Accounting Considerations in AI Projects

Introduction

The rapid advancement of artificial intelligence (AI) technologies is causing more and more companies to consider developing or purchasing software that leverages AI to enhance internal processes or to be incorporated into their products and services to customers. Generative AI technologies have become integral to applications ranging from customer service chatbots to advanced data analytics. However, as organizations invest heavily in developing or licensing these sophisticated models, they face complex accounting challenges. Traditional accounting frameworks that were designed for more linear software development processes can be challenging to apply to the iterative and data-intensive nature of AI model and application development.

In this article, part of our ongoing AI Accounting Insights series, we explore the accounting considerations for entities developing generative AI technology, specifically discussing how to apply existing U.S. GAAP accounting frameworks to the costs incurred in developing foundational models or AI applications. Our goal is to provide insights on how management can bridge the gap between evolving AI technologies and traditional accounting models, helping financial leaders to ensure accurate financial reporting and compliance.

Development of AI Technologies

Generative AI refers to a type of AI that can produce new content — whether it’s text, images, code or even music — based on the patterns it has learned from existing data. Generative AI is capable of generating content that is novel, coherent and contextually appropriate. This technology is powered by advanced machine learning techniques that enable computers to understand and mimic the complexities of human creativity and language.

One of the key innovations behind generative AI is the development of large language models (LLMs). LLMs, such as GPT-4 and similar models, are designed to process and generate human-like text. They are built using deep learning techniques and trained on vast amounts of data, which may include books, articles, websites and other written content. The model learns the structure of language, including grammar, context and the relationships between concepts, which enables it to generate relevant and contextually accurate responses. By leveraging LLMs through generative AI applications, businesses can automate and enhance a variety of language-intensive tasks, leading to increased efficiency and innovation.

LLMs form the backbone of generative AI applications and are increasingly being integrated into various business applications. The development of LLMs and generative AI applications requires substantial investments in both resources and infrastructure. Developing and training LLMs and generative AI applications differs from more traditional software development and includes acquiring large datasets used for training models, employing advanced computing power and utilizing highly specialized talent.

LLMs can be developed internally, which can be extremely labor- and resource-intensive, or entities can fine-tune existing LLMs in their applications rather than create their own. Entities building their own LLMs to power their AI applications can expect several key stages in designing and developing their LLM, each with associated costs:

1. Conceptualization and Design: Defining the model’s purpose, architecture and algorithms. Costs include salaries for AI researchers, data scientists and engineers engaged in exploratory studies and model experimentation.
2. Data Acquisition and Preparation: LLMs require enormous datasets, often sourced from the internet, books and other texts. Costs include purchasing or licensing datasets, data cleaning, labeling, preprocessing and ensuring compliance with data usage rights. In addition, entities may incur significant storage costs for these datasets, through the purchase or leasing of hardware or purchase of cloud storage services from third parties.
3. Training: The most resource-intensive phase, requiring substantial computational power, often through high-performance hardware or cloud-based solutions. Training can take weeks or months, depending on the model’s size and complexity, and is one of the largest cost drivers.
4. Fine-tuning: Adapting the model for specific tasks or industries by retraining on a smaller, specialized dataset. This enhances the model’s performance for specific business applications.
5. Testing and Validation: Ensuring accuracy, reliability, and alignment with intended use. This phase identifies and corrects biases or errors, ensuring the model is ready for deployment in real-world scenarios.
6. Deployment and Maintenance: Implementing the model into production environments and ongoing maintenance to update and improve the model as needed.

Licensing an existing LLM to power generative AI applications through fine-tuning for specific purposes accelerates time to market and reduces costs, though it may come at the expense of control and level of customization. Building AI applications through an existing model shortens the process and can include the following stages and costs:

1. Licensing: Obtaining a license for an LLM from provider. Licensing fees vary depending on usage needs but are generally lower than the costs of developing a model.
2. Fine-tuning: This typically involves training the LLM for the entity’s purposes by fine-tuning it using specific datasets. Costs include acquiring and preparing custom data, computational resources for retraining, and paying AI specialists and data scientists.
3. Testing and Validation: Ensuring the customized model’s outputs are accurate and reliable, checking for performance issues and biases.
4. Deployment and Maintenance: Implementing the model into production environments and ongoing updates as business needs evolve.

In addition, entities may also incur more traditional software development costs when creating a generative AI application, such as costs related to developing the software application user interface, infrastructure, graphics, and content.

Current Accounting Framework

As generative AI applications and LLMs are a form of software, we believe that general software development accounting considerations as outlined under ASC 985-20 and ASC 350-40 would apply to developing these technologies.

Under software development, different guidance applies depending on whether the software (in this case the LLM or application) is being developed for internal use or for external sale or marketing. AI technologies developed for internal use or to be sold via a hosting arrangement would apply the capitalization guidance in ASC 350-40 on internal-use software. If the AI technology is being developed to sell or market via software license, ASC 985-20 would be applied.

AI Technologies Developed for Internal Use

Software is considered internal use if it is not intended for sale or external licensing, even if customers access it. Similar to SaaS applications, generative AI applications or LLMs can be hosted on the entity’s servers (or via hosted cloud providers) and provided as a service, meaning the entity retains control over the software and model.

Under the current ASC 350-40 internal-use software framework, costs tend to relate to software developer and engineer labor largely and are categorized based on the stage of development:

1. Preliminary Project Stage: Activities include conceptualization, evaluation of alternatives and determining technology requirements. Costs incurred during this stage are expensed as incurred.
2. Application Development Stage: Activities include design of the chosen path, coding, hardware installation and testing. Costs incurred during this stage are capitalized.
3. Post-implementation Stage: Activities include training and maintenance. Costs incurred during this stage are expensed as incurred.

In applying the current internal-use software development guidance to the development of generative AI applications and LLMs for internal use, entities will need to carefully evaluate processes and map to the three stages outlined in ASC 350-40. Costs incurred during the conceptualization stage in which the model’s purpose and architecture is being determined and designed should be expensed, as these activities most clearly map to the preliminary project stage under guidance. The preliminary project phases for many AI applications may be longer than those for other software development projects given the novelty of the technologies being used and due to higher-risk development issues affecting the successful completion of the project. These factors can also increase the development risks of these projects, which may impact the cost capitalization of the project.

Once management has judged that the planning and design phase is complete, the key concepts and framework of the LLM or generative AI application should be agreed upon, and management should commit to funding the project through completion. This typically triggers the start of the application development stage, where entities will need to identify and capitalize the direct internal and external costs incurred in developing the AI application or LLM. The training phase of an LLM can be compared to coding traditional software during its development phase, as both involve building a system to perform specific functions. While traditional software relies on manual coding in which developers specify how the software should operate, AI technologies rely heavily on machine learning algorithms that enable the model to learn from data; the “behavior” of the model emerges from the training process by analyzing vast amounts of data to recognize patterns through guided feedback.

The significant costs incurred during LLM training typically include compute and infrastructure expenses, data acquisition and preparation costs, and compensation for data scientists, AI engineers, and researchers involved in the process. To the extent that these costs are directly related to the development of the LLM, they may be capitalized under ASC 350-40. This generally includes compensation for those directly engaged in training the LLM, as well as compute and infrastructure costs that can be clearly identified as incremental to the model’s training.

Certain costs may be capitalizable under separate accounting guidance, such as hardware costs (servers, computers, GPUs) capitalized as property and equipment or data acquisition costs, which may be capitalizable as an intangible asset (see our separate article, Accounting for Data Acquisition Costs).

Judgment is often necessary for other costs incurred during the application development phase. For instance, ongoing hosting costs that would be incurred regardless of the development phase would be expensed as incurred. In contrast, excess compute costs specifically attributable to the LLM training process would be capitalized. Overhead expenses not directly tied to development activities, regardless of stage, are expensed as incurred.

Once the primary training is completed, some level of fine-tuning and testing is typically required prior to the model being ready for deployment; however, fine-tuning often continues post-deployment and throughout the model’s life. Under current GAAP, refinements made before the model meets its intended objectives and is ready for release can be classified as part of the application development stage. Once the model is considered ready to be deployed, any further testing or refinement expenses and ongoing maintenance should be expensed as incurred.

Capitalized costs are then amortized over the software’s useful life once it’s ready for use (see Determining the Useful Lives of AI Assets), with regular assessments for impairment to ensure the carrying value reflects its current utility (see Impairment Considerations for AI Assets).

After the initial release, entities will likely improve the application’s functionality through additional software development and fine-tuning. Consideration will need to be given to these costs and whether they are associated with an upgrade or enhancement and can be capitalized or related to ongoing maintenance and fine-tuning and expensed as incurred.

AI Technologies Developed to be Sold or Marketed Externally

If the generative AI application or LLM will be licensed externally, the development of such technology would be within the scope of ASC 985-20, and the costs of development cannot be capitalized until “technological feasibility” is established. Technological feasibility tends to be a higher bar than reaching the application development stage in internal-use software development project. Generally, it occurs toward the end of the development period when all high-risk development issues have been resolved. Costs incurred prior to technological feasibility are considered research and development costs and are expensed as incurred. As a result, minimal costs tend to be capitalized when software is developed to be marketed or sold externally, unless the costs are subject to other GAAP (such as hardware costs, or data acquisition costs).

Research and Development

Software developed to be used in research and development activities is subject to ASC 730, Research and Development (ASC 730). Under ASC 730, the costs incurred in developing software for use in research and development activities are considered research and development expenses and are charged as expenses as they are incurred. LLMs developed for research and development purposes therefore would be excluded from capitalization and discussion herein, and costs would be expensed as incurred.

Proposal to Modernize the Accounting for Software Costs

In June 2022, the FASB added to its agenda a project on modernizing the accounting for software costs, which was further updated in March 2024 to limit the project to making targeted improvements to ASC 350-40 while keeping the guidance and including ASC 985-20 within the scope of the project. The following changes are expected to be made to ASC 350-40:

• Removing all references to stages, which is expected to make the guidance more operable for entities that develop a software in a nonlinear manner.
• Retaining a probable-to-complete threshold for capitalization and including considerations related to significant development uncertainties and unresolved high-risk development issues the software project’s completion is unclear and whether the software will function as intended.

An exposure draft on the proposed guidance is expected to be issues sometime in the fourth quarter of 2024 for a 90-day comment period.

Conclusion

As organizations increasingly invest in generative AI technologies, aligning these innovative projects with existing accounting standards is essential. By applying frameworks like ASC 350-40 for internal-use software and ASC 985-20 for software intended for sale or marketing, entities can navigate the significant costs associated with AI development. This involves carefully mapping development activities to the appropriate accounting stages and making informed judgments about cost capitalization versus expensing. Understanding these accounting considerations enables financial leaders to accurately reflect the economic impact of AI investments on financial statements. Ultimately, bridging the gap between emerging AI technologies and traditional accounting models ensures compliance, supports strategic financial planning, and positions organizations to capitalize on the efficiencies and innovations that AI offers fully.

At CBIZ ARC, we specialize in providing top-tier technical accounting and financial consulting services to growth-oriented companies, including those leading in AI innovation. Our expert teams bring a unique combination of deep accounting knowledge and a clear understanding of emerging technologies, helping companies navigate complex financial, systems, and data management challenges. From addressing technical accounting issues and preparing for IPOs to unlocking deeper insights with AI-driven tools, we deliver customized solutions that ensure compliance, financial transparency, and performance optimization. Renowned for our commitment to seamless execution and responsive client service, CBIZ ARC provides the strategic support to drive your company’s sustained growth and long-term success.