Your dashboard says 65% gross margin. Your board deck says the unit economics work. Both can be true and you can still be losing money on your best customers. AI agent gross margin is the most misleading number in your business because it's almost always reported as a blend, total revenue minus total compute, divided by revenue, and a blend hides the handful of power users quietly consuming a disproportionate share of your variable cost. Anthropic learned this the expensive way, ending its unlimited $200-a-month plan after discovering a single customer had run up tens of thousands of dollars in compute. Here's the math behind that and how to find your version of that customer before it scales.
Why AI breaks the blended-margin assumption
For two decades, blended gross margin was an honest number in SaaS, because the marginal cost of one more user was effectively zero. A power user who logged in fifty times a day cost you nothing extra; they were free upside. Reporting an average was fine because the average described everyone.
AI products break that assumption at the root. The marginal cost of one more action is real, tokens, tool calls, retries, compute, so a power user isn't free upside, they're a cost center. Bessemer found a cohort of fast-scaling AI startups averaging just 25% gross margin in their early stages; AI products generally land at 50 to 60% against the 80 to 90% that classic software enjoyed. As Cursor's Jacob Jackson put it, when a customer pays you $10, you can't turn around and spend $10 on compute. With variable costs that real, the distribution of usage matters far more than its average, and usage in AI products is almost always a power law, where a small slice of customers drives a large share of the runs. Blend a power-law cost curve with flat pricing and your reported margin is being carried by light users subsidizing heavy ones.
The math of a negative-margin customer
A simple model shows how a healthy average hides a sick tail. Take a flat $200-a-month plan with 100 customers. Ninety of them use the product lightly and cost you about $20 each to serve, a 90% margin. The other ten are power users who cost you $260 each, a margin of negative 30%, or a loss of $60 every month, per customer.
Now blend it. Revenue is $20,000. Cost is (90 × $20) + (10 × $260) = $1,800 + $2,600 = $4,400. Blended gross margin is ($20,000 − $4,400) ÷ $20,000 = 78%. That number looks excellent. It would survive any board meeting. And it is actively concealing that you lose $600 a month across ten customers, the ten who, because they get the most value, are the most likely to expand, refer others like themselves, and show up in your next cohort. As you sell to more of your happiest customers, the loss compounds and the blend slowly deteriorates. The number that looks healthiest is the one hiding the problem.
Why flat and unlimited pricing guarantee the leak
Flat and unlimited pricing don't just risk this outcome, they guarantee it, because they sever price from cost completely. The customer who extracts the most value, and therefore costs the most, pays exactly what the lightest user pays. Worse, there's a selection effect working against you: power users are disproportionately drawn to unlimited plans precisely because they intend to use them heavily. You are advertising directly to your least profitable segment.
This is why Anthropic ended its unlimited tier and why a string of AI-era products, GitHub, Replit, and others, bolted usage components onto flat plans once their margins dove. Flat pricing is comfortable to sell and quietly expensive to keep. The eventual fix is usually a move toward usage or hybrid pricing, but you can't make that move intelligently until you can see which customers are underwater.
How to find your negative-margin customers before they scale
The fix starts with looking at the right number. Four steps surface the problem. 1. Compute margin per customer, not blended: each customer's revenue minus their fully-loaded cost to serve. That requires attributing cost down to the customer, which means knowing the true cost per agent and per customer underneath the blend. 2. Rank customers by margin and look hard at the bottom decile. In a power-law product, the bottom decile is often deeply, not marginally, negative. 3. Segment by behavior. Find what drives the cost, long context windows, heavy retrieval, high retry rates, expensive tool calls, so you know what you're actually paying for. 4. Watch the trend, not just the snapshot. A customer drifting from positive to negative margin as they "adopt more" of your product is your single highest-priority signal: your success is making them unprofitable.
Fixing it without a clumsy price hike
The instinct is to raise prices across the board. Don't, you'll punish the profitable light users who are subsidizing everything and still bleed on the heavy ones. Negative-margin customers are a targeting problem, not a list-price problem. The better moves are surgical: * Add a usage or outcome component above an included allotment, shifting the heavy users toward paying for what they consume. * Set fair-use ceilings or throttles on the long tail, so the most extreme usage can't run unbounded. * Re-scope the expensive use case itself: route easy queries to a cheaper model (tiered routing can cut per-query cost 60 to 80%) and cache aggressively (prompt caching can cut input cost 70 to 90%). * Re-price the specific negative segment, not the whole base.
Every one of these depends on the same prerequisite: seeing per-customer margin in the first place. You cannot fix a customer you can't find.
A healthy blended AI agent gross margin is not proof you're profitable. It's proof you haven't looked closely enough. As 2025's AI pilots hit their first real renewals in 2026, the companies that survive the squeeze will be the ones watching margin per customer, catching the negative tail early, and fixing it precisely, while the ones reading only the blended number keep selling harder to the customers losing them the most money. (For the full picture of how this connects to your pricing model, start with the complete playbook on pricing AI agents.)
Paygent shows AI agent gross margin per customer, per agent, and per use case in real time, so the power user who loves your product doesn't quietly become the one bankrupting it.
