The AI Gross Margin Trap and Outcome-Based Pricing

01From zero marginal cost back to variable COGS

In traditional SaaS, engineering is an upfront fixed cost; once shipped, whether a user clicks once a day or a hundred times a day barely moves the cloud hosting bill. Power users were pure upside.

AI applications invert that dynamic. Reviewing a contract, refactoring a module, or executing a multi-step agent workflow consumes tokens billed on every call. The more engaged the user, the higher the variable cost of goods sold (COGS) — turning gross margin back into an operational discipline.

02The power-user margin trap under flat seat pricing

Consider a seat-priced AI product charging $30/month: 90% of users are light users costing $5/month in inference, while 10% are power users costing $90/month:

Against a $30 seat price, the blended gross margin looks like a respectable 55% — yet the top 10% of users are quietly consuming two-thirds of total inference spend:

Under flat per-seat pricing, your most engaged champions become your biggest money-losers — paying $30 a month while burning $90 in tokens. Worse still, as product adoption matures and more light users graduate into power users, overall gross margin actively deteriorates.

03How three pricing models allocate cost risk

Escaping the power-user trap requires aligning the pricing unit with value and cost. The three dominant pricing models distribute margin risk very differently:

Flat seat pricing clears procurement easily but leaves the vendor exposed to uncapped token burn. Pure usage pricing protects vendor margins but makes enterprise budgeting unpredictable, causing customers to throttle adoption.

Outcome-based pricing — such as charging per resolved customer support ticket — offers the strongest margin leverage. Pricing is anchored against the cost of human labor, making ROI obvious to the buyer, while every backend optimization the vendor makes (model routing, caching, distilling smaller models) drops straight to gross margin.

04Falling token prices won’t save unoptimized workflows

While per-token API prices for a fixed capability tier continue to fall rapidly, token consumption per task is growing even faster.

As products transition from single-turn chat to test-time reasoning models and multi-step agent loops, a single user request can trigger dozens of internal model calls across massive context windows. A 50% drop in per-token price paired with a 10x jump in tokens per task still leaves you with a larger compute bill.

The windfall from cheaper inference accrues specifically to teams that price on outcomes and aggressively engineer their token efficiency.

05Four checks we run on AI gross margins

• Inspect cost by user decile. Break down inference COGS for the top 10% of active users against their revenue contribution to spot hidden power-user subsidies.
• Check the pricing anchor. Determine whether pricing is anchored to the human labor being replaced or merely marking up raw token consumption.
• Audit the cost-down engineering roadmap. Verify whether the team is actively deploying tiered model routing, prompt caching, and task-specific fine-tuned small models.
• Verify cohort unit economics as engagement scales. Ensure that rising usage within an enterprise account expands contract value faster than it expands inference bills.