I have spent the better part of 2025 talking about how AI changes our workflows. We look at prompts, we look at Python scripts, and we look at automation. But there is a massive side of this conversation I haven’t touched on enough: the accounting.
Companies are pouring billions into internal AI tools. Until recently the rules for accounting for that spend were stuck in the past.
I went through the latest RDCP update to understand how the Financial Accounting Standards Board (FASB) is finally catching up to reality. Specifically, I looked at ASU 2025-06 and how it amends ASC 350-40.
Here is what I found and how it changes the balance sheet.
The Old Way: Linear Logic
To understand the change you have to look at the old logic. Under the traditional ASC 350-40 rules for internal-use software costs, everything was governed by stages. It was a waterfall approach.
- Preliminary Stage: You are planning and evaluating. You expense these costs.
- Application Development Stage: You are coding and installing hardware. You capitalize these costs.
- Post-Implementation Stage: You are training and maintaining. You expense these costs.
This works fine if you are building a standard SQL database or a CRM. You plan it, you build it, you use it.
The Reality of AI Development
The problem is that AI development is not linear.
I know from my own experience building tools that it is messy. You might pick a model, test it, realize it hallucinates, and start over. You might switch from OpenAI to Gemini halfway through because the token cost is better.
The RDCP transcript highlights this conflict perfectly. The iterative nature of training, failure, and retraining broke the “stage” model. If you are constantly looping between planning and developing, tracking capitalization becomes a nightmare.
The Fix: ASU 2025-06
The new standard changes the metric. It removes the requirement to track distinct stages for AI software.
Instead of asking “What stage are we in?” the new question is “Is completion probable?”
The focus shifts from a rigid timeline to a judgment call. This acknowledges that generative AI and machine learning are fundamentally different from traditional software engineering.
The Decision Matrix
I broke down the guidance into a simple decision matrix to determine if you should capitalize or expense these costs.
The Test for Capitalization:
- Has management authorized the project?
- Is the technology feasible?
- Are resources committed?
The Outcome:
- If No: The project is treated as R&D or a pilot. You expense it.
- If Yes: Completion is “probable.” You capitalize external data, payroll, and cloud compute costs.
Important Note: Even if completion is probable, you still expense maintenance and ongoing training after deployment. That hasn’t changed.
Implementation Timeline
I looked at the dates to see when this actually hits the books.
- Public Companies: Fiscal years beginning after December 15, 2025.
- Private Companies: One year later.
You also have a choice on how to adopt this. You can go prospective (applying it only to new projects) or retrospective(revising past financials).
Tinfoil Hat Corner
Here is my speculation. Companies are spending absolute fortunes on compute power and data cleaning. If all of that had to be expensed because the project was “iterative” and didn’t fit the old development stage definitions, earnings would take a massive hit.
By shifting the rule to “probable completion,” it allows companies to move massive amounts of AI spend from the P&L to the Balance Sheet. This protects net income while they burn cash on GPU clusters. It feels like a move designed to keep tech valuations stable during a heavy investment cycle.
Key Takeaways
- Linear is Dead: The old stage-based accounting model doesn’t work for the iterative nature of AI.
- Probability is Key: Capitalization now depends on whether completion is “probable,” not which development stage you are in.
- Scope: This applies to internal-use software, specifically targeting generative AI and machine learning.
- Adoption: Public companies need to be ready by late 2025.
- Collaboration: Accountants can’t do this in a silo. You need to talk to the tech teams to know when a project shifts from “experiment” to “probable asset.”
Want to earn CPE for this topic?
- Compare Options: See how we stack up against others in our 2025 Flexible CPE Guide
- Understand the Format: Read how Nano-Learning works for CPAs.
- Check Your State: Ensure you are compliant with our State Requirements Guide.
- What is EverydayCPE?
Related Courses:
