Summary: AI is reshaping how companies develop software, and that spend is climbing fast. Guidance on how those costs fit the R&D tax credit hasn’t kept pace, so real dollars often go unclaimed. Learn how, with the right documentation, portions of your AI and cloud expenditures may be eligible for the R&D credit.
Does Your AI Spend Qualify for the R&D Tax Credit?
Almost every product and engineering team is now spending real money on artificial intelligence (AI), from cloud graphics processing units (GPUs) for model training and metered access to large language models to coding assistants, and increasingly, agents that carry out development work once done by people. That spend is climbing fast, and it raises a fair question: If this work is genuine research, why wouldn’t the AI costs behind it count toward federal and state R&D credits under IRC Section 41?
If you can tie the dollars firmly to qualified research at the business component level, then a portion of your AI compute can boost your R&D tax credit if it meets the criteria explained below. The risk of leaving an R&D credit unclaimed because AI-related spend was overlooked can be costly, underscoring the value of working with a proactive advisor.
How to Identify Qualified AI Costs
There is no Internal Revenue Service (IRS) guidance on how AI costs fit within the R&D tax credit and Section 41, and there is scant case law on these exact facts. Because these technologies are evolving faster than the guidance around them, the positions that follow rely heavily on the existing statutory framework of Section 41, the longstanding computer-use and contract-research regulations that predate these technologies, analogies to the established treatment of cloud-computing costs, the available substantiation case law, and seasoned professional judgment. As guidance develops, specific positions may need to be revisited or require the consultation of a knowledgeable tax advisor.
Start with the Activity, Not the Invoice
AI costs qualify only if they are consumed in the conduct of qualified research. To determine what constitutes as qualified research, the first step is to identify the initiatives that meet the four-part test of Section 41:
- Permitted Purpose – You are developing or improving the function, performance, reliability, or quality of a business component (e.g., product, process, platform, application, technique, formula, invention, or software).
- Technological in Nature – The work relies on the principles of physical or biological sciences, engineering, or computer science.
- Elimination of Uncertainty – You face uncertainty at the outset about capability, method, or design.
- Substantially all the work is a process of experimentation – You are forming and testing alternatives rather than just building a solution to a known spec.
When an AI tool quickly produces workable code or a credible prototype, it can be harder to show the technical uncertainty and process of experimentation the credit requires. However, that uncertainty usually does not disappear; it shifts. The related costs can still qualify where the taxpayer is working through questions such as:
- The appropriate design of the prototype(s),
- The reliability of the output,
- Whether it integrates securely, and
- The scalability and validation against real-world requirements.
A team that uses AI to generate several approaches, tests them, measures performance, and rejects elements of prototype models on technical grounds can still tell a strong experimentation story. Conversely, accepting the first AI-driven solution without evaluating alternatives is less likely to have a defensible case for AI-related costs, even if the result is novel or improves the overall product. Consider whether your AI tools are being used for development, testing, or both.
Taking 100% or 65%
After determining that AI was used in support of qualified research activities, the next important distinction is whether AI spend is treated as computer use at 100% inclusion or contract research at 65% inclusion. Keeping these straight is where a well-run study earns its keep.
Cloud and metered compute generally fit the computer-use framework and are included at 100% under IRC Section 41. That’s because the taxpayer is acquiring computing capacity it operates directly, rather than purchasing a research service performed by a provider . Reserved instances, compute commitments, and savings plans do not change that character; they are prepaid ways to pay for the right to use computers, allocated to qualified use on the same basis as pay-as-you-go.
Where a third-party performs research on your behalf and happens to use AI to do it, the contract-research rules govern. Under these rules, 65% of the qualified portion can be included, subject to the funded research, rights, and risk tests of Treasury Regulation Section 1.41-2(e). The contractual structure and who bears the economic risk are more important than the fact that the contractor used AI.
Types of AI Costs
A single AI or cloud invoice can bundle very different charges, and they do not all follow the same rules. The habit that helps most is simple: break “AI tools” into separate categories of costs before you quantify anything, because a single catch-all account makes support far harder. Here is the quick reference for some of the costs to be segregated:
| Cloud & AI Cost Typ | Section 41 Treatment | Key Considerations |
|---|---|---|
| Cloud and GPU or TPU compute (IaaS or PaaS) | Right to use computers, for the qualified pre-production portion. | Must be vendor-owned, off-site from taxpayer, and available for shared use; |
| API or token usage (metered access to a model) | Best positioned as computer use when it is access to the provider’s computing resources tied to development. | Document how queries drove the research |
| AI SaaS subscription or coding assistant (flat fee) | Generally, a software license and therefore excluded; a qualified portion is defensible only where usage data supports a computer-use characterization. | A flat-fee license is a historical exclusion. If the spend is genuinely usage-metered, it may be more properly considered API or token usage (row above). |
| Bots and agents that do development in place of headcount | Test for computer use eligibility first; if it does not fit, test for contract research eligibility. | Metered billing points toward computer use. Contract research requires that the taxpayer bear the risk and keep the rights; a fixed fee for a guaranteed result is generally not qualified contract research. |
| Owned or self-hosted hardware and chips | Generally, not a QRE. Property depreciable in the hands of the taxpayer is not supply or computer rental QRE. | Evaluate the inclusion of employees and/or contractors who are using hardware and chips to perform qualified research. |
| Third-party performing research for you using AI | Contract research, for the qualified portion | Must evaluate for substantial rights and economic risk |
Deductibility, Book Treatment, and Timing
Three questions travel alongside the credit:
- When is the cost deducted for tax?
- When is it capitalized or expensed for book?
- When does it count as a QRE?
These questions are governed by separate rules and often diverge; however, permissible book-to-tax differences are normal. The goal is to understand how each rule treats a given cost, helping to identify all credit-eligible costs and support a consistent, defensible position.
For the credit, the questions that matter most are when the qualified research activity occurred and when the associated expenditure is incurred. For domestic research, qualifying R&D costs, including the AI costs discussed here, are generally deductible in the year incurred under Section 174A. For tax purposes, costs tied to research conducted outside the U.S. are recovered over time instead, which is one more reason the location of the R&D matters. The finer mechanics of the deduction sit in Section 174 and related guidance; for a cost-side credit analysis, the aim is to coordinate the QRE timing with the tax treatment of the underlying costs.
The timing issue that comes up most with AI spend is prepaid versus consumed. Reserved capacity, compute commitments, and prepaid token bundles are paid up front but used over time. The more defensible position is to treat them as QREs in the year the compute is consumed, rather than the year it was paid for.
- For accrual-method taxpayers, the economic-performance rules for the right to use property generally point the same way, so the deduction typically falls in the year the compute is consumed.
- For cash-method taxpayers, it generally falls when the amount is paid.
- For book treatment purposes, treatment is governed by GAAP and will not always match the tax result.
- For credit computation, the credit should generally follow the year when the cost is deducted for tax purposes.
Importance of Supporting Documentation
Every position above depends on tying the spend to qualified activities and business components. For AI and cloud compute, the records that carry the most weight are:
- Allocation method: an objective driver, such as instance-hours, GPU-hours, environment tags, or project codes, that splits qualified experimentation from production, maintenance, and general business use. Claim the allocated portion, not the whole bill.
- Usage and invoice detail: provider invoices tied to the tagged projects or environments, with the consumption records behind them.
- Contemporaneous technical support: design documents and experiment plans showing objectives, uncertainties, alternatives tested, and how results were evaluated, mapped to specific business components.
- Contracts and statements of work: for any vendor performing research on your behalf, show the pre-agreement, non-contingent payment, and your rights to the results.
- Updated research narratives: where AI assisted with the work, the narrative should say where the uncertainty remained, what alternatives were evaluated, and how the team tested and refined the solution.
One commonly overlooked point is that the same user-level usage data that supports the compute claim also helps substantiate qualified wages. Traditionally, wages have been the largest QRE category for most companies, and supportability often relies on the ability to show which specific people performed, supervised, or supported qualified research. For example, a log that ties a named engineer to hands-on work on a specific business component, through their token or tool usage, is strong contemporaneous evidence that their time belongs in the wage base. Capturing who is using the AI and for what purpose pulls double duty: it supports the AI cost, and it corroborates the wages behind it. Companies that fail to capture these costs and provide documentation to support them can leave significant savings on the table.
The case law rewards this kind of recordkeeping. Courts have allowed reasonable estimates when a taxpayer can prove qualified research occurred and provided credible evidence (e.g., as in the cases of Suder, McFerrin, and Fudim) to support the claim. By contrast, courts have denied credits based on estimates where the reconstruction was unreliable or lacked a credible tie to specific qualified activities (e.g., as in the cases of Eustace and Shami). Contemporaneous records, paired with credible explanations, can help move a claim from the second group to the first.
State Credits & the Apportionment Trap
In addition to the federal R&D credit, most businesses can also benefit from one or more of the many state-specific R&D credits. These state credits often build upon the Section 41 definitions, but the details vary widely: different bases, caps, refundability, and different conformity to the federal rules. That patchwork can change both the state result and its timing. A federal position is not automatically a state position.
For cloud and AI compute specifically, the sourcing question deserves careful attention. States can differ in how they treat where qualified cloud costs are incurred, and the answer is not always intuitive. For example, it is still unclear whether qualified cloud and AI compute costs should follow the location of the research team or the location of the infrastructure. Because this area remains largely uncodified, taxpayers should not assume a federal position or one state’s approach, which is not guaranteed to apply in another state. Because this area is nuanced and can move the numbers, these decisions belong in a careful state-by-state analysis.
Moving Forward: How to Document AI Spend for R&D
- Breakdown your AI spend into categories, cloud and GPU compute, API and token usage, SaaS subscriptions, on-premises operations, and contract research, rather than one “AI tools” account.
- Tag compute spend to environments and business components so qualified experimentation can be separated from production, maintenance, and general use.
- Keep the queries, usage logs, and experiment records that show how the tools drove the research, not just that they were paid for.
- Map where development happens, since Section 41 does not reach research conducted outside the U.S.
- Run the state-by-state review, including sourcing of cloud costs, rather than assuming the federal classification governs.
- Partner with an experienced Aprio R&D tax advisor to validate cost treatment, strengthen documentation, and identify opportunities that may otherwise be overlooked.
Final Thoughts: Strengthen Your R&D Tax Credit Position for AI Spend
AI has changed how research is performed, and Section 41 framework can still account for it. The value of claiming AI compute costs lies in classifying each cost correctly, documenting the research behind it, and aligning the credit with its related federal and state considerations.
This is the kind of detail that is easy yet expensive to miss, and it is worth a proactive look before year-end. Contact Aprio’s R&D team today, so we can help you Account for Anything™.
