Spotting R&D Tax Credit Opportunities in Clients' Custom AI Projects

Somewhere in your client list is a distributor that stopped buying document-processing software and started building it. An insurance agency is testing ways to pull and validate data out of carrier documents. None of these companies think of themselves as technology companies, and their tax questions still arrive as entity elections and depreciation.

Yet some of that work may be qualified research under Section 41, and the facts that would support it are being created right now, in tickets and test results nobody has been asked to keep.

The credit rewards whoever is in the room before the work happens. Specialty study shops arrive after year-end and reconstruct. You are already in the room.

The law stopped fighting you

Many clients quietly stopped asking about the credit during the Section 174 years, and it is worth being precise about why. From 2022 through 2024, domestic research expenditures had to be capitalized whether or not anyone claimed a credit, so declining the credit never restored the deduction. The cash-flow hit soured clients on the subject, and for smaller claims the credit did not look worth the study cost. Different problem, same result.

Section 174A permanently restores immediate expensing of domestic research expenditures for tax years beginning after December 31, 2024, with an election to capitalize over not less than 60 months, and Section 41 qualified research is now defined by reference to it. Transition rules apply to previously capitalized amounts, so this is not simply everyone deducts everything. Foreign research sits under two separate rules: 15-year amortization, and independent exclusion from qualified research under Section 41(d)(4)(F). Where a client’s developers sit is not only a cost question.

One timing point belongs in every planning file. The qualified small business payroll offset is capped at $500,000 per year for up to five taxable years, though the cap is not the number, since the election cannot exceed the credit actually available. Eligibility turns on gross receipts, not on what a founder calls revenue: generally less than $5 million for the election year, and none before the five-taxable-year period ending with that year. That is a taxable-year test, not five years from a first sale. The election must be made on a timely filed original return. An amended return may still support an ordinary income tax credit claim under the refund claim rules, but it is not where this election can first be made, so for an early-stage client a missed filing is cash that does not come back.

One question starts the analysis

You do not need to become a technologist to screen this. You need one distinction, applied consistently: is the client configuring, or building?

Buying an AI license does not create qualified research. Neither does deploying a standard chatbot or configuring a commercial platform to do what it already supports. The analysis changes when the purchased product cannot meet the requirement and the client’s own people start developing something new around it. That is where technical uncertainty moves in-house, which is where the analysis starts, not where it ends.

Ask the client four things, in their own language. What were you trying to make the system do that it could not already do? At the start, did you know whether it was achievable, or which approach would work? What alternatives did you try, and how did you measure which was better? What did the results change?

Thin answers mean adoption, and you should say so plainly. But treat strong answers as a preliminary screen and nothing more, because it does not test two things that decide cases. The uncertainty must be technological, resolved through a process fundamentally relying on engineering or computer science, not commercial doubt about whether anyone will buy the result. And Section 41 excludes work that looks identical from the outside: adapting an existing business component to a particular customer, duplicating something that exists, research after commercial production begins, and ordinary debugging, testing, and quality control. Substantially all of the activities, generally at least 80 percent measured on a reasonable basis, must constitute elements of a process of experimentation. The Seventh Circuit’s decision in Little Sandy Coal is worth an hour on how little generalized novelty and estimated time allocations actually buy.

Qualification also attaches to activities and business components, not to companies, so “we are an AI company” proves nothing and “we are a distributor” excludes nothing. And while a failed project can contain qualified research, failure by itself proves nothing either. Projects die for budget and data-access reasons that never involved experimentation.

The two questions that decide who holds the position

Once a project looks real, two issues determine whether your client is the taxpayer who can claim it. Both are easier to resolve before contracts are signed.

The first is internal-use software. AI developed primarily for general and administrative functions must also clear the high threshold of innovation test in Treas. Reg. 1.41-4(c)(6): innovative, in the regulation’s sense of a substantial and economically significant measurable improvement in cost or speed rather than merely customized or impressive; significant economic risk; and not commercially available for the intended purpose. Practitioners overreach on the exception. Software developed to be sold or licensed generally is not internal-use software, and software letting third parties interact with the client may fall outside it too, but software developed for both purposes is dual-function software, presumed internal-use unless a third-party subset can be identified, with a 25 percent safe harbor where anticipated third-party use is at least 10 percent. Intent at the start of development matters, and someone using the software to support the client’s own administrative functions is not a third party here. Bolting a customer portal onto an internal system does not convert it.

The second is funded research. When a client pays an outside agency, the client may generally treat 65 percent of the qualifying portion of eligible contractor payments as contract research, but only where the agreement predates the research, the work is on the client’s behalf, the client bears the expense even if the research fails, and the client has a right to the research results. A right, not exclusive ownership: the contractor can retain intellectual property and the client can still qualify, which is the opposite of what most people assume on reading an IP clause. The 65 percent also applies to the qualifying portion rather than the invoice, and a line item reading “development” establishes nothing about the underlying activities.

Run it from the other side when your client is the developer, keeping two tests separate. Did the developer bear the risk of research failure, meaning payment depended on technical success rather than effort? And did it retain substantial rights to use or exploit the results? Bearing risk does not cure the absence of rights. A fixed price or capped fee is not by itself research-failure risk, which is where Geosyntec turned, because ordinary cost-overrun exposure is not the statutory risk. “Fixed fee,” “time and materials,” and “the client owns the IP” are useful facts, and none answers the question alone. Risk and rights can also be allocated so poorly that neither party has a clean position.

Where documentation actually comes from

The evidence is abundant while development happens and thin afterward. Architecture decisions, evaluation datasets, benchmark results, abandoned approaches, tickets, and contractor invoices all sit in the client’s systems while the project runs. By the following spring, what is left is a memory and an invoice total.

There is no contemporaneous-documentation safe harbor, and testimony and reconstruction can be considered. What gets rejected is the generalized narrative and the estimated allocation, which is exactly what reconstruction produces. So the test for a client’s records is not whether they show activity. It is whether they connect a specific business component to the uncertainty at issue, the alternatives evaluated, the method and results, the people who performed or directly supervised the work, and the associated wages, supplies, and contractor amounts. That is also the shape of the data the redesigned Form 6765 Section G calls for, generally required for tax years beginning after 2025 with optional reporting for several categories of smaller filer. Section G has been revised more than once, so check the instructions for the filing year rather than a rule of thumb.

One more thing, before the client starts multiplying. The credit is incremental, computed against a base, so qualified expenditures do not convert into a flat percentage of research spending, and a Section 280C election changes the amount again. Only certain categories are QREs at all. An architecture document is evidence and a cloud invoice is a cost, and neither is automatically a qualified research expenditure.

Adding it to the practice without becoming a study shop

None of this requires your firm to perform credit studies. The role that fits a general practice is narrower and more defensible: recognize the pattern early, screen it honestly, preserve the facts, and hand a clean file to a specialist for computation and the filing conclusion.

The practical version is small. Add the configure-or-build question to your year-end planning agenda. When a client mentions an AI project, ask who is doing the development and where they sit. When a client is about to sign a development agreement, ask to see it before signature rather than after. Flag the payroll offset deadline for every early-stage client that might qualify.

The most valuable output of that screen will usually be a no, delivered with reasons. Most AI adoption does not qualify, and saying so credibly is what makes the yes worth something. We are putting our own firm’s AI development through the same screen, and partial qualification is the realistic expectation there too. The clients who need this conversation are not asking for it, because they do not know the work they are already doing has a name in the Code.

This article is general educational information current as of August 2026 and is not a determination that any taxpayer, project, activity, or expenditure qualifies for a credit. Eligibility depends on specific facts, contracts, activities, expenditures, locations, and applicable law, and the authorities cited arise in particular circuits.

—–

Chea Romine, CPA, is managing partner of RHW CPAs, a 30-person firm with offices in Ohio and California, and the founder of an AI automation company. He works with firms and their clients on recognizing research credit opportunities in custom AI development.