The Unit Economics Nobody Shows on Their AI SaaS Pitch Deck

I've reviewed more than 30 AI SaaS pitch decks in the last year. Every single one shows gross margin of 70–80%. Every single one is wrong — not because the founders are being dishonest, but because they're using a traditional SaaS cost model that breaks the moment LLM API spend becomes significant.

The error is consistent: LLM costs are modeled as infrastructure — fixed, amortized, scaling sublinearly with users. They're not. LLM API spend is variable COGS that scales directly with feature usage. When a user engages more, the cost goes up. When a feature becomes core to the product, the cost grows with every interaction. Traditional gross margin calculations that bury this in "hosting" or exclude it from COGS are showing investors a number that doesn't reflect reality.

Why Traditional SaaS Economics Break for AI Products

Traditional SaaS gross margin analysis is built on a stable assumption: infrastructure costs are largely fixed and decline as a percentage of revenue as you scale. AWS costs don't go up because a user spends more time in your app. Database costs don't double when usage doubles (at most scales). So you model COGS as a roughly fixed percentage and forecast healthy 70–80% gross margins at scale.

LLM API costs violate every part of this model:

• They're linear with usage. Every query costs tokens. More active users, more tokens, higher bill. There's no amortization.
• They vary by feature, not by user count. One feature can cost 50x more per interaction than another on the same platform.
• They grow with product success. When users find value in an AI feature and use it more, costs grow in lockstep — before any pricing change can respond.
• Power users distort averages. The top 10% of users by usage can account for 40–60% of LLM costs. Blended-average cost per user masks negative-margin cohorts.

None of this is fatal — it's a different cost structure, not a broken one. But it requires different metrics, different modeling, and different conversations with investors than traditional SaaS requires.

The Real Waterfall: From $29 MRR to Actual Gross Margin

Take a concrete example: an AI writing assistant at $29/month. Standard SaaS COGS math shows ~85% gross margin. Here's what it actually looks like when LLM API costs are modeled correctly as variable COGS:

Same product, same users, same engagement. Unoptimized: 41% gross margin. Optimized with routing and caching: 61%. The 20-point gap is the difference between an AI SaaS that passes investor due diligence and one that doesn't.

Unit Economics by AI SaaS Category

LLM cost as a COGS component varies significantly by product type. Here are realistic unoptimized gross margins by category — these assume typical pricing and average (not power) user behavior:

Agentic tools show the widest variance because each user action can trigger 5–15 LLM calls. At high engagement — which is what product success looks like — agentic features can push individual users into negative margin territory without feature-level cost caps.

The Five Metrics Your Pitch Deck Is Missing

Standard SaaS metrics (MRR, churn, CAC, LTV) don't surface the AI-specific economics. Here's what you need to add:

Four Scenarios: From Healthy to Crisis

To make this concrete, here's how the unit economics play out across four real-world usage patterns for a $99/month AI tool:

Scenario C is the one that kills companies quietly. The power users who trigger it are often your most engaged, most likely to refer, most vocal about product value — and each one is generating a loss. Without per-user cost attribution, you don't see it until the aggregate margin is already compressed.

What the Pitch Deck Should Actually Show

Investors who understand AI unit economics are starting to ask the right questions. Here's what a defensible AI SaaS unit economics slide looks like in 2026:

Show gross margin by tier, not blended. "Our enterprise tier runs 72% gross margin; our starter tier runs 48%" is a far more credible and useful number than a blended "65% gross margin." Blended averages hide cross-tier subsidy and don't tell you where margin improvement dollars should go.

Show the margin improvement trajectory. Current unoptimized state → planned state with routing and caching → target state at scale. The gap between current and optimized is credible evidence that the business model works — you just haven't deployed the optimizations yet. This is the AI equivalent of traditional SaaS infrastructure cost curves.

Show power user economics explicitly. "Our top decile of users by usage breaks even at $X/month. We've implemented usage caps at that level on the starter tier, and our enterprise pricing is designed to capture this segment profitably." This demonstrates you've thought through the problem — which most AI founders haven't.

Show LLM cost per unit of delivered value, not just per user. Cost per document processed, per task completed, per query answered — these tie LLM spend to business outcomes rather than presenting it as a raw cost item.

The Optimization Path: From Current to Defensible

Most AI SaaS companies are presenting their current, unoptimized LLM cost structure as the steady state. They're not. The optimization path is well-understood and the improvements are significant:

• Model routing (20–40% reduction): Route simple tasks to cheap models. A classification, yes/no determination, or short extraction doesn't need a frontier model. Identify your top 10 endpoints by cost and assess each one. Most teams complete initial routing in a few days.
• Prompt caching (15–25% reduction): Cache exact duplicate requests. SHA-256 hash the prompt, return cached response on match. Zero false positives, sub-millisecond overhead. Typical production apps send 15–20% duplicate requests.
• Usage caps by tier (structural protection): Hard limits on queries per day or month per pricing tier. Alerts fire after the fact; caps prevent negative-margin users from existing. Implement at the proxy layer — no application code changes.
• Feature-level attribution (visibility, not reduction): Tag every LLM request with the feature that triggered it. This doesn't reduce costs directly — it gives you the data to make routing and capping decisions. Without it, everything else is guesswork.

Teams that deploy all four typically move from unoptimized to optimized within two to four weeks. The resulting margin profile is what you should be building your unit economics model around, not the current baseline.

Gaurav Dagade

Founder of Preto.ai. 11 years engineering leadership. Previously Engineering Manager at Bynry. Building the cost intelligence layer for AI infrastructure.

LinkedIn · Twitter