Boards calibrated on the classic SaaS playbook are now asking questions the old benchmark decks don't answer. Three metrics have moved the most: gross margin, where AI economics rewrote the reference range; gross dollar retention, where the median quietly slipped; and internal AI usage, where boards want a number and almost nobody has a credible one. Here is where each actually sits, and how to report them.

What is a good gross margin for an AI company?

Roughly 50 to 60 percent for an AI-native product in 2026, against 75 to 85 percent for traditional SaaS. That is not a broken business; it is a different business, because inference sits in cost of goods sold. ICONIQ survey data shows AI-native gross margins improving from 41 percent in 2024 to 45 percent in 2025 to a projected 52 percent in 2026, and recent analyses put inference cost at roughly 20 cents or more per dollar of AI product revenue.

What matters for a board is the trajectory and the discipline, not the SaaS anchor. Report gross margin with and without inference, track inference cost as a percentage of AI revenue as its own line, and show the levers being worked: model routing, caching, prompt efficiency, and renegotiated unit pricing as model costs fall. A 55 percent margin with a visible path to 65 reads very differently from a 55 percent margin nobody can explain.

What is a good gross dollar retention?

Benchmark against segment, not against the blended median. Recent survey data puts the median B2B SaaS gross revenue retention near 84 percent, down from 88 percent a year earlier, with the top quartile around 91 percent. By customer segment, the long-standing SaaS Capital ranges still hold: 85 to 88 percent for SMB-focused companies, 90 to 93 percent for mid-market, and 94 to 97 percent for enterprise.

Two disciplines matter more in the AI era. First, gross retention before net: NDR flatters, GDR confesses, and every investor knows it. Second, cohort AI SKUs separately. AI products bought from experimentation budgets churn differently than core platforms, and blending them hides the signal a board most needs to see.

How should internal AI usage be measured?

Not by adoption percentage. Surveys now show nearly universal AI adoption alongside a striking absence of measured return, which makes "percent of employees using AI" a vanity metric. Usage is not the outcome; recovered hours are.

A board-credible framework has three parts. Coverage: the share of recurring workflows with AI in the loop, counted from a defined workflow inventory. Intensity: hours saved per employee per month, measured against a calendar rather than a feeling. Yield: AI spend as a percentage of operating expense, set against the value of the hours recovered. One slide, three numbers, each with an owner. Workflows that fail to save measurable hours within a month get killed, a rule covered in more depth in the note on where AI earns its keep in a finance function.

Benchmarks are calibration, not targets. The purpose of these three numbers is to have the conversation with investors on accurate footing: margins that acknowledge inference economics, retention measured where it is honest, and AI usage reported as recovered hours rather than enthusiasm.