Accounting for AI Development Costs: Expense vs Capitalize (GAAP)
AI work looks like “software,” “R&D,” “data,” and “cloud” all at once. That is exactly why accounting for AI development costs can get messy fast.
The goal is simple: record AI spend in a way that is accurate, consistent, and defensible in diligence. That means knowing when costs must be expensed now versus when they can be capitalized and amortized over time under U.S. GAAP.
Why founders should care
From a founder’s perspective, AI spend impacts three things immediately:
- Your burn and runway story (expense hits the P&L now, capitalization moves it to the balance sheet).
- Your fundraising narrative (investors often ask how much is “true R&D” vs “platform build”).
- Audit and diligence readiness (unsupported capitalization is a common red flag).
Capitalization is not a financial “hack.” It is an accounting conclusion that needs support.
The baseline rule: most AI research is expensed
Under U.S. GAAP, R&D costs are generally expensed as incurred. This includes many activities that AI teams call “development,” especially early experimentation and model iteration.
Common AI activities that are usually expensed (R&D):
- Exploring model architectures and training approaches
- Early prototyping and proof-of-concepts
- Evaluating data sources and labeling strategies
- Experiments to resolve performance uncertainty
- Failed training runs or abandoned approaches
If your team is still answering “Can this work?” you are usually in expense territory.
The key split: internal-use AI software vs software to be sold
To decide whether some AI-related costs can be capitalized, first determine which software accounting model applies:
1) Internal-use software (including most SaaS builds)
If you are building AI capabilities that power your internal systems or your hosted SaaS platform, internal-use software guidance is typically relevant.
Qualifying costs can be capitalized when you are past preliminary planning and you are building the software for use.
Also note: the FASB issued an Accounting Standards Update in 2025 that modernizes internal-use software guidance by removing explicit “project stage” references and focusing on a capitalization threshold tied to management commitment and “probable-to-complete,” effective for fiscal years beginning after December 15, 2027, with early adoption permitted.
2) Software to be sold, leased, or marketed (licensed software)
If you are developing software that will be sold or licensed (not hosted), a different model applies. Costs incurred before technological feasibility is established are expensed. In practice, that means a lot of early and mid-stage costs remain expense until late in the process.
Founder takeaway: Most AI startups offering SaaS apply internal-use software logic for development costs. AI startups selling licensed software often expense more for longer. Your monetization model matters.
What costs can be capitalized
Assuming you are in a capitalizable phase for internal-use software, focus on direct costs tied to building the AI feature or platform.
Often capitalizable (when criteria are met):
- Direct engineer compensation for building production code and integration
- Third-party developers working on the product build
- Certain cloud costs directly attributable to development activities (facts matter)
- Testing and implementation activities tied to getting the AI feature into production
Usually not capitalizable (still expensed):
- Research, experimentation, and early prototyping
- Ongoing model retraining that maintains current performance (maintenance)
- Sales, marketing, customer success, and go-to-market costs
- General overhead not directly tied to development
AI-specific nuance: model training and data labeling
Model training can be either research-like experimentation or production build, depending on intent and maturity.
A practical lens:
- If training is still resolving major uncertainty (accuracy, feasibility, performance requirements), it tends to behave like R&D expense.
- If training is part of a defined build (you have a committed plan, stable requirements, and you are preparing the model for production deployment), some costs may align with capitalizable development, depending on your policies and facts.
For many startups, it is cleaner to expense most training and labeling unless you have strong documentation and a stable build process.
A simple decision framework for founders
Use this as a monthly checkpoint.
Step 1: What phase are we in?
- Exploring feasibility: expense
- Building the production feature: potentially capitalize
- Maintaining or tuning after launch: expense
Step 2: What is our delivery model?
- Hosted SaaS or internal tools: internal-use software logic
- Licensed software sold to customers: software-to-be-sold logic
Step 3: Can we prove it?
If you cannot document the phase, the costs, and the approval, assume an auditor will push you to expense.
Documentation that makes capitalization defensible
Capitalization lives or dies on documentation. This is what you want in place before you scale spend:
- A written capitalization policy (what qualifies and what never qualifies)
- Project approval evidence (budget approval, product roadmap sign-off)
- A method to allocate engineering time by initiative
- A monthly capitalization memo or tracker showing:
- total eligible costs
- what was capitalized
- what was expensed
- why
This also improves your internal decision-making. AI spend becomes measurable, not just emotional.
If you want help building a capitalization policy, tracking AI build costs, and producing GAAP-ready schedules that investors and auditors can rely on, Ridgeway Financial Services supports high-growth teams with GAAP-ready accounting and software capitalization support.
FAQs
Can we capitalize AI model training costs?
Sometimes, but only if the work is part of a committed build with resolved uncertainty. If it is still experimentation, it is typically expensed as R&D.
If we capitalize, do we become “more profitable”?
Not economically. Your cash spend is unchanged. You are only changing timing of expense recognition.
What happens if we abandon the project?
Capitalized costs tied to an abandoned initiative may need to be written off.
Does SaaS mean we always capitalize?
No. SaaS often falls under internal-use software logic, but the phase and documentation still matter.
Reviewed by YR, CPA
Senior Financial Advisor