The R&D tax credit for AI companies

Machine-learning development rests on computer science and applied mathematics, which satisfies the technological-in-nature requirement, and it proceeds through hypothesis and experiment: try an architecture, measure it against a baseline, refine or discard. That is a process of experimentation in the literal sense.

The uncertainty is genuine and documented in the work itself. Ablations, failed training runs, and abandoned approaches are the record of resolving questions whose answers were not knowable up front.

Developing new model architectures, training and evaluation pipelines, retrieval or inference systems, and novel data-processing methods generally qualifies when the outcome is uncertain at the outset.

Calling an existing model through an API with no development of your own, or routine integration work, does not. As with any claim, the defensible version separates the experimental work from the plumbing.

The experimentation that makes AI work qualify also makes it hard to reconstruct later: runs are numerous, branches are many, and little of it is written up for tax purposes.

R&D Binder derives the audit trail from the GitHub history directly, mapping qualifying work to business components and the four-part test without slowing the team down.

• What is a qualified research activity?
• What is the substantially-all (80%) rule?
• The Section 41 four-part test
• R&D tax credit for software companies

• 26 U.S.C. § 41, credit for increasing research activities - Cornell Law School, Legal Information Institute
• Treas. Reg. § 1.41-4, qualified research and the process of experimentation - Cornell Law School, Legal Information Institute
• IRS, Instructions for Form 6765 - Internal Revenue Service