SAP's CEO Christian Klein said it plainly: "It would be foolish to still charge subscription base, because AI is so powerful that it will automate a lot of tasks." SAP lost roughly 20% of its market value as investors tried to figure out whether the transition would work. The revenue model that built your company may not survive the product you're building next — and CFOs who haven't rebuilt the numbers are walking into board meetings blind.
What breaks in your financial model when you add an AI product?
Adding AI features converts fixed COGS into variable COGS. Gross margin can drop 15 percentage points before the customer's price changes.
Traditional SaaS COGS — hosting, support, amortized development — doesn't vary with usage. An AI product's COGS does. Every prompt, agent action or generated output triggers a real compute cost. The SaaS CFO modeled it: $100 in SaaS revenue with $20 in traditional COGS gives 80% gross margin. Add AI features. COGS rises to $35. Margin drops to 65%.
The second break is in how you measure revenue growth. Seat-based ARR tracks access, not usage or value delivered. An AI product delivers value at the moment of use. Seat count can stay flat while actual value delivery doubles, and ARR misses that entirely. It can also move in the wrong direction: if AI agents handle work that previously required human users, seat count falls even when the product is succeeding.
Why are AI product gross margins lower than traditional SaaS?
AI product companies average 50 to 60 percent gross margins in 2026, compared to 70 to 80 percent for traditional SaaS, per ICONIQ's 2026 State of AI survey.
The gap comes from inference costs: the per-use compute expense of running a language model, embedding or agent workflow.
GitHub Copilot's early economics made the problem concrete. The subscription was $10 per user per month. For heavy users, compute costs ran up to $80 per month. The product was working. The economics were not. For every dollar of AI product revenue a company books, a share walks out the door as inference cost before a single employee gets paid — and that share shifts with prompt length, model choice and usage patterns without any price or contract change.
| Metric | Traditional SaaS | AI-Native Product |
|---|---|---|
| Gross margin (2026) | 70–80% | 50–60% |
| COGS structure | Fixed infrastructure + support | Variable inference + fixed infrastructure |
| Primary revenue driver | Seat count | Usage volume or outcomes delivered |
| Primary health metric | ARR + NRR | Gross margin per interaction + NRR |
| Pricing model (2026) | Per-seat subscription | Hybrid (base fee + usage) or outcome-based |
Subscription, consumption or outcome-based: which model fits AI products?
Hybrid pricing outperforms both pure approaches. A base fee plus a usage layer ties revenue to actual delivery while keeping customer forecasts predictable.
The numbers support the shift. Hybrid pricing jumped from 27% to 41% of companies within a single year, according to getmonetizely's analysis of 2025 pricing trends. By 2025, 92% of AI product companies were running some form of mixed model. Companies still on pure per-seat pricing reported gross margins roughly 40% lower than those using usage or outcome-based approaches.
Outcome-based pricing — charging per result rather than per use — offers the highest theoretical margin but creates the hardest forecasting problem. Sixty-four percent of SaaS finance executives cite unpredictability as their top concern with outcome-based models. For a CFO building a board forecast, that unpredictability has a direct cost. Replit grew from roughly $2 million to $144 million in ARR and improved gross margins significantly by moving to usage-based pricing rather than pure subscription.
Nexairi Dispatch
Get the next AI move before it turns into a workflow problem.
Join the free newsletter for concise AI news, practical checklists, and the decisions practitioners need to make next.
Free. No spam. Unsubscribe anytime.
What metrics should CFOs track for an AI product line?
Three metrics take over from seat-count ARR as the primary health signals for an AI product.
The first is gross margin per AI interaction. This is the unit economics test for your product. If the margin on each interaction improves as usage scales, the economics are working. If it's flat or declining at scale, you have a token tax problem or a pricing problem — and you need to know which before the next board meeting.
The second is Inference Efficiency Ratio: revenue generated per dollar of inference cost. A rising ratio means the product is delivering more revenue value for each compute dollar spent. A declining ratio at scale is an early warning that pricing isn't keeping pace with usage growth. CFOs tracking AI product lines are starting to use this alongside gross margin as the two-number health check.
The third is net revenue retention from expansion. NRR matters more for AI products than traditional SaaS, because successful AI products deepen usage rather than add seats. If customers use the product more but revenue per customer doesn't grow, the pricing model has a structural leak. Salesforce Agentforce hit $800 million in ARR by February 2026, growing 169% year over year, with more than 60% of Q4 bookings from existing customers expanding.
Doesn't token cost deflation fix the margin problem over time?
Token costs fell roughly 98 percent between early 2024 and 2026, and that deflation does improve margins — but it does not eliminate the model reset.
But it doesn't eliminate the model reset. AI product companies were still seeing 15 to 23 percentage points of margin compression in 2025, when most boards started asking hard questions. And product complexity tends to grow to match cost reductions: as models get cheaper, products use longer context windows and more agent steps. The cost per token shrinks, but total inference spend often holds steady.
What goes in front of the board when you're mid-pivot?
Three numbers need to be ready before the board asks whether the transition is working or you're trading durable margin for uncertain revenue.
First: gross margin broken out by product line. Not blended. If AI products and traditional SaaS products are combined into one gross margin figure, the board cannot see where the compression is coming from or whether it's improving.
Second: AI COGS as a percentage of AI revenue. Payhawk's 2026 CFO framework names this explicitly as a required step — reclassify AI costs from OpEx into Cost of Goods Sold to expose true gross margins. If AI infrastructure spend is sitting in OpEx, the P&L is hiding the real picture from everyone who reads it.
The CFOs closing the gap aren't the ones with the best products. They're the ones who rebuilt the financial model before the board meeting, not after it. That's the move. The numbers are available. Most CFOs just haven't reorganized them yet.
The blended model is where visibility disappears
The most dangerous position mid-pivot is reporting blended gross margins that average a declining AI product line against a healthy legacy SaaS business. The board sees 72% gross margin and stops asking questions. Then legacy SaaS matures, growth slows and the AI margin problem surfaces when there's no cushion left. Separate the product lines before the board forces it. That move buys time to fix the economics rather than explain them.
For CFOs building the broader AI ROI case alongside product revenue, Nexairi's How to Measure AI ROI as a CFO covers the metrics that hold up under board scrutiny. If the AI pivot is still early stage, CFOs Funded the AI Revolution. Most Didn't Get One. maps why execution determines whether the investment pays off — and why the sequence matters more than the technology.
Sources
Related Articles on Nexairi
Free Assessment
Is your firm ready for AI?
A 5-minute governance check for CPA firms using ChatGPT, Copilot or AI accounting software. Get your score and your top gaps — free.
Curated insights from the NEXAIRI editorial desk, tracking the shifts shaping how we live and work.
