Can you capitalize AI software costs under US GAAP?
Every quarter our Los Angeles office fields the same question from controllers at software, SaaS, and life-sciences clients. The company has spent heavily on AI tools — coding agents, API tokens, fine-tuning time, model licenses — and management wants to know how much can be capitalized. Behind the question is a reasonable fear that the auditor will sweep it all into operating expense, and the P&L will show the cost without any of the offsetting asset. The answer under US GAAP has not changed because the technology has. AI software costs get the same software cost guidance that has been on the books for a decade — but the documentation bar is higher, and the way AI tools are consumed makes the application harder than it looks. This post walks through the current rules, the two common fact patterns, and the records an auditor will want to see.
What US GAAP standard governs AI software costs? 📘
US GAAP has no AI-specific standard. AI software costs are evaluated under the existing software cost guidance in ASC 350-40 or ASC 985-20, depending on how the software is used.
The FASB sets US GAAP through the FASB Accounting Standards Codification. The two codification topics that control software costs are:
- ASC 350-40 (Internal-Use Software). Applies to software developed or obtained for internal use — enterprise software, SaaS platforms the company itself operates, internal data and reporting systems. This is the pathway that most AI capitalization questions fall under.
- ASC 985-20 (Software to Be Sold, Leased, or Marketed). Applies to software that will be sold, leased, or otherwise marketed. Different cost model: capitalization begins only once technological feasibility is established.
- ASC 730 (Research and Development). Applies to research and development activity. Costs that are R&D in substance are expensed as incurred under US GAAP.
The AI question rarely sits inside a single topic. A company training an internal AI model used in its own product delivery is in ASC 350-40 territory for the system itself and potentially in ASC 985-20 for parts of the product. A company researching new models before committing to a specific project is in ASC 730 territory and is expensing. The classification step comes first, and it decides the rest of the accounting.
What are the ASC 350-40 project stages, and when can capitalization begin? 🔁
ASC 350-40 defines three stages — preliminary project, application development, and post-implementation. Capitalization is permitted only during the application development stage, after management commits to the project and it is probable the software will be completed and used.
The stage framework is the single most important control in internal-use software accounting, and it is where AI projects most often go wrong. The rules on this point are mechanical and have not changed with the technology.
| Stage | Typical activities | Accounting |
|---|---|---|
| Preliminary project | Feasibility studies, vendor evaluation, high-level requirements, pilot testing to decide whether to proceed. For AI projects: proof-of-concept experiments, early benchmarking of different model families, informal prompt tuning before any commitment. | Expensed as incurred. |
| Application development | Begins when management authorizes the project, funds it, and it is probable the software will be completed and used for its intended function. Direct costs during this stage — including external services, direct employee compensation, and software licenses — may be capitalized. | Capitalized if directly attributable to the project. |
| Post-implementation / operation | Software is substantially complete and ready for its intended use. Training, routine maintenance, bug fixes, and ongoing operation of the system. | Expensed as incurred (with narrow exceptions for upgrades and enhancements that add functionality). |
A fourth question lurks in every AI engagement: how to allocate shared costs. An AI coding agent used by five developers, three of whom are on a qualifying project and two of whom are doing general support, is partially capitalizable only to the extent the usage can be tied to the qualifying project. Usage logs, API call records, and seat-level billing matter here. An unsupported pro-rata assumption is the kind of adjustment auditors propose at year-end.
Which AI costs are capitalizable, and which are expensed? 💡
Project-specific incremental AI costs tied to a qualifying internal-use software project during its application development stage may be capitalized. Broad, enterprise AI subscriptions used across general business functions are expensed.
| AI cost | Typical treatment | Why |
|---|---|---|
| Enterprise AI chat tool used company-wide | Expensed as incurred | Not project-specific; crosses the preliminary and operation stages for many business functions at once. |
| AI coding agent used by a development team on a specific qualifying project | Capitalizable to the extent usage is directly tied to the project during the application development stage | Project-specific incremental cost; requires usage logs or seat-level attribution. |
| API tokens for a model embedded in a specific internal application being built | Capitalizable during the application development stage | Direct cost of the asset being built; the tokens are consumed to produce the software. |
| Fine-tuning a foundation model for a specific internal use case | Generally capitalizable during the application development stage | Directly attributable to a specific internal-use software project. |
| Pre-commitment research into which model to use | Expensed as incurred | Preliminary project stage under ASC 350-40; also R&D under ASC 730. |
| General AI training for employees | Expensed as incurred | Training costs for internal-use software are expensed in all stages. |
| Post-launch monitoring and retraining of a deployed AI feature | Generally expensed | Post-implementation / operation stage; capitalizable only if the retraining is an enhancement adding new functionality. |
How do you document capitalized AI software costs for an audit? 📝
Capitalized AI costs need written project authorization, direct-cost evidence tied to the project, allocation methodology for shared tools, and an impairment review at year-end.
- Project authorization memo. A written record — steering-committee minutes, approved funding, or an equivalent — showing when the preliminary project stage ended and the application development stage began. The date is important; capitalization cannot begin earlier than this.
- Feasibility documentation. Enough evidence that it is probable the software will be completed and used for its intended function. Specifications, user stories, or an architecture document all qualify.
- Direct-cost evidence. Contracts, invoices, and payment records for AI tools (coding agents, API usage, token purchases, model licenses, dedicated development services) tied to the specific project.
- Allocation methodology. For shared AI tools, a documented method — usage logs, seat counts tied to project assignments, metered API calls, or a defensible pro-rata allocation. The method should be applied consistently across projects.
- Internal time records. Developer and engineer time tracked by project, with the portion assigned to qualifying work capitalized at the fully-loaded cost.
- Impairment review. At each fiscal year-end, confirm that the capitalized amount is still recoverable. A project that has been abandoned, suspended indefinitely, or fundamentally rescoped triggers an impairment analysis under ASC 350-40.
From our practice: the documentation gap is the real risk
In our practice, the question that decides whether capitalized AI costs survive an audit is almost never the characterization. It is the records. A controller who can show a dated authorization memo, a project code on every AI invoice, and a usage log that ties token consumption to specific developers is in a strong position. A controller who comes to the audit with a spreadsheet of “AI spend, allocated 70% to development” is not. Build the record now; it is cheaper than defending it later.
What are the common AI accounting pitfalls? ⚠️
Capitalizing preliminary-stage experimentation, missing allocation evidence for shared tools, treating general AI subscriptions as project costs, and skipping the year-end impairment review are the four most common issues.
- Pulling capitalization forward. Starting to capitalize before management has formally authorized the project. The dated memo is the control.
- Treating enterprise subscriptions as project costs. A ChatGPT or Copilot subscription used by the whole engineering team is not project-specific. A portion may be allocable, but only with records to back the allocation.
- Confusing R&D with application development. ASC 730 research and development is expensed. Early AI model experiments typically fall here even if the eventual product reaches the application development stage.
- Omitting the operation-stage line. Costs after the software is ready for its intended use go back to expense, even if the project team is still active.
- Skipping impairment at year-end. Abandoned pilots and rescoped projects require an impairment analysis. Writing off capitalized AI costs the following year without an impairment memo is an audit finding.
US GAAP, not SEC rulemaking, controls AI capitalization
The SEC’s proposed rules on semiannual reporting and filer status modernization attract attention, but they do not change ASC 350-40 or ASC 985-20. US GAAP is set by the FASB, not the SEC. Document AI capitalization decisions under current FASB guidance, and track FASB and SEC projects separately.
How does AI capitalization fit into year-end close? 📅
Capitalized AI costs land on the balance sheet as internal-use software (or, for ASC 985-20 projects, software to be sold), amortized over the estimated useful life beginning when the software is ready for its intended use.
At year-end, the capitalized balance is reviewed for both completeness and recoverability. Completeness: were all qualifying costs captured, including developer time at fully-loaded rates and direct AI tool consumption? Recoverability: do current facts still support the assumption that the software will generate future benefits? If a project has been shelved, the capitalized asset must be written down.
Amortization begins when the software is ready for its intended use, not when the project started. Useful life depends on the software’s expected period of benefit; three to five years is a common range for internal-use software, though AI-powered systems that depend on a rapidly evolving model base may support a shorter life. The useful-life estimate itself needs documentation.
Summary: AI software costs under US GAAP
- No AI-specific standard — ASC 350-40 (internal-use software), ASC 985-20 (software to be sold), and ASC 730 (R&D) still control.
- Capitalization is limited to the application development stage of a specific project, after management authorization and probable completion.
- Project-specific AI costs (API tokens, dedicated coding agents, fine-tuning for an application) may qualify; broad enterprise subscriptions do not.
- Documentation decides the outcome: authorization memo, invoices tied to project codes, allocation evidence for shared tools, year-end impairment review.
- SEC proposed rules on semiannual reporting and filer status do not change the FASB guidance; track them separately.
Frequently asked questions about capitalizing AI software costs ❓
Q. Can AI software costs be capitalized under US GAAP?
Sometimes. US GAAP has no AI-specific standard. The existing software cost guidance in FASB ASC 350-40 (internal-use software) and ASC 985-20 (software to be sold) still controls. If the AI spend is directly tied to building qualifying software and is incurred after the project reaches the capitalization threshold, it may be capitalized. If it is a general enterprise AI subscription used across many business functions, it is generally expensed as incurred.
Q. What are the three project stages under ASC 350-40?
ASC 350-40 divides internal-use software projects into the preliminary project stage, the application development stage, and the post-implementation/operation stage. Costs incurred in the preliminary stage are expensed. Only direct costs incurred during the application development stage, once management has committed to the project and it is probable the software will be completed and used for its intended function, qualify for capitalization. Costs incurred after the software is substantially complete and ready for its intended use are again expensed.
Q. How does the SEC-rulemaking landscape affect AI capitalization decisions?
The SEC’s proposed rules on semiannual reporting and filer status modernization do not change ASC 350-40 or ASC 985-20. US GAAP is set by FASB, not the SEC. But comment letters on the SEC proposals, combined with ongoing FASB projects on software costs, mean that both the reporting frequency and the capitalization framework could evolve. Companies should document AI capitalization decisions on the current rules and track future FASB and SEC activity separately.
Q. What documentation does an auditor expect for capitalized AI costs?
Written project authorization or steering-committee minutes showing when the preliminary stage ended, a technical feasibility assessment, contracts and invoices for AI tools (coding agents, API usage, tokens, model licenses) tied to specific projects, usage logs or metering data allocating consumption, time records for internal developers assigned to the project, and a clear impairment review at the end of the fiscal year. Weak or missing records are the most common reason capitalized AI costs are proposed as audit adjustments.
Q. Does a general-purpose AI subscription qualify for capitalization?
Generally no. Broad, enterprise-wide subscriptions to AI chat tools or coding assistants used across general business functions — marketing, HR, finance, support — do not meet the project-specific standard in ASC 350-40 and are expensed as incurred. The same tool can be partly capitalized only when a measurable, project-specific portion of the usage is tied to qualifying software development, with records to support the allocation.
Q. Are AI research costs the same as AI software development costs?
No. Costs that are research or development under FASB ASC 730 — basic investigation, feasibility work on new models, costs incurred before the application development stage begins — are expensed as incurred. Costs that cross into the application development stage of a specific internal-use software project move under ASC 350-40, and some may be capitalizable. The characterization decides the accounting.
This article is general information, not accounting advice for your situation. The application of ASC 350-40 and ASC 985-20 depends on specific project facts, contracts, and usage patterns. If your company is deciding how to account for AI tool costs, contact SW Accounting & Consulting Corp for a confidential review.
