Capitalizing AI and Machine Learning Development Costs

AI and machine-learning development doesn’t fit neatly into one accounting box. Whether you capitalize the cost depends on what the AI is for and how it’s delivered — and ASU 2025-06‘s new uncertainty test hits novel AI work especially hard.

Which framework applies

Coding vs. exploration

Within development, there’s a line that matters: coding an algorithm into software — and the compute, people, and cloud cost of that coding — is a capitalizable direct cost. Non-coding algorithm exploration by researchers, which is really R&D for a separate intangible, is expensed under ASC 730. (Note that “training” an AI model is a different thing from the expensed “training” of personnel.)

Training-data costs (unsettled)

How you treat data acquired to train a model is genuinely contested, with three views in play:

• Capitalize as a separate intangible (ASC 350-30) — when the data is separately identifiable, controlled, and has alternative future use (e.g. usable to train multiple models);
• Expense (ASC 730) — when acquired for R&D with no alternative future use;
• Capitalize within ASC 350-40 — as a direct cost of an internal-use project without alternative use.

Major firms (Deloitte, PwC, Crowe) diverge here, and all agree that data with no alternative future use can’t be recognized as a standalone intangible. This is a live area — document your position and involve your auditors.

Why ASU 2025-06 hits AI hard

AI projects routinely involve novel, unproven functionality and requirements that keep shifting — exactly the significant development uncertainty that defers capitalization under ASU 2025-06. Expect more cost expensed early on novel AI builds, until the technology is proven through coding and testing and the scope settles.

Document it

Because so much rides on judgment, set auditable policies up front: separate capitalizable coding and compute from expensed research exploration; track training-data acquisition by alternative-use status; and record when novel functionality is proven through coding and testing. That contemporaneous evidence is what an audit will ask for — and what Quantify captures from your delivery tooling.