Should AI Agent Development Costs Be Capitalized or Expensed? What It Does to Your ROI Case

Quick answer

Under both US GAAP and IFRS, whether AI agent development costs are capitalized or expensed generally follows the same rules used for internally developed software: costs incurred during the preliminary project and post-implementation stages are expensed as incurred, while costs incurred during the application development stage, once technical feasibility is established, can typically be capitalized. Licensing fees for a third-party AI platform are usually expensed unless the arrangement includes a significant internal-use software component. The accounting treatment changes when costs hit the P&L, which materially affects the near-term ROI and margin picture shown to the board, even though it doesn't change the underlying cash spent.

This is an accounting question layered under a familiar ROI question

Our posts on calculating AI agent ROI before greenlighting a project and the cost of not adopting an AI agent both work from the assumption that "cost" is a known, fixed number you're weighing against a benefit. In practice, the same dollar of spend can appear very differently in the P&L depending on whether it's capitalized or expensed, and that treatment decision has a real, sometimes decisive, effect on how the ROI case looks to a board that's watching quarterly margins.

The general framework, following existing internal-use software guidance

Most accounting standards, including US GAAP's guidance on internal-use software (ASC 350-40) and the equivalent IFRS treatment, break a development project into three stages:

1. Preliminary project stage: evaluating alternatives, deciding on a vendor or build approach. Expensed as incurred.
2. Application development stage: once the project has cleared feasibility and management has committed to funding it, coding, configuration, and integration costs can generally be capitalized.
3. Post-implementation stage: training, maintenance, and minor upgrades after go-live. Expensed as incurred.

For a custom-built or heavily customized AI agent, this maps reasonably cleanly: research and vendor evaluation gets expensed, the build phase gets capitalized, ongoing maintenance and prompt tuning after launch gets expensed.

Where AI agents complicate the traditional framework

Licensing a third-party platform. If you're primarily paying for access to a vendor's AI agent platform rather than building custom software, that's usually a straightforward operating expense, similar to any SaaS subscription, unless your contract includes a significant internal-use software license component that meets capitalization criteria on its own.

Ongoing model and prompt iteration. Unlike traditional software, a production AI agent often undergoes continuous prompt refinement, retrieval tuning, and model swaps well past what would traditionally be considered "post-implementation." Whether meaningful post-launch changes qualify as a new capitalizable development effort or routine maintenance is a judgment call that should be made with your accounting team and applied consistently, not decided ad hoc each time a change ships.

Compute and inference costs. Ongoing inference cost to run the agent in production is an operating expense, not a capitalizable development cost, regardless of how the initial build was treated. This distinction matters directly for the cost modeling covered in our post on AI agent inference cost at scale, since that ongoing cost hits the P&L immediately no matter how the upfront build was accounted for.

How this changes the ROI case in practice

Capitalizing eligible development costs spreads the expense over the asset's useful life instead of hitting the P&L in the period the work happened, which improves near-term reported margins and can make an ROI case look more favorable in the first year specifically. This is a real accounting effect, not a way to manufacture ROI that isn't there: the total cash spent is identical either way, and finance and the board should understand the ROI case is about total value versus total cost, with the capitalization treatment only affecting when the expense lands on the P&L, not whether the investment was worthwhile.

What to bring to the board conversation

When presenting the ROI case, be explicit about which costs were capitalized versus expensed and over what useful life any capitalized costs are being amortized. A board member comparing this year's AI initiative margin impact to a prior year's initiative with a different accounting treatment will draw the wrong conclusion if the treatment difference isn't called out plainly. This pairs directly with the board-level metrics dashboard covered in our post on AI agent metrics for the board and CEO, where the accounting treatment should be a footnote on any cost trend line, not an invisible assumption.

FAQ

Does the capitalize-versus-expense decision affect the actual cash we spend? No. It only affects the timing and P&L presentation of that spend, not the underlying cash outflow, which is identical either way.

Who decides whether AI agent development costs should be capitalized? This should be a decision made jointly by finance and accounting, applying your existing internal-use software capitalization policy consistently, not a decision engineering or the project team makes unilaterally based on how they'd like the ROI case to look.

Can ongoing prompt engineering and fine-tuning after launch be capitalized? Generally this falls into the post-implementation stage and should be expensed, similar to routine software maintenance, unless it constitutes a substantial new development effort that itself clears the feasibility bar, which is a judgment call to make with your accounting team rather than a default assumption either way.

Does capitalizing AI development costs create risk if the project is later abandoned? Yes. If a capitalized AI agent project is abandoned or becomes obsolete, the remaining capitalized balance typically needs to be written off immediately, which creates a one-time expense hit at exactly the point the project is already viewed as a failure. This is worth flagging to the board as a risk of capitalization, not just a benefit.