AI-Assisted Software Development: Do Programmer Wages Still Qualify for the R&D Tax Credit?
By Jon Dolen, EA, Tax Controversy & Strategic Account Advisor, DST Advisory Group
AI is changing software development at a remarkable pace.
Developers can now use generative AI and coding agents to generate code, troubleshoot problems, propose alternative architectures, run tests and identify potential solutions—often in a fraction of the time previously required.
For taxpayers claiming the Research and Development (R&D) Tax Credit under IRC Section 41, this raises an increasingly important question:
When AI becomes part of the development process, do the programmer’s wages still qualify as research expenditures?
There is currently no specific Treasury or IRS guidance addressing how generative AI affects qualification under Section 41. But that does not mean taxpayers are without a framework.
The better question may not be whether AI was used. It is how AI was used, what technological uncertainty existed, and what the developer continued to do to evaluate alternatives and resolve that uncertainty.
The Rules Haven’t Changed—But the Development Process Has
AI does not create a new test for qualified research.
Software development must still satisfy the requirements of Section 41. There must be a permitted purpose, the research must be technological in nature, uncertainty must exist regarding the capability, method or appropriate design of the business component, and substantially all of the research activities must constitute elements of a process of experimentation.
It is the last two requirements—uncertainty and process of experimentation—where AI-assisted development becomes particularly interesting.
Consider two developers faced with similar technical uncertainties.
The first asks an AI tool to develop a solution. The AI generates working code, and the developer implements it with little additional technical evaluation, testing or modification.
The second uses AI to generate several possible approaches. The developer evaluates those alternatives, identifies limitations, modifies the proposed architecture, conducts performance testing, rejects unsuccessful approaches, changes parameters and continues iterating until the technical uncertainty is resolved.
Both developers used AI.
But their activities may look very different under Section 41.
The regulations contemplate a process in which uncertainty is identified, alternatives intended to eliminate that uncertainty are identified, and those alternatives are evaluated through modeling, simulation, systematic trial and error or another evaluative process.
AI can potentially participate in that process without replacing it.
Where Did the Experimentation Occur?
This may become one of the most important questions for taxpayers using AI-assisted development.
AI can dramatically shorten the time between identifying a technical problem and arriving at a potential solution. Coding agents can increasingly perform tasks that historically required substantial manual work: writing code, debugging, running tests and even revising their own output.
That efficiency does not necessarily eliminate qualified research.
But it can change where the experimentation occurs and who—or what—is performing it.
If the AI generates alternatives and the developer meaningfully evaluates, tests, modifies and rejects those alternatives, there may still be a human-led process of experimentation.
If the AI generates a solution that the developer simply accepts and implements, demonstrating that process may become more difficult.
The distinction is not that AI is permitted in one situation and prohibited in another. The distinction is whether the taxpayer can demonstrate that the activities performed satisfy the requirements of Section 41.
The Documentation Challenge
AI-assisted development also creates a practical challenge: the traditional evidence of experimentation may be changing.
Historically, software development often left a recognizable trail—source-code changes, development tickets, design documents, debugging records, test results and successive versions of code.
AI can compress portions of that activity into a much shorter period of time.
A developer may evaluate several alternatives through an AI tool in minutes rather than manually coding each alternative over several days. Agentic development may compress that process even further.
For taxpayers, this makes contemporaneous technical documentation increasingly important.
The question should not simply be:
Did the developer use AI?
A stronger R&D file should help answer:
What technological uncertainty existed?
What alternatives were considered?
How was AI used in developing or evaluating those alternatives?
What did the developer test, analyze, reject or modify?
What technical criteria were used to evaluate the results?
How was the uncertainty ultimately resolved?
Those questions are not fundamentally new. But AI may make the answers less visible in traditional development records.
Documentation May Need to Evolve With Software Development
Taxpayers should consider whether their existing R&D documentation processes adequately capture AI-assisted development.
Depending on the development environment, relevant contemporaneous evidence might include prompts or instructions provided to AI tools, alternative solutions or architectures generated, developer comments, testing protocols, performance benchmarks, failed approaches, modifications to AI-generated code and records of subsequent iterations.
This does not mean taxpayers should archive every AI prompt simply because an employee may eventually be included in an R&D credit study.
The objective remains the same: retain sufficient evidence to establish the underlying qualified research activities and demonstrate the connection between those activities and the wages being claimed.
In fact, AI could ultimately improve R&D documentation.
Prompt histories, model outputs, testing logs and development records may provide contemporaneous evidence showing what alternatives were considered, how results were evaluated and how a developer’s approach evolved.
For tax departments, the opportunity may be to work with engineering and software development teams now—while companies are establishing their AI policies and development practices—to ensure that valuable technical evidence is not inadvertently lost.
What Should Taxpayers Be Thinking About Now?
Companies do not need to wait for specific AI guidance before evaluating their development activities.
The fundamental Section 41 questions remain familiar:
Did a technological uncertainty exist?
Were alternatives identified?
Was there a process for evaluating those alternatives?
What role did the developer play in that process?
And can the taxpayer substantiate that activity?
AI should not automatically disqualify a programmer’s wages simply because it accelerated the research process or generated potential solutions.
At the same time, the use of sophisticated AI technology does not transform routine development into qualified research. If AI produces a solution that is accepted and implemented without meaningful technical evaluation or experimentation, taxpayers may have greater difficulty demonstrating that the associated activities satisfy Section 41.
Until the Treasury or the IRS provides more specific guidance, taxpayers should continue applying the existing statutory and regulatory framework to the substance of the development activity—not simply the technology used to perform it.
AI may be changing how experimentation occurs. The challenge for taxpayers will be demonstrating where the experimentation occurred, who performed it, and how it resolved the technological uncertainty. Reach out to DST Advisory Group with questions.
About the author: Jon Dolen is an enrolled agent and retired from the IRS after 35 years of service, including 26 years involving extensive R&D tax credit experience.