R&D credit for AI development projects: what the records show and what to do next

Does AI development qualify for the R&D credit?

Some AI development work can support a credit claim, and some cannot; the answer depends on the facts of each project and on the tax rules, which the engagement team must apply. What advisers can control is the evidence. Experiment logs, evaluation results and decisions about what to try next are what a study reads, and they also describe how mature a client's engineering records are.

This guide is for CPA and advisory teams whose clients are building AI features. It covers the records a study examines, how to run the project interviews, and how to raise a licensing introduction separately and with permission. Trade press has covered how AI is reshaping software credit eligibility; this page does not restate eligibility rules and takes no position on them.

This is general information, not legal, tax or financial advice. Confirm with your own counsel, tax adviser or professional body before acting.

What records does an AI project study usually examine?

AI feature work leaves a different trail from conventional software. A study is likely to ask for these.

Note that coding assistants and hosted model APIs mean more of the trail now sits in vendor dashboards. Ask the client what is retained, for how long and who can export it before a tool subscription lapses.

Which of these records signal depth for licensing?

Licensing buyers look for records of real work with outcomes: multi-step workflows, decisions, tool use and results. Of the list above, pull requests with review threads, tickets with resolutions and incident notes are closest to that. A client that has kept several years of them in connected systems may be worth a quiet conversation. A client whose AI work started last quarter is not a candidate yet.

Raw model weights, training sets that were scraped or bought, and anything the client holds under a third-party license are not what this program is about. Rights come first: the client must have created the records or hold the right to license them.

The three-question depth check

Use this at the end of a study, from information you already have.

• Which systems: can the client name where tickets, code review, chat and incident notes live, and are they connected by ticket IDs or links?
• How long: do those systems go back several years, including any archived ones?
• Who owns it: is there a person who can run exports, and a sponsor (owner, CEO, CFO or authorized representative) who could approve a license?

Add the company baseline: 50+ full-time employees at peak (contractors excluded), several years of documented operations and rights to license. The company fit checker offers a non-binding screen, and who qualifies has the full list.

What permissions and promises limit what can be licensed?

AI companies often make privacy and confidentiality commitments to customers. In January 2024 FTC staff wrote, in guidance that is not a rule, that a company's promises not to use customer data for undisclosed purposes such as training models can be enforceable whether made in privacy policies, terms or marketing. For a client, the practical step is to read its own commitments before deciding what is in scope.

Rights review is between the company and SourceX. The adviser does not review, copy or describe the records.

What should the interview with the engineering lead cover?

Keep the credit interview and the depth check in one sitting, but record them in separate notes. Ask who ran each experiment, what the team expected, what happened, and what changed as a result. Then ask where the evidence lives today: which tracker, which repository, which channel. Ask whether any of it was moved when a tool was replaced, because migrations are where history gets lost. Finally, ask who would own exports if the company ever needed one. These answers fit on half a page and let you stop without opening a single record.

How do you raise it separately from the credit engagement?

1. Finish the credit work and fee discussion first; never tie a licensing introduction to the credit engagement.
2. Ask the CFO or owner for permission to mention an unrelated topic.
3. Explain in two sentences: companies license records they created for a one-time payment, keep ownership and approve scope, price and terms.
4. If they say yes, give your referral link or submit basic company details through the referral form.

You can also see how the same thinking applies to manufacturers and to engineering interview questions. For a CPA firm's wider role, see referral opportunities for accountants and the CAS growth guide.

How do rewards work?

Partners earn 25% of the eligible platform fees SourceX actually collects from the referred company's licensing deals, capped at $100,000 cumulative per referred company. Rewards are payable only after the buyer pays and SourceX receives its fee, and no reward is guaranteed. The reward is never deducted from the company's proceeds. Check your firm's independence and fee policies and your state board's rules first, and read the program terms.

When to skip it

• The AI work is under a year old and has no connected history.
• The client's training data or code is mostly owned by customers or third parties.
• Nobody can export tickets, code review or chat history.
• The data was already licensed for AI training.

Next step

At the next AI-feature credit study, note which systems hold the engineering trail and how far back they go. If the client qualifies and agrees, register as a partner and make the introduction.