A software company ships an AI feature. The compute bill lands somewhere in the financial statements. No line item says where.

The disclosure gap is structural

Software companies announce AI in product language. The cost arrives in accounting language, usually later and rarely under its own name.

Public companies must discuss known trends and uncertainties that are reasonably likely to have a material effect. They do not have to itemize every cost driver. The SEC has also said that existing rules may require AI-related disclosure when its use or risk is material to the business. That still leaves a wide gap between “material to the company” and “useful for understanding one product.”

Item 303 of Regulation S-K sets the broad MD&A requirement. The SEC’s AI disclosure review statement explains how those existing obligations can apply to AI.

Read three lines together

No single line resolves the question. The useful signal is the relationship among three.

Gross margin

Gross margin shows whether the cost is being treated as a cost of delivering the product. If an AI feature is included in a subscription, metered through a third-party API, or bundled into usage, inference can appear in cost of revenue and compress product margin.

A lower margin does not prove AI caused it. Hosting contracts, acquisitions, service mix, and pricing can move the same line. The direction tells you where to look, not what verdict to reach.

Free cash flow against operating income

If a company buys or builds inference infrastructure, the cash may leave before the expense reaches the income statement. Capital expenditure can widen the gap between operating income and free cash flow, especially during a buildout.

Again, the gap is not uniquely AI. It is a clue that the economics may sit below gross margin and outside the quarter’s operating expense.

R&D

R&D is where compute can sit while a feature is still being developed, before it is embedded in the product or used at meaningful scale. When the feature reaches production, some of that cost may move closer to cost of revenue.

None of the lines says “AI compute”

That is the point. The company does not have to name the ingredient for every product. You infer its likely location from movement, management commentary, and what the architecture would require.

The quiet case may be the most interesting. If AI usage rises and no line moves, the cost may still be immaterial, may be offset elsewhere, or may not yet have reached production scale. “No movement” is an observation, not proof of free inference.

Margin quality is a defense score

The score in the tearsheet is not a performance ranking. It asks how defensible the current price-to-cost relationship may be if inference becomes a commodity.

A high score suggests pricing power that does not depend on today’s cost of compute. A low score suggests gross margin may be supported by inference costs that still have somewhere to go. It is a way to rank questions, not companies.

The simulator separates use from payment

Usage attach and paying attach are different. A feature can be widely used and poorly monetized. It can also be narrowly used by customers who pay enough to cover inference. Combining both into one attach rate hides the decision a pricing team actually controls.

The default values are starting points, not industry averages. Change them. The value of the simulator is not the answer it produces; it is the assumption that breaks first.

What this study cannot prove

The table uses public filings and reported company figures. It cannot attribute a basis-point move to AI without company disclosure. Acquisition mix, restructuring, hosting contracts, pricing, and accounting policy can all produce the same pattern.

Yext’s fiscal 2026 revenue was $446.6 million and gross margin was 74.5%, down from 77.1% in fiscal 2025. Its first quarter fiscal 2027 filing reported $107.9 million of revenue and $78.7 million of gross profit. Those figures update the row. They do not prove why it moved.