AI and exit value · Answered by Golden Section from more than 400 B2B software companies observed
Keep blended gross margin above 65% with a path to 72% or better, and expect the AI line inside it to run lower while it matures. We underwrite AI-delivered revenue separately at 50% to 60% gross margin, against roughly 76% to 80% for traditional SaaS, and AI products run at roughly 53% in our research. A low AI margin is acceptable when it sits on its own line, is priced to cover inference and improves as per-token costs fall. It becomes a problem when inference is buried in hosting and grows faster than the revenue it serves. Rent model access from providers, keep self-hosted open-weight models as the fallback, and fund AI work against a named customer outcome. Start by separating AI inference revenue and cost, then model margin at three times current usage.
Report AI revenue and inference cost as their own lines and judge the AI margin on its trend, not its first quarter. Hold the blend above 65%, rent model access with self-hosted open-weight models as the fallback rather than owning hardware, and spend on AI only where it moves an outcome the customer already pays for.
| Metric | Value | What it means | Source |
|---|---|---|---|
| Median AI product gross margin | roughly 53% (2026), from 41% in 2024; 59% projected for 2027 | gross profit ÷ revenue on AI-delivered productsICONIQ State of AI, about 300 AI product leaders per wave, as cited in our research; traditional SaaS runs about 78% in the same comparison | Golden Section, publishedInvesting in Software addendum, September 2026 |
| How Golden Section underwrites gross margin | 65–85% blended; AI-delivered revenue separately at 50–60% | gross margin after all cost of revenue, including inferencecost of sales now assumed at 15–25% of revenue with an explicit inference line, up from 10–15% | Golden Section, publishedInvesting in Software addendum, September 2026 |
| Effect of the AI mix on the blend | about 68% | 42% of revenue at roughly 53% margin blended with 58% at 78%industry-level arithmetic in our research; margin compression is worst below $5M ARR, down 7 points below $1M and 4 points at $1–5M (High Alpha) | Golden Section, publishedInvesting in Software addendum, September 2026 |
| Committed model capacity | low single-digit % of revenue | provisioned capacity bought from the primary model providerbudget it as insurance sized to the revenue that depends on AI, not as a discount | Golden Section playbookAI Vendor Continuity |
| Engineering and product allocation | about 45% core / 45% AI-native features / 10% skunkworks | share of engineering and product resourcesour suggested split for incumbent software companies | Golden Section, publishedThe Enhancement Doctrine, March 2026 |
Seat-based software had a nearly fixed cost of delivery, mostly hosting, so every new dollar of revenue arrived at 75% to 85% margin. Inference changes that. It is a variable cost that rises with usage, so the agentic feature that lifts contract value also brings in a cost you do not control, and a flat-fee price with unlimited use can turn the AI line negative as customers adopt it. That is why the line has to stand alone. Buyers and lenders will ask whether AI revenue scales and sticks, and only a separated line answers.
The P&L should carry inference in cost of revenue from the first invoice, and unit economics should carry the cost to serve per account. Pricing decides whether margin holds: the AI pricing play covers when to bundle, meter or price the outcome. On build versus buy, the vendor continuity play is direct: build portability and buy committed capacity in proportion to revenue at risk. Owned hardware turns a margin question into a balance sheet question, and it rarely pays at this scale.
The figures are invented. A company at $5M in annual revenue earns $4.4M from its core subscription at 78% margin and $600K from an AI add-on sold at a flat fee per location, costing $330K in inference, a 45% AI margin. Blended margin is 74%. Modeling three times current usage, inference rises to about $990K, the AI line loses $390K, and blended margin falls to about 61%, below the floor. The company keeps the flat fee for new customers but adds a usage allowance with credits above it at each renewal, and routes routine requests to a cheaper model that passes its evaluation suite. Nothing changes mid-contract.
Outcome-priced AI that replaces services labor can earn a margin above the subscription, because the price follows the value rather than the token count. And a company that has just launched AI features may see the line run below 50% for a quarter or two while usage settles; that is fine if it is separated and improving.
From the Golden Section mistakes list, each paired with the play that prevents it.
An AI price set without reference to customer value or inference cost is how a flat fee turns into a negative margin line.
Margin assumptions break when a model vendor raises prices, tightens limits or deprecates the model your product is tuned to.
Model and hosting contracts with uncontrolled escalations and auto-renewals flow straight into cost of revenue.
In the order we would run them. Each is on its own page, most with a free Excel template.
Puts inference in cost of revenue so the AI margin is visible from the first invoice.
Carries cost to serve per account into payback and lifetime value.
Chooses a pricing model that covers inference and sequences any change to the existing base.
Answers build versus buy: portability first, committed capacity sized to revenue at risk, hardware almost never.
Vendor plays AI puts a third party inside your cost of revenue, and these plays cover the contracts and continuity that keep that cost under control.
Buy model access and build portability. Route every call through one abstraction, keep a second provider carrying a small share of live traffic, and buy committed capacity in proportion to the revenue that depends on AI. Treat a self-hosted open-weight model as the fallback you have tested, not the default. Reserved GPUs or owned hardware make sense only after a vendor has actually restricted you or volume would keep the hardware genuinely busy.
Size it as a share of engineering rather than as a separate budget: our suggested split is about 45% protecting the core, 45% building AI features into existing workflows and 10% on a skunkworks. Budget inference as cost of revenue, and committed capacity as an insurance premium. Every AI project should name the customer outcome it moves before it starts.
Show AI revenue, inference cost and AI margin as their own lines beside the blended figure, monthly, with the trend. Add attach rate and retention for AI accounts. A buyer will compute the blend either way, so the separated view is what lets you explain it.
Gross margin after inference is one of the first numbers we check: our lending starts at 65%, and 68% for term loans. AI work that has not yet reached revenue is a product investment for equity, not something to finance with debt.
Growth capital lending →Reviewed by Dougal Cameron, CEO & Co-Founder on 2026-10-22. Golden Section observations are labeled separately from external benchmarks and illustrative arithmetic.