AI product economics

Classical SaaS COGS is hosting and support. AI features add inference COGS: tokens in and out, sometimes tools, sometimes GPU time. Gross margin math matters more than in pure software. See gross margin.

A feature that sends 8k tokens of context per action at a blended cost of a fraction of a cent to a few cents can be fine at high seat prices — or fatal if every keystroke triggers a call. Caching, retrieval narrowing, smaller models for easy cases, and "run only on demand" are product decisions as much as infra ones. See latency and cost.

Walking through a power user

Seat price $30/month. A power user triggers 200 agent runs/day × $0.05 COGS = $10/day = ~$300/month. That one seat destroys margin.

Fixes are product-shaped: cap included runs, charge credits, route easy tasks to a cheap model, require confirmation before expensive tools, or redesign the workflow to fewer calls.

Budget eval and staging spend separately so experiments do not look like production runaway. Finance will ask for a gross-margin story on the AI line item. Percentile COGS is how you keep the feature shipped instead of hastily capped after a scary bill.

Pricing the work, not the tokens

Do not default to "price like tokens" unless you are selling tokens. Models meter tokens; applications should meter recognizable work or outcomes.

Credits should map to jobs ("1 research run", "1 document") even if internally you reconcile tokens. Pass through expensive reasoning models when costs are volatile; keep stable pricing on the platform layer you control.

Packaging a research agent: (A) $40/seat unlimited — simple sales, lethal if 5% of users run continuously; (B) $25/seat + 200 credits/month + $0.20/credit overage — teaches cost, needs a credit UX; (C) $2 per completed brief — aligns to value, invites disputes over "completed." Most B2B teams land near (B) or hybrid A+B. Revisit after you have real COGS percentiles.

Common mistakes

• Unlimited seats on an agent. One power user inverts margin.
• Pricing in tokens to customers. They buy jobs, not context windows.
• Looking only at average COGS. The p95 user is the one who blows the bill.

Check yourself

$30 seats, 200 runs/day at $0.05 each. What is monthly COGS for that user?

About $300 — ten times the seat price.

Why include credits on top of a seat fee?

The seat pays for the workflow and SSO. Credits stop power users from making COGS exceed revenue.

Key takeaways

• Watch inference COGS per outcome, at p50 and p95, not just total tokens.
• Price seats, credits or outcomes — recognizable work, not tokens.
• Caps, cheap-model routing and confirmation before expensive tools are product levers.

Further learning

You are not a model. Don't price per token (a16z)Charging per token is becoming the default for AI companies. For most AI applications, it is a mistake.Andreessen Horowitz

Next, defend the product in AI product moats.