Contents
Key takeaways Why is AI shrinking your capitalized software balance? Can you capitalize AI spend as ASC 350-40 internal-use software? When should you capitalize software under ASU 2025-06? What is capitalizing AI spend worth in EBITDA? What will your auditor ask for? How do you track AI spend for capitalization? Should you adopt ASU 2025-06 early? What should finance do in the next 90 days? What does AI spend capitalization look like in practice? Frequently asked questions

Quick Answer

Yes, in part. Under ASC 350-40, AI software costs can be capitalized when they're direct costs of building a specific, authorized internal-use software project. Deloitte and EY both say a shared AI bill split by headcount doesn't qualify. Without usage records tied to a project and a development activity, the whole bill is expensed and your AI ROI is understated.

CloudZero, the AI ROI company, tracks cloud and AI spend for finance and engineering teams, and it has run into this problem in its own books. The company has run AI-first across the business since early 2026, and Dan Carducci, its VP of Finance and RevOps, wrote on September 23, 2026 that AI can now reach every line of the P&L, R&D included. For a while, his forecast for the AI line was the run rate plus a cushion.

That changed once his team could see the spend by the work it paid for. Emily Allen, CloudZero’s director of finance and accounting, has since capitalized the AI spend that went into new feature development. Carducci puts the condition plainly: “You cannot capitalize what you cannot attribute.“

Next quarter’s audit committee meeting has a version of this problem in it. Engineering shipped more than it did a year ago. The AI tools bill is up sharply. Capitalized R&D is down, operating expense is up, and someone on the committee wants to know why EBITDA missed when the roadmap didn’t.

The explanation is in how software capitalization works. Finance counts engineers’ hours on qualifying projects and capitalizes that labor. Now a coding agent does a share of the work, and an agent doesn’t fill in a timesheet. Its cost arrives as a vendor invoice with no project on it, so it lands in opex.

Key takeaways

  • Deloitte published seven worked examples on capitalizing AI spend, dated August 3, 2026. EY followed with a Technical Line on September 24, 2026.
  • ASU 2025-06 removes project stages from ASC 350-40. It is mandatory for annual periods beginning after December 15, 2027, and early adoption is allowed.
  • AutoNation adopted early on January 1, 2026, and recorded a $10.5 million after-tax reduction to retained earnings.
  • Company-wide AI seat subscriptions are overhead. Production inference is cost of revenue. Neither is capitalized.
  • Capitalizing lifts EBITDA in the period and leaves cash unchanged. The cost returns as amortization.

Why is AI shrinking your capitalized software balance?

AI is shrinking capitalized software balances because the cost of building software is moving from payroll, which finance knows how to capitalize, to AI usage, which is much harder to trace. Capitalized labor rests on hours logged to qualifying projects. When an agent writes the code, engineers may log fewer of those hours, and nothing takes their place on the balance sheet.

The company that ships fastest with AI can end up reporting the weakest capitalized R&D, unless its records change with its tools.

Scott Castle, CloudZero’s chief product officer, makes the same point about a large part of the AI bill. It isn’t cost of goods sold, he argues. It’s coding agents producing the software, and no bill maps that spend to what it produced. As he puts it: “Cost-to-serve isn’t the problem. Cost-to-produce is.“

Finance has had years of practice with cost-to-serve. Cost per customer and cost per feature for a production workload are familiar numbers. Producing software is harder to pin down. One feature may be built by two different coding agents across dozens of sessions, and the provider’s invoice shows tokens by model and by day. The project is nowhere on it.

Counting tokens doesn’t fix that. Amazon reached a similar view about measuring its own engineers in a memo it published on October 27, 2026, calling token usage, adoption rates, and code-change counts “deeply flawed” ways to measure a team. CloudZero covered the memo in Amazon told its engineers to stop tokenmaxxing. A token count tells you how much AI ran. It can’t tell a new feature from a bug fix, and the accounting turns on that difference.

The AI line is now large enough to draw scrutiny. CloudZero’s AI Economics Pulse for September 2026 found that the median company in its 430-company customer panel spent 2.66% of its cloud bill on AI in August, while the 75th percentile reached 11.12%, up from 9.63% in July, a finding covered in why so few companies can explain the AI bill.

Can you capitalize AI spend as ASC 350-40 internal-use software?

Yes, when the charge is a direct cost of developing a specific internal-use software project and the work is development. That is the conclusion of Deloitte’s August 3, 2026 guidance on AI in software development and of EY’s Technical Line, published on September 24, 2026.

ASC 350-40, the US GAAP guidance on accounting for software a company develops for internal use, has always told companies to capitalize the external direct costs of materials and services consumed in development. That rule predates AI.

An AI service gets the treatment a contractor would. Show that the invoice belongs to the project and that the work built something new. Deloitte applies the same test to token charges, subscriptions, and agent fees, and adds that the pricing model does not settle it. A metered token charge is not automatically capitalizable, and a fixed fee is not automatically excluded.

Deloitte then works through seven examples. Most of them look like contracts your procurement team has already signed.

Deloitte exampleThe situationTreatment
1Token charges from a coding agent used only by one project’s team and tracked by project IDCapitalize
2Enterprise per-seat subscription used company-wide for many purposesExpense as overhead
3Six-month fixed subscription bought solely for one project, with access restricted to itCapitalize
4Mixed contract with project-level token charges plus a fixed shared-capacity chargeCapitalize the project tokens. Expense the fixed charge
5One session that codes for the project, then works a production incidentCapitalize only the project tokens. If the split cannot be supported, expense the whole session
6$1 million minimum commitment for up to 5 billion tokens, with about 600 million expected and shared useDo not capitalize a per-token share of the fixed payment
7$1 million of prepaid credits expected to be fully consumedCapitalize credits used on the project, expense maintenance and research credits, hold unused credits as prepaid

Example 2 will sting for anyone on a company-wide seat deal, such as a flat Claude Code plan for the whole engineering org. The pricing structure of each AI coding tool now has an accounting consequence as well as a budget one.

Allocation is ruled out too. Deloitte says spreading a shared tool’s cost across projects by headcount, labor hours, or budget does not make it a direct project cost. EY’s term for the accepted method is specific identification: usage logs that attribute consumption to one software project. If you cannot track usage to a project, you expense all of it.

Coding tools are not the whole AI build budget. Crowe’s August 26, 2026 analysis says fine-tuning compute can be capitalized when usage logs and run manifests tie it to a qualifying project. Data labeling and cleaning in a research program is expensed. Third-party API costs used to serve customers are generally cost of revenue.

When should you capitalize software under ASU 2025-06?

ASU 2025-06 changes when to capitalize software. Under the new standard, capitalization starts once management has authorized the project and committed funding, and it is probable the project will be completed and used as intended. That test is not in force yet. Until a company adopts, the old project stages still govern when capitalization begins, so a calendar-year company is working to the stage model through 2027 and to this one from 2028.

FASB issued the update on September 18, 2025, and it deletes the old project stages. It takes effect for fiscal years that start after December 15, 2027, and companies can adopt early.

Old modelASU 2025-06
What starts the clockThe project leaves its preliminary stageFunding is authorized and committed, and completion is probable
Fit with how teams buildAssumes sequential stagesWorks for iterative delivery
Test for uncertaintyNone spelled outSignificant development uncertainty, with two defined factors
Which costs are eligibleExternal direct costs and directly associated payrollUnchanged

The change suits how engineering teams work. Stages assumed a waterfall plan, and agile delivery did not fit them. A team shipping every two weeks could not say when “preliminary” ended, so finance often drew the line by convention. The new test asks a question a product organization can answer: was it funded, and will it ship?

AI work gets a harder version of that question. Capitalization cannot begin while there is significant development uncertainty. FASB defines it two ways: the software has novel, unique, or unproven functions that coding and testing have not resolved, or its significant performance requirements have not been identified or keep being substantially revised.

Much AI work can fit one or both. Crowe’s view is that a program built on novel AI functionality may expense most of its costs, much as research would. Compare two illustrative projects. Document summarization through a commercial model API, with acceptance criteria fixed at kickoff, has a path to capitalized software. An autonomous agent whose accuracy target has moved three times this quarter probably does not yet.

AutoNation has already shown what that test can do to a balance sheet. The car retailer, with 244 stores as of March 31, 2026, adopted early on January 1, 2026. Its first-quarter 10-Q reports that in-process software projects failing the new requirements were derecognized, meaning removed from the balance sheet, a $10.5 million after-tax reduction to retained earnings.

What is capitalizing AI spend worth in EBITDA?

Capitalizing AI spend lifts EBITDA by the full amount you capitalize, in the period you spend it. An enterprise that capitalizes $7.2 million of a $12 million AI coding bill reports $7.2 million more EBITDA that year, with no change to cash. Capitalized software costs move off the income statement and onto the balance sheet, and their amortization sits below the EBITDA line. Spend you cannot trace is expensed whatever your policy says.

It is also missing from the AI ROI you report, because the build cost never lands on the thing it built.

Take an illustrative enterprise with $12 million a year in usage-based AI coding charges. Its records tie 60% of that usage to development work on authorized projects. It capitalizes $7.2 million. With a three-year useful life and software that goes live at mid-year, first-year amortization is $1.2 million.

Illustrative year oneWith project-level recordsWithout them
AI coding spend$12,000,000$12,000,000
Capitalized$7,200,000$0
Expensed$4,800,000$12,000,000
Amortization$1,200,000$0
EBITDA vs. no records+$7,200,000Baseline
Operating income vs. no records+$6,000,000Baseline

Who feels the difference depends on ownership. For a public company it flows through to earnings per share. For a private-equity-backed company it changes covenant EBITDA, the figure its debt agreements are tested against.

CloudZero’s capitalization guidance for finance teams carries a caution: analysts and diligence teams compare capitalization policies and reverse amounts they judge aggressive.

No benchmark says what share of an AI bill should qualify. It depends on how much of the work is new development, how the contracts are priced, and how much usage you can trace.

What is an example of a capitalized software journal entry?

For the same illustrative company, the year’s journal entries record the invoices first and the amortization once the software is in service.

AccountDebitCredit
Capitalized software in process$7,200,000
AI services expense$4,800,000
Accounts payable$12,000,000
Amortization expense, first year$1,200,000
Accumulated amortization$1,200,000

Software amortization starts when the software is ready for its intended use, which is typically when the code goes into production.

The gain is timing. A board should hear it described that way. A board that hears it as profit will ask where the profit went in year two. Cash out the door is the same with or without capitalization. The cost comes back as amortization, and an abandoned project becomes an impairment. A feature customers will pay for over three years should not be charged entirely to the quarter it was built in.

Amortization also has to land somewhere, and for product software that is often cost of revenue. Expensify’s 2025 annual report shows it. The company held $36.4 million of gross capitalized software development costs at year-end, $13.6 million net, amortizes over three years, and records that amortization in cost of revenue.

AI products have less margin to absorb it. ICONIQ’s July 2026 State of AI report, a survey of more than 300 executives at software companies building AI products, put average gross margin on AI products at 45% in 2025, with 53% projected for 2026 and 59% for 2027. Those margins are improving, but they still sit well below the 70% to 80% a traditional SaaS model assumes, and amortization lands on top of inference. Any forecast of AI gross margin needs the amortization line next to inference.

CloudZero’s free demo measures how much of your AI and cloud spend already maps to a product or project.

What will your auditor ask for?

Your auditor will ask for records that tie each capitalized AI charge to a specific project and a development activity, made when the work happened.

Deloitte’s list of acceptable evidence starts with “AI-platform reporting or billing details, project-management systems, development tickets.” A spreadsheet assembled at close from engineers’ recollections is the weak version of all three.

What the auditor asksWhere the answer livesWhere it tends to break
Was the project authorized and funded?Budget approval and roadmap sign-offApproval covers a program, not the project
Was the hard part proven?Design decisions and test resultsNobody wrote the assessment down
Who used the AI service, on what?Usage logs joined to a ticket and repositoryShared API keys, no ticket
Was it development or support?Ticket type and the work performedOne session does both
Does the cost match the invoice?Provider billThe bill is one line with no project

For a public company, these judgments sit inside internal control over financial reporting. Someone has to own the start date, the uncertainty assessment, and the split between development and support, and each control has to leave evidence. Under ASU 2025-06, capitalized internal-use software also follows the disclosure requirements for property, plant, and equipment.

How do you track AI spend for capitalization?

Look for a tool that records AI usage against a project as the work happens and ties it to the invoice. Engineering time trackers capture labor hours. Provider consoles show tokens by model and API key. Neither connects an AI charge to a ticket, an activity, and an invoice line, and that chain is what an auditor follows.

CapabilityWhy it matters for capitalization
Attribution by ticket and repositoryIt is specific identification. Headcount allocation does not qualify
Mixed sessions split by activityDevelopment is capitalized. Incident and support work is expensed
Cost taken from the provider invoiceThe capitalized amount is what you paid
Records created when the work happensContemporaneous evidence holds up better than a spreadsheet rebuilt at close
Policy written as rulesThe same treatment applies every period, and the rule shows on each line
Coverage across providersAI build spend is spread across OpenAI, Anthropic, AWS, Azure, and others
Export for audit supportThe auditor gets the sessions behind each number

CloudZero built its capitalization workflow to that list. Finance writes its policy as rules. Each AI coding session is then recorded with its user, tool and model, ticket, repository, class of work, and invoice cost as the work happens. The same engine is what lets finance allocate AI spend to customers and features.

The AI capitalization datasheet shows a sample close view. Three projects under development are capitalized at $184,210. An incident fix and a batch of support investigations are expensed at $62,880. Each line opens to the sessions behind it. CloudZero’s capitalization workflow for finance covers the setup, and the same records classify every AI task into the business activity it belongs to, so the capitalization view and the department view come from one source.

Should you adopt ASU 2025-06 early?

Adopt early if you can already trace project-level spend, and wait if you cannot. Adopting early drops stage tracking sooner, but it forces an immediate retest of every project in flight. That is the call to put to your controller and auditor. AutoNation chose the modified prospective approach, which derecognizes in-process projects that fail the new test on the adoption date. That retest produced its write-down.

Advance Auto Parts is weighing the same move. It said in its annual report for the year ended January 3, 2026 that it was evaluating the standard, including possible early adoption in fiscal 2026. Neither company sells software. Any business that builds its own systems is in scope.

The standard offers three transition methods, and all three run on project-level history, so the choice is easier for a company that can already trace its spend.

Transition methodWhat it applies toWhat it can cost you
ProspectiveNew costs only; existing capitalized balances stay on the books and are tested for impairmentTwo sets of rules running in the same year
Modified prospectiveIn-process projects that fail the new test are derecognized on the adoption date, through retained earningsA write-down, as AutoNation recorded
RetrospectiveThe same derecognition, applied as of the beginning of the earliest period presentedFull project history for every period presented

What should finance do in the next 90 days?

Start by finding out how much of your AI build spend is traceable today. A calendar-year company must adopt by January 1, 2028, and the evidence has to exist from the day the money is spent, so 2027 planning is the moment to set this up.

  1. Sort the AI bill by contract type. Metered usage, flat seats, committed minimums, and prepaid credits are treated differently, as Deloitte’s examples show. Managing AI spend starts with knowing what you signed.
  2. Write the capitalization policy. Define what a project is, what counts as authorization, who signs the start date, and how the uncertainty assessment is recorded. FASB left “project” undefined, so your capitalization policy has to.
  3. Tie usage to tickets. Every AI session or API key should carry a project, a ticket, and a repository.
  4. Dry-run last quarter. Apply the draft policy to spend you’ve already incurred and see what share would have qualified.
  5. Take it to the audit committee. Bring the policy, the dry-run number, and a view on adoption timing.

If the dry run shows 20% of spend is traceable, tracking is the first project, and it pays twice: the same records that support capitalization are the ones that show what each AI dollar produced. If it shows 80%, the conversation moves on to useful life and impairment.

What does AI spend capitalization look like in practice?

A finance team that gets this right can answer, in minutes rather than at close, which project each AI charge paid for. CloudZero’s own finance team is the nearest example. Emily Allen, its director of finance and accounting, first went into the platform to find out why AI spend had roughly doubled in the P&L.

She traced most of the increase to heavy use of a premium model by one account development rep. Moving part of that usage to a cheaper model was worth $13,000 to $30,000 a month. The analysis took her 10 minutes.

Allen has since built a view of AI spend against revenue, gross margin, and each department’s salary base, alongside the capitalization work. Dan Carducci, CloudZero’s VP of Finance and RevOps, says AI now works like every other cost category in his business: someone brings a finding and a recommendation, and he approves it.

Capitalization is one use of those records. They also answer the board’s question about AI ROI and set next year’s budget.

Scott Castle, the company’s chief product officer, notes that the allocation engine underneath predates AI coding and has supported financial audits at companies like Coinbase, Nubank, Shutterstock, and Duolingo. It now does the same job for AI spend.

CloudZero’s 30-minute capitalization demo records your live AI coding sessions by ticket and work type, applies a sample policy, and shows the close view with the sessions behind every line.

Frequently asked questions

Note: This article describes accounting guidance in general terms and is not accounting advice. Whether a cost is capitalized is a decision for your controller and auditor under your own accounting policy.