AI Unit Economics and the Margin Trap: When Your Best Customer Loses You Money

Your heaviest user is often your least profitable one. AI product gross margins ran 45% in 2025, against the 70% to 80% that classic software taught the market to expect. The cause is not model pricing. It is AI unit economics under a flat rate. Cost tracks usage while revenue does not, so the customer getting the most value is the one draining the margin.

This post is written for a founder or a director who owns the P&L on an AI feature. It picks that reader because the fix is a pricing decision, not a prompt engineering decision, and only one of those people signs it off.

The trap, stated as arithmetic

A flat-rate product charges a fixed price and absorbs variable cost. That works while the variable cost is small and bounded. Hosting a spreadsheet is both.

Model inference is neither. Every action carries a metered cost, and nothing in the product stops a determined user from taking a thousand actions a day. The price is capped. The cost is not.

Anthropic published the clearest example anyone has. When it introduced weekly rate limits in July 2025, it said some subscribers were running Claude Code continuously in the background, 24 hours a day. One user had consumed tens of thousands of dollars of model usage on a $200 plan. That is a contribution margin of roughly negative 10,000%, on one account, inside a product the market treats as a success.

Sam Altman said something similar about ChatGPT Pro in January 2025. He picked the $200 price himself, expected it to make money, and reported that OpenAI was losing money on it because people used it much more than expected. Two of the best-informed pricing teams in the industry both mispriced the same thing in the same direction.

Why blended gross margin hides it

Most teams look at one number: revenue minus cost of revenue, expressed as a percentage. That number is an average, and averages are exactly the wrong instrument here.

A book of business can report 68% gross margin while a fifth of its accounts lose money on every renewal. The light users subsidise the heavy ones, silently, until the mix shifts. Nothing in the reported percentage tells you when that shift is coming.

The instrument that works is contribution margin per account. It is not complicated, it is just not the number most dashboards were built to show. This is the same measurement gap that shows up on the cost side, covered in the piece on why inference costs sit inside AI gross margins differently to classic software costs.

What AI changed about the flat-rate bargain

Flat rate was never a gamble. It was a bet on a ceiling, and the ceiling was human attention.

A per-seat CRM could safely offer unlimited records because a salesperson can only type so much in a day. The heaviest user of a traditional tool might consume five times the median. That is a rounding error against a 75% margin.

The ratio in AI products is not five. Ramp's data on the businesses buying these products shows a 99th percentile spending about 370 times the median. Whatever the exact multiple is inside your own product, it is not five.

Agents removed the ceiling on effort

An agent works while the user sleeps. That single property breaks every assumption behind a flat rate, because consumption stops being a function of attention and becomes a function of ambition.

Cursor described the same effect from the vendor side. When it restructured its Pro plan in June 2025, it explained that new models spend more tokens per request on longer-horizon tasks, and the hardest requests cost an order of magnitude more than simple ones. Two users on identical plans can differ by 10 times on cost per request, before you count how many requests each one makes.

Multiply a 10 times cost-per-action spread by a 100 times actions-per-month spread and you have a customer base whose cost distribution spans four orders of magnitude, priced with one number. That is the trap. It is arithmetic, not mismanagement.

The contribution margin model you can copy

Here is the whole model. It fits in six columns and it will tell you more than any dashboard you currently own.

The six inputs, and the one everybody gets wrong

Five of these are easy. The sixth, the overhead multiplier, is where most models understate the problem by a third.

Your provider bill is not your feature bill. It also contains retries after timeouts, evaluation and guardrail calls, context re-sent on every turn of a conversation, and runs the user abandoned halfway. None of that is visible in a per-action price list, and all of it lands in cost of revenue.

Set r honestly by dividing last month's total provider invoice by the count of actions your product logged as successful. If that ratio is 1.4, your r is 0.4. Do not guess it.

The breakeven usage line

Rearranged, the model gives you one number worth putting on a wall. Breakeven usage is (P minus F) divided by (c multiplied by (1 + r)).

With the illustrative values above, that is $40 divided by $0.125, which is 320 actions per month. Above 320 actions, the account destroys value. Below it, the account funds the company.

Every figure in this table is derived from the illustrative inputs in the previous table, not measured from a live product. The structure is the point. Substitute your own P, F, c and r and the shape holds, only the crossover moves.

The cross-subsidy ratio

Run the model across your base and one ratio falls out. How many median accounts does it take to fund one heavy account?

With the values above, it takes six. A heavy account at minus $210 needs six median accounts at $35 each to get back to zero. That sounds survivable, and it is, while heavy accounts are 1% of the base.

Now solve for the tipping point. With those inputs the whole book goes negative once heavy accounts pass 14.3% of customers. Adoption campaigns, agent launches and successful onboarding all push that share up. Your growth plan and your margin plan are pulling the same lever in opposite directions, which is the second-order effect nobody puts on a slide.

What the usage distribution actually looks like

Before modelling anything, get the shape. Most teams have never plotted it, and the plot is the argument.

Percentiles, not averages

Ramp's figures describe buyers rather than any single vendor's seats, so treat them as directional for your own product. The shape is what transfers, and the shape is the finding.

Pull the same five percentiles for your own accounts this week. If your median account uses 40 actions and your 99th uses 2,000, your pricing page is a bet you have never priced. That is also the mechanism behind the pattern described in falling token prices arriving alongside rising bills.

One warning about definitions. Count billable actions, not sessions or seats, and count them per paying account rather than per user. An enterprise account with 200 seats and three obsessive users is a heavy account, and per-user averages will never show you that.

What vendors do once they see the curve

Every serious AI vendor has now run this arithmetic, and the record of what they did is public. The pattern is remarkably consistent, and the dates tell you when each one hit the wall.

The specifics are worth having, because they show four different instruments. Anthropic used hard caps, adding weekly limits on 28 August 2025 that it said would touch under 5% of subscribers. Figma used metered seats, announcing in December 2025 that full seats carry 3,000 to 4,250 credits a month and that limits would be enforced from 18 March 2026. Microsoft used a second currency, moving agent billing to Copilot Credits from 1 September 2025 at a pay-as-you-go rate of $0.01 per credit. GitHub used complexity multipliers, pricing extra premium requests at $0.04 each and assigning code review a multiplier of 13.

My position is that the fourth row is the only unacceptable one. Caps, credits and overage are all defensible pricing choices that a buyer can plan around. Quietly routing a paying customer to a cheaper model is a product change sold as a service level, and it is the fastest way to turn a pricing problem into a credibility problem.

The market has already moved in the same direction. ICONIQ found that among the AI builders it surveyed, consumption-based pricing rose from 35% to 42% adoption and outcome-based pricing from 18% to 23% in six months, with companies blending 1.7 models on average. That is the same conclusion reached in the analysis of why per-seat pricing stopped matching delivered value, arriving from the cost side rather than the value side.

Where this argument is weakest

Three genuine objections, and the third one is the strongest.

Costs per unit of work are falling

The same ICONIQ survey reports gross margins rising from 45% in 2025, with respondents expecting further improvement as they route most workloads to smaller or fine-tuned models and escalate only hard tasks. Two-thirds of the companies surveyed said per-query unit economics had improved.

If that continues, the breakeven line moves right every quarter and today's unprofitable account becomes next year's profitable one. A vendor that panics and caps hard in 2026 may have priced against a cost curve that no longer exists in 2027. That is a real risk and I would weight it seriously before writing caps into a contract.

The customer you would never fire

The heavy user is not a parasite. They are usually the account with the highest retention, the deepest workflow integration, the best case study and the largest expansion potential. Contribution margin at the account level is a one-period measure, and it says nothing about lifetime value.

Firing them, or capping them into leaving, can be the most expensive margin improvement you ever book. The honest version of this analysis compares contribution margin against acquisition cost and retention, not against zero. I would tolerate a negative-margin account that renews for four years and closes two referrals, and I would say so out loud in the pricing review.

There is also a survey limitation worth stating plainly. The ICONIQ margin figures are self-reported by roughly 300 mostly sub-$100 million companies, and self-reported margin definitions vary in what they put above the line. Treat 45% as directional evidence of a category, not as an audited benchmark.

If you are on the buying side of this

The same arithmetic runs in reverse, and it tells you what your renewal will look like.

Any vendor selling you unlimited AI on a flat seat is either capping it later or losing money on you now. Neither is stable. The useful question in a sales conversation is not what the price is, it is what happens when your team doubles its usage.

Ask for three things in writing. The included allowance expressed in a unit you can measure, the overage rate once it is exhausted, and the notice period before allowances change. A vendor that will commit to all three has done this model. A vendor that answers with the word unlimited has not, and you are looking at a repricing in the next 18 months.

Then instrument your own side. Most buyers cannot say which of their teams generated 80% of last month's AI actions, which makes negotiating an allowance guesswork. The discipline is the same one described in the analysis of where measurable AI return has actually appeared. It applies just as well to the build against buy decision on coding agents, where the vendor's cap becomes your ceiling.

Frequently asked questions

Where to start this week

One measurement, one decision, in that order.

The measurement takes an afternoon. Export last month's provider invoice, divide it by the count of successful actions your product logged, and you have c and r together. Then pull actions per paying account and sort descending. The top 1% of that list is your exposure, and you can read the answer off the screen.

The decision is what to do about the accounts above your breakeven line. I would not cap them first. I would call the top ten and ask what they are using it for. The question is whether you are looking at a pricing failure or an underpriced product tier that those customers would happily pay for. In my experience the second is more common than founders expect, and it is the only response to this trap that increases revenue instead of defending it.

Related reading