Can you capitalize AI development? Internal-use software rules applied to models, prompts and agents

In brief

US GAAP has no AI-specific rule, so internal-use software guidance decides. What can be capitalized for agents, fine-tuning, data and tokens, with an example.

Steve Elliott

Steve Elliott · · 7 min read


On this page
  1. Does US GAAP have specific rules for AI development costs?
  2. When does capitalization start for an AI project?
  3. How do the rules apply to each type of AI cost?
  4. What about the tokens your developers spend on AI coding tools?
  5. An illustrative project: what gets capitalized?
  6. What evidence do the firms expect AI teams to keep?
  7. Sources

Yes, some of it. US GAAP has no separate rule for AI, so an AI agent or model your company builds for its own use is internal-use software: qualifying coding, prompt engineering and testing labor, and project-specific external costs, can be capitalized once the recognition threshold is met, while compute, data cleansing and staff training are generally expensed.

In our view, the judgment that moves the most money is when capitalization starts; the illustrative project below shows why. Under ASU 2025-06, a novel agent doesn't pass the threshold until coding and testing resolve its uncertainty. This post sets out what the FASB and the Big Four firms say, cost type by cost type.

Does US GAAP have specific rules for AI development costs?

No. The FASB considered it and declined. In the basis for conclusions to ASU 2025-06, issued in September 2025, the Board notes that "Subtopic 350-40 does not provide specific guidance for AI development costs." Some respondents asked it to clarify how the guidance applies to training an AI model and to data conversion. The Board decided not to, calling that beyond the scope of its targeted improvements.

The firms fill the gap by treating AI as ordinary software. Deloitte's 2024 Technology Spotlight on generative AI products calls it "essentially a form of software." KPMG's February 2026 Hot Topic on software data costs says AI development follows the same guidance as any other software. PwC's Q3 2026 Closing Statements, published September 17, 2026, says AI doesn't create a new accounting model. EY's Technical Line 2026-04, published September 24, 2026, advises splitting each AI arrangement into its parts, such as software development, software access, data, implementation and hardware, and assessing each under its own guidance. That is why the table below treats an AI project as several cost types, not one asset.

When does capitalization start for an AI project?

Under ASU 2025-06, capitalization begins once management with the relevant authority commits to funding and completion is probable. The threshold isn't met while significant development uncertainty exists: novel or unproven functions not yet resolved through coding and testing, or significant performance requirements still unidentified or being substantially revised.

EY's illustrations apply this to AI. A retailer building an internal customer-inquiry chatbot, a widely available kind of software, does not trigger the novelty factor. A company building an agent that runs contracts from drafting to finalization, where vendors describe such end-to-end agents as novel and unproven, does, so under the ASU its costs are expensed until coding and testing resolve the uncertainty. The FASB accepted a similar result for SaaS companies developing novel or unproven software: in BC57 it said most of those development costs may be expensed, and called that a reasonable outcome.

The ASU doesn't define a software project, so the unit of account is a judgment: Deloitte's Heads Up on the ASU works an AI example two ways, as one project and as two, with different start dates. Until a company adopts the ASU, the old stage model applies and only application development stage costs are capitalized, as EY notes.

How do the rules apply to each type of AI cost?

As of November 2026, the published guidance points this way. Where a row says "our reading," none of the sources addresses that item by name.

AI costWhere the guidance pointsSource
Per-token API fees for a model in productionUsage of a hosted service; expensed as incurred (our reading; Deloitte says ongoing usage costs aren't deferred)Deloitte 2024
Up-front payment for future model accessLikely a service contract; could be a prepaid assetDeloitte 2024
Implementing a hosted AI service you don't possessQualifying implementation costs capitalized, amortized over the hosting termASU 2018-15; EY 2026
Employee labor on prompt engineering, coding and testing an agent or RAG pipelineSoftware development cost, capitalizable after the threshold is metEY 2026; RAG by name is our reading
Token fees directly attributable to one projectExternal direct cost of that project, if the threshold is metEY 2026; PwC 2026
Company-wide AI subscription used across functionsTypically expensedEY 2026; PwC 2026
Fine-tuning or retrainingCapitalizable if it adds functionality; maintenance, expensed, if it keeps a model current (fine-tuning by name is our reading)Deloitte 2024; KPMG 2026
Data licensed with other usesSeparate intangible under ASC 350-30, not part of the software assetKPMG 2026; EY 2026
Cleansing, reconciling or creating data for the new systemData conversion; expensed, except software that converts or accesses old dataASU 2025-06
Cloud compute and storage, including to train a modelGenerally expensedEY 2026
Owned GPUs and serversProperty, plant and equipment under ASC 360EY 2026
Training people to use the systemExpensedASU 2025-06

Two rows carry the most judgment. The first is fine-tuning. Deloitte says data and training intended to keep an AI application current would most likely be maintenance, while training that creates new functionality might be an upgrade. KPMG says the development team's involvement will often be necessary to make that assessment. The second is infrastructure. EY says compute and storage fees are generally expensed as incurred. KPMG treats data-related infrastructure as generally indirect, unless it was obtained exclusively to house data whose only use is training a specific internal-use AI application.

Hosted model access sits under a separate test. Under ASC 350-40, a hosting arrangement includes a software license only if the customer has a contractual right to take possession of the software without significant penalty and could feasibly run it on its own hardware or through an unrelated host. Deloitte expects most entities to conclude that the foundation models they use are provided under a service contract.

What about the tokens your developers spend on AI coding tools?

Deloitte's 2026 Technology Spotlight on this question notes that attribution gets harder when AI tools serve several projects, productivity work or other business functions.

EY's view is that pricing models, tokens or subscription, "do not determine the accounting for the AI costs"; the activity does. In its illustrations, consumption-based token fees for an LLM service used to code a specific claims application are external direct costs of that project, subject to the company confirming the capitalization threshold is met. A flat company-wide subscription is expensed. EY prefers specific identification, such as project-level token tracking or usage logs, accepts a systematic allocation that reflects actual consumption, and rejects headcount-based or broad corporate allocations. If costs can't be tracked reliably, they are expensed.

PwC reaches a similar place. It says incremental costs for AI coding agents, API usage, tokens or dedicated AI development services may be capitalized when project-specific, measurable and incurred after the threshold is met, while broad enterprise-wide subscriptions should typically be expensed.

In our view, this makes the AI gateway an accounting control as much as a cost control. A dev-only key per project, logged by request, is the kind of evidence EY and PwC describe. A shared key across teams is not.

An illustrative project: what gets capitalized?

The numbers below are made up to show the mechanics. They don't come from any company, and an auditor could sort individual lines differently.

A company builds an internal claims-triage agent on a third-party model, with retrieval over policy documents. Funding is approved in January. The team treats the agent as novel, and coding and testing resolve the uncertainty at the end of April. It goes live in October and has run for three months by year end.

CostBefore uncertainty resolvedAfterTotalTreatment
Staff labor: coding, prompt engineering, integration, testing$300,000$600,000$900,000Capitalize $600,000
Contractor build fees$50,000$150,000$200,000Capitalize $150,000
Token fees on a dev-only project key$10,000$30,000$40,000Capitalize $30,000
Share of company-wide AI coding seats$25,000Expense
Cleansing and reformatting historical claims data$120,000Expense
Cloud compute and vector database during development$60,000Expense: EY treats compute and storage as generally expensed, unlike project-attributable token fees
Training claims staff$50,000Expense
Production token fees after go-live$90,000Expense

Total spend is $1,485,000. Capitalized: $600,000 + $150,000 + $30,000 = $780,000, about 53%. Expensed: $705,000.

Now change one judgment. If the agent were as routine as EY's chatbot and capitalization started at approval, the $360,000 incurred before April ($300,000 + $50,000 + $10,000) would also be capitalized, taking the asset to $1,140,000. That single date moves about 24% of the project's cost between the income statement and the balance sheet.

The asset is amortized straight-line over its useful life, which EY says requires judgment about obsolescence. On an assumed three-year life, $780,000 becomes $260,000 a year, about $65,000 a quarter from go-live, so year one carries about $65,000.

What evidence do the firms expect AI teams to keep?

PwC says AI "does raise the bar on documentation." We think four kinds of record matter most:

  1. The capitalization decision. Deloitte says entities must use judgment on what counts as a software project and document their reasoning. PwC says to document if and why AI costs qualify. For a novel agent, that includes when coding and testing resolved the uncertainty.
  2. Usage attributed by project. EY's preferred evidence is project-level token tracking or usage logs. Dev-only keys or workspaces per project produce it. This is AI cost allocation with an audit trail.
  3. Engineering records for each training run. KPMG expects the development team's input on whether retraining adds functionality. Record that call at the time.
  4. Support for useful life. EY says useful life requires judgment about obsolescence. Ask your auditors how replacing or retiring a model early would affect the asset.

EY says it does not believe entities should capitalize costs without appropriate processes and controls. Book treatment also differs from tax, which our post on book capitalization vs Section 174 covers. For the labor-versus-token split behind these lines, see AI initiative costs: tokens vs labor.

In our view, capitalization depends on facts finance rarely holds, such as which project a token served and when a prototype became a build, which is the case for keeping AI spend, projects and funding in one operating model.

This post summarizes published guidance and is not accounting advice. Confirm the treatment of your own projects with your auditors and accounting advisers.

Sources

Frequently asked questions

Can you capitalize the cost of calling an AI model's API?

Usually not in production, where, in our reading, per-token fees are operating costs of a hosted service. EY and PwC describe an exception: token fees directly attributable to coding and testing a specific software project, incurred after the capitalization threshold is met and tracked reliably, can be capitalized as external direct costs of that project. A company-wide AI subscription is typically expensed.

Can fine-tuning a model be capitalized?

Sometimes. Deloitte and KPMG treat training or retraining, which includes fine-tuning, as a possible upgrade when it adds new functionality, capitalizable if the internal-use software criteria are met, and as maintenance, expensed, when it only keeps a model current. Data bought for training that has other uses is a separate intangible asset. KPMG says the development team's involvement will often be necessary.

Does ASU 2025-06 make it harder to capitalize AI agents?

For novel agents, it can. Under ASU 2025-06, capitalization can't begin while significant development uncertainty exists, such as unproven functions not yet resolved through coding and testing. EY's illustrations treat an end-to-end contract agent as novel and a standard chatbot as not. The FASB said most costs of developing novel SaaS software may be expensed. The ASU applies to annual periods beginning after December 15, 2027.

Is data preparation for an AI model capitalizable?

Mostly not. Under ASC 350-40, data conversion, including purging, cleansing, reconciling and creating new data, is expensed, though software built to access or convert old data can be capitalized. Licensed data with other uses can be a separate intangible asset under ASC 350-30. KPMG notes that internally generated data costs are generally expensed. The FASB chose not to add AI-specific data guidance.

Free playbook

Run the enterprise from any AI.

Read The Headless Enterprise
Steve Elliott

Steve Elliott

Published November 22, 2026

More on AI Economics

Tokens, inference cost, cost per task.