion, debugging, and generating code snippets. The training of LLMs follows the ASC guidance applicable to coding; by analogy, therefore, the internal and external AI software development cost, as well as the cost attributed to the training of the LLMs, validation of outputs, and optimization of activities, are to be capitalized during the development phase of the project.

Capitalization stops when the AI software is substantially complete and ready for use, generally after major testing is finished and the project is nearing integration and deployment.

Continuous monitoring and maintenance are necessary to ensure the model remains accurate and effective over time. This stage involves tracking the model's performance, updating it with new data, and addressing any issues that arise. None of these activities provides additional functionality, but they are intended to keep the AI model relevant and ensure it remains accurate and effective. The related costs are to be expensed as incurred, except if additional functionality is introduced, which is to be accounted for in accordance with the capitalization guidance.

How Do You Account for AI Software Sold, Leased or Marketed (ASC 985-20)?

GenAI software that will be sold or marketed externally is in scope of ASC 985-20 guidance and to be classified as an amortizable intangible asset. All costs are to be expensed as research and development cost in scope of ASC 730-10 until technological feasibility is accomplished, which is typically met upon. completion of the product design and working model, confirmed by testing, or the completion of a detailed program design for the software, which is less commonly used. The expectation is that research and development costs associated with AI models will be significantly higher, and that technological feasibility will be accomplished at a much later stage than their regular software counterparts, due to the novelty and high risk of failure associated with new technology.

Achievement of technological feasibility marks the start of capitalizing on the direct and indirect cost increments, clearly attributable to the AI software. Direct costs include software engineers’ payroll, benefits, and related payroll taxes, as well as external developer costs; LLM training costs; and indirect costs to be capitalized include overhead costs related to the AI project’s programmers and the IT infrastructure they use for the initiative. If software is purchased to train the LLM associated with the AI project, overhead is to be allocated, whereas general and administrative costs are not permitted to be allocated at the entity or department level. Capitalization of costs stops on the general release of the AI software to the customers. The amortization of capitalized costs is generally recorded in the cost of sales.

For AI software marketed or sold as a hosting arrangement, and the customer has the contractual right to take possession anytime during the hosting period without significant penalty, and it’s feasible for the customer to run the software on its own hardware or contract with another party unrelated to the entity to host the software, the AI software hosting arrangement is covered under ASC 985-20 and not ASC 350-40.

Foundation Models and Large Language Models (LLM’s)

Gen AI technology includes LLMs, multimodal AI (text, image, audio, and video), and AI agents that execute multi-step workflows autonomously. For all AI applications involving one or more foundation or LLM models, entities need to determine whether the arrangement entered into with a vendor is for a service, or for the transfer of a foundation model/LLM license.

Foundation models or LLM’s obtained as a hosting arrangement is considered a service whereby only the implementation cost is capitalizable under ASC 350-40, when the reporting entity as the customer is not able to take possession of the foundation model/LLM software, or in the event it has the contractual right during the hosting period to do so at a significant penalty or at a significant diminution in utility or value.

In the contrary, when the entity can contractually take possession of the foundation model/LLM software during the hosting period without significant penalty and has the ability to do so without incurring significant cost and it is feasible for the entity as the customer to either run the software on its own hardware or contract with another party unrelated to the vendor to host the software, while having the ability to use the software separately without a significant diminution in utility or value, the hosting arrangement is not considered a service and ASC 350-40 applies in its entirety. Contracts entered into with a vendor for the transfer of a foundation model/LLM License apply ASC 350-40.  Most entities enter into service agreements.

Most development-type activities that involve obtaining or creating code/configuration specific to the entity’s environment for the software to perform as intended are generally capitalizable. In contrast, user- training, preliminary project-phase activities, and post-go-live activities are expensed as incurred. In all scenarios except for service agreements, the capitalized  cost is normally disclosed as part of the intangible asset, while for a service agreement it is normally considered prepaid cost within other assets in the Balance Sheet and amortized straight-line over the non-cancellable term, increased by renewal periods the entity is reasonably certain they would extend the foundation model/LLM access service by, determined at contract inception.

The amortization cost associated with implementation cost is typically presented in the same line item as LLM fees on the Income Statement/Profit and Loss Statement . The prepaid asset raised is subject to impairment if the related service period is shortened or the foundation model/LLM service is abandoned.  

For the non-service agreement scenarios, the amortization cost is based on the expected life of the asset and presented as such in the Income Statement/Profit and Loss Statement, and ASC 360 is applied for impairment.

What’s Ahead: FASB ASU 2025-06

In September 2025, the FASB issued ASU 2025-06, Intangibles – Goodwill and Other Internal-use Software, effective for all entities with annual reporting periods beginning after December 15, 2027, and for interim reporting periods within those annual reporting periods, with early adoption permitted.

Currently, capitalization of cost starts when both the following occur:

1. The preliminary project stage is completed, and
2. Management, with relevant authority, implicitly or explicitly authorizes and commits to funding a project, and it is probable that the project will be completed and the software will be used to perform the function intended.

The pronouncement eliminates the need for clearly defined, sequenced project development stages, thereby neutralizing the guidance for easy application to any software implementation method, whether it is a waterfall, agile, or another strategy. The revised standard introduces a threshold for “probable to complete,” allowing capitalization of development costs once the significant development uncertainty associated with the software is overcome.

Two criteria are introduced to determine if significant uncertainty exists:

1. Software driven by technological innovations or novel, unique, or unproven functions or features needs to resolve any identified uncertainty related to those features or functions through coding and testing.
2. Significant performance requirements (for example, functions or features) need to be identified with no significant revisions anticipated.

This introduces additional judgment in determining when and whether to capitalize costs. It does not change the nature of the capitalizable costs. The new standard eliminates intangible disclosures for capitalizing internal-use software; instead, the disclosure requirements of ASC 360-10, Property, Plant, and Equipment, are to be followed.

How EisnerAmper Can Help

Accounting for AI development costs requires significant judgment, and the guidance continues to evolve as standards such as ASU 2025-06 take effect. EisnerAmper's Technical Accounting Advisory team helps organizations evaluate which standards apply, determine when costs should be capitalized or expensed, and prepare for new reporting requirements.

To discuss how these considerations apply to your organization, contact our Technical Accounting Advisory team using the form below.