If agents are building your app, who gets the W-2?

Accounting for Code That Writes Itself

Autonomous coding agents are no longer experimental. They can move a feature from design through deployment without human intervention. While the technology has outpaced expectations, the accounting framework for the cost of that work has not fully caught up. Under current U.S. GAAP, certain third-party software costs can be capitalized during the application development stage if they are a direct cost of creating software. In practice, developer tools have historically been treated as overhead because their costs were not directly traceable to capitalizable work. AI agents change that calculus: their usage is logged with enough detail to tie costs to specific development activities.

The Rules Already Fit: ASC 350-40’s Three Stages

The GAAP framework for internal-use software under ASC 350-40 divides development into three phases with distinct accounting treatments:

• Preliminary project stage: Research, planning, and feasibility costs are expensed as incurred.
• Application development stage: Costs for coding, integration, and testing before launch may qualify for capitalization, recorded as an asset and amortized over time.
• Post-implementation stage: Maintenance, bug fixes, and minor enhancements are expensed.

When a human developer codes during the application development stage, their salary is a capitalizable direct cost. The same logic applies to AI agents. If their usage is tracked and linked directly to coding, integration, or testing, those costs should qualify for capitalization under the same rule. The key enabler is granular logging: agent platforms can record exactly how much work occurred during a phase that GAAP considers capitalizable, which was nearly impossible with traditional overhead tools.

Why the Overhead Habit Persists

GAAP does not prohibit capitalizing AI-related development costs. The disconnect is an accounting convention left over from the pre-agent era. Tooling was historically spread across all three project phases and could not be allocated to a specific stage with any precision. As a result, finance teams expensed it all as overhead without revisiting the assumption.

AI usage data breaks that pattern. When an agent’s work is attributable to a single feature or project, its cost can be isolated and evaluated for capitalization on its own merits. That distinction converts what was a blanket expense into a direct dev cost — providing the accounting treatment already afforded to salaried engineers.

The Case for Treating Agents Like Staff Engineers

Capitalizing qualifying AI agent usage would align financial reporting with the underlying work being performed. Four benefits stand out:

• Economic substance: Both human engineers and AI agents produce long-lived software assets during development; identical work should not get different accounting treatment based on who is doing it.
• Transparency: Investors see a more accurate picture of asset creation when all qualifying development activity — human or machine — is capitalized.
• Comparability: Normalizing the treatment of labor and agent costs prevents distorted cost structures and makes project-level economics easier to compare.
• Consistency with existing GAAP: ASC 350-40 already contemplates capitalizable third-party software costs; agent usage meeting the criteria should fall within that rule.

What the Numbers Say

The balance-sheet effect is tangible. For a project generating $500,000 in AI-agent development costs that qualify under GAAP, the impact is significant relative to expensing everything immediately:

Objections and Answers

Agents are a service, not employees — can their costs still be capitalized?
Yes. GAAP’s test is the nature of the work, not the payee. Costs that meet the stage-of-work criteria qualify whether performed by a human or an autonomous agent.

Isn’t usage tracking too difficult?

Tracking was equally hard for human workers, and companies solved it when capitalization mattered. Agent platforms generally produce precise usage logs that are more reliable than self-reported time sheets.

Will capitalizing agent costs inflate assets?

Only if rules are misapplied. The capitalization guardrail requires work to clearly fall within the application development stage; that restriction holds regardless of whether a human or an agent performed it.

Closing the GAAP Gap

Treating all AI agent costs as overhead is a carryover from an era when tool usage could not be measured or allocated. That limitation no longer exists. Agent platforms provide the data needed to meet GAAP’s capitalization requirements for direct development costs.

The accounting treatment should reflect the work rather than the worker. Under ASC 350-40, AI agent capacity is a direct development cost when it performs capitalizable activity. The standard already permits it; finance teams now have the visibility needed to apply it.