Accounting for AI Founders

AWS, GCP, and Azure spend dropped into a generic Software Subscriptions account. That's fine for cash accounting, but it kills the §41 R&D study because nobody can tell which compute was used for qualified research and which was production inference.

Without that distinction in the books, the §41 study has to be rebuilt from invoices line by line. Often six figures of qualified compute gets missed because the data doesn't survive the trip from AWS console to general ledger.

Most firms treat all engineers the same. The IRS doesn't. Research engineers, ML engineers, and applied scientists generally qualify under §41. Customer support engineers, deployment engineers post-launch, and sales engineers usually don't.

Without segmentation in payroll, the QRE calculation is a guess. We've seen §41 studies built on a flat 80% allocation across engineering. That number is indefensible if the IRS asks for project-level documentation under the new Form 6765 Section G requirements.

AI products often charge per token, per call, or per compute-hour. ASC 606 treats these as variable consideration with specific recognition rules. Most bookkeepers recognize on cash, which is wrong, or recognize the contract value upfront, which is also wrong.

The downstream effect: deferred revenue is misstated, gross margin looks wrong, and the first investor who runs a revenue audit during diligence finds the gap.