Since artificial intelligence expenses appear in most engineering budgets, taxpayers claiming a federal tax credit for increasing research activities often raise two questions: How should AI-related costs be treated in the calculation? And how does AI affect whether the work still qualifies for the credit?
There is no IRS guidance on AI in this context, and Internal Revenue Code Section 41 doesn’t address it. Practitioners are left to work from a 2004 Treasury Department regulation and older cloud-computing analogies. That framework remains usable but creates interpretive risk when AI costs don’t fit neatly into traditional qualified research expense categories.
The research and development credit was built for a world where people performed research and computers served as tools. Today, work is performed with AI that can accelerate design, coding, testing, and analysis.
The Section 41 framework can still accommodate that shift, but only if practitioners are deliberate about classifying costs and the story they tell about the underlying research.
AI Spend Questions
Section 41(b) allows only a narrow set of qualified research expenses. Whether an expense fits within that category depends on how the service is delivered, billed, and used in the research process. Some examples include:
Cloud GPU and training compute.Spending on AWS, Azure, etc., for training or running AI models is the clearest case because it closely resembles traditional cloud-computing costs long treated as qualified research expenses. If the underlying activity qualifies, 100% of the expenditure should be included as a computer rental expense.
Application programming interface-based AI tools.Usage-based charges paid to providers can qualify if they’re described as paying for access to computing resources on the provider’s systems. The risk is that an examiner could view the charge as payment for a finished service rather than computer use.
Taxpayers should document how the tool was used in the development process and include documentation of the underlying and any follow-up queries. If engineers repeatedly use an application programming interface to assess different model outputs while building a new internal decision tool, that usage is easier to analyze than a general subscription with no project connection.
Software-as-a-service AI subscriptions.If the taxpayer is paying for a cloud-based tool and can connect part of that cost to qualified research, many practitioners include an allocated portion. If the arrangement looks more like a software license, the historical exclusion for software licenses may apply. If a development team uses a coding assistant partly for qualifying prototype work and partly for routine bug fixes, consider only the portion tied to qualified research for a tax exemption.
On-premises and self-hosted AI.If the taxpayer owns the hardware or runs the model on its own equipment, the analysis changes. Owned graphic processing unit clusters and self-hosted models don’t give rise to rental expenses, and depreciation on owned hardware isn’t a qualified research expense under Section 41. Wages paid to engineers working on the qualifying research may still count.
AI-assisted contractors.Where a third party is performing research and uses AI, contract research rules govern that 65% of the expense can be included as deductible qualified research, subject to the funded-research and rights-and-risk tests of Treasury Regulation Section 1.41-2(e). What matters is the contractual structure and the economic risk borne by the taxpayer.
Allocation is critical. A team using Copilot to write production patches, respond to support tickets, and prototype a machine-learning pipeline is engaged in qualified research only part of the time. A defensible claim requires a contemporaneous allocation method, whether based on seat-level usage data, project tracking, or engineer time studies.
What Counts?
The question isn’t whether AI costs exist, but whether qualified research does. To satisfy Section 41:
- The taxpayer must be trying to develop or improve a product, process, technique, formula, invention, or software
- The work must rely on principles of the hard sciences or engineering
- The taxpayer must face technical uncertainty
- Substantially all of the work must involve testing, evaluating, or refining alternatives (a process of experimentation)
AI affects the last two requirements:
Elimination technical uncertainty.When AI tools successfully produce workable code or a draft design, it becomes harder to argue that the taxpayer faced the same technical uncertainty that existed before these tools were available.
Uncertainty doesn’t disappear. Instead, it shifts to different questions, such as which model to use, how to assess reliability, and whether the result performs accurately in real-world conditions. If AI quickly generates a first draft of code for a new feature, the uncertainty may lie in whether that code can be integrated safely, scaled efficiently, and validated against technical requirements.
Process of experimentation.This asks whether the taxpayer evaluated alternatives through a process of testing and refinement. An engineer who simply accepts an AI-generated answer without comparing options may have a weaker case, even if the result is new.
By contrast, a team that uses AI to generate several approaches, tests those approaches, measures performance, and rejects some of them on technical grounds has a stronger experimentation narrative. The key point is documentation. Design documents, test results, code review notes, and records showing why one approach was selected will become important.
Track the cumulative effect.Two opposing trends are at play: AI tool spend will continue to grow as a share of the QRE base, while qualifying activity may narrow as routine engineering becomes quicker to conduct. Firms that document their approaches may still have meaningful claims. Companies reliant on AI without documentation may see qualified percentages decline as technology spend rises.
Going forward, taxpayers should create a clear AI cost framework. Separate recurring charges into cloud computing, API usage, SaaS subscriptions, on-premises operations, and contract research. Those categories don’t all follow the same Section 41 treatment, and a single catch-all “AI tools” account will make support harder.
They also should update research narratives. If AI assisted with the research, the study should explain where the uncertainty remained, what alternatives were evaluated, and how the team evaluated and refined solutions.
This article does not necessarily reflect the opinion of Bloomberg Industry Group Inc., the publisher of Bloomberg Law, Bloomberg Tax, and Bloomberg Government, or its owners.
Author Information
Martin Karamon is the tax credits and incentives advisory leader at Cherry Bekaert.
Michael Wronsky is managing director of Cherry Bekaert’s national tax practice.
Interested in writing? Review ourauthor guidelines,and submit pitches toInsights@bloombergindustry.com.
