Your llm cost per user is not what your provider dashboard says. It is the invoice at the end of the month divided by the number of people who actually finished something useful. Everyone publishes the token-price table. Nobody publishes what a shipped consumer AI product costs per user at volume.
We can. Here is the real number from Parcoursup Zen (/work/parcoursup-zen), the French-language orientation platform we built: under €4 average AI cost per student, across 12,000+ students onboarded in season one.
That is the whole pipeline per student, not one call. It includes the profile-to-formations matching against a corpus of 23,000+ programmes, the deadline tracking, the motivation-letter drafting, and the orientation chatbot—all of it. And the number is not an accident of scale or a lucky model choice. It is an architectural outcome.
The Number That Actually Matters: llm cost per user
A bare per-token cost tells you nothing about whether your product is viable. The question is what you pay for a completed journey.
| Metric | Typical AI-SaaS | Parcoursup Zen |
|---|---|---|
| Unit of cost accounting | Per API call | Per completed student flow |
| Cost-visibility problem | Heavy power-user tail invisible until invoiced | Bounded, structured flow caps per-user ceiling |
| Completion rate | Often unmeasured or gamed | 87% guided-flow completion |
| Average user rating | — | 4.8★ |
| Average AI cost per user | Variable, tail-driven | < €4 (pipeline total) |
An 87% completion rate is not a vanity number; it is an economic one. Abandoned sessions are the cheapest sessions. If you optimise for cost-per-session, you are optimising for failure. What we measured—and what any team shipping an AI feature should measure—is cost per finished journey. That is what the €4 buys.
The stack that produced this number is unexotic: Next.js, LangChain and LangGraph for orchestration, OpenAI models, and Stripe for the paid tier. No fine-tuned open-source model on a custom inference cluster. The cost control does not come from model choice; it comes from flow design (AI services and architecture).
Why a Bounded Flow Avoids the Margin Cliff
Published guidance from the AI-SaaS space keeps flagging the same danger point: on a flat-rate plan, a tiny slice of power users drives the majority of inference spend. The top 1–2% of users often account for 40–50% of total LLM cost. That tail makes unit economics unpredictable month to month and turns growth into a margin problem.
A structured, guided flow is the exact shape that avoids that tail. Parcoursup Zen is seasonal and scoped: a student goes through a known set of steps with a bounded corpus of 23,000+ formations. There is no open-ended prompt box where a power user can chain a hundred-agent debate and burn €3 in a single session. The architecture enforces a cost ceiling per user, and that ceiling is what makes the <€4 number stable across the cohort.
This is not a generalisation to every product. Do not take the €4 figure and plug it into your own model. The transferable lesson is the method:
- Design the flow to be bounded, not open-ended.
- Measure cost per completed journey, not per API call.
- Price the expensive human-reviewed tier separately, instead of averaging it into everyone's plan.
The Hidden Cost Line No Budget Spreadsheet Captures
Every prompt change is a liability until it passes an eval suite. For Parcoursup Zen, that suite is 240 cases, and it gates every prompt update. The cost of that gating is real and frequently omitted from AI-feature budgets:
- Inference cost of the eval runs themselves. Every prompt candidate gets run against 240 scenarios. That is token spend with zero user-facing output.
- Human time to curate the cases. The 240 cases are not auto-generated. They were built, reviewed, and maintained by someone who understands French post-bac orientation. That is a recurring operational cost, not a one-off setup.
The payoff is that a "cheaper prompt" cannot silently degrade recommendation quality. Without an eval gate, cost-optimisation becomes a drift toward mediocrity. With it, you pay a known, fixed QA tax on every change, and that tax is what protects the 4.8★ rating.
The Commercial Structure That Solves the Margin Problem
The premium tier in Parcoursup Zen is the actual answer to the AI-feature margin question. There is a free AI-guided tier that carries the predictable <€4 per-student cost. Separately, there is a paid tier where a human reviews the output—for parents who want that reassurance (Parcoursup Zen case study).
This is not a new commercial structure, but it is the right one for AI features. The AI cost is low, fixed, and per-user-capped, so the free tier's unit economics are known in advance. The expensive part—the human review—is priced separately rather than amortised across every user's plan. Most AI features hit a margin wall because they bundle the variable cost and pray usage is low. Unbundling the high-cost tier is the fix.
If you are pricing your own AI feature, start with a free LLM API pricing calculator to model your per-call cost at projected volume. Then multiply by worst-case flow length, not average. That is your ceiling. Then ask whether your pricing covers that ceiling for every user. If it does not, you have not an AI problem but a packaging problem.
FAQ
How do I measure llm cost per user accurately?
Tag every LLM call with user ID, feature, and workflow. The metric that matters is total LLM spend per user over a billing period, divided by the number of users who completed the core action—not just logged in. Provider dashboards show aggregate spend; they do not show the per-user distribution or the p95 tail.
Is €4 per user a good benchmark for AI features?
No. It is a real number from a specific product shape—seasonal, structured, guided, bounded corpus, French-language. A general-purpose chatbot with unbounded input will look nothing like this. The benchmark is not the number; it is the method of capping per-user cost through flow design and measuring per completed journey.
What is the biggest mistake when pricing an AI feature?
Bundling high-cost human review into a flat-rate plan. The small share of users who need it will blow up your unit economics for everyone else. Price the expensive tier separately, as Parcoursup Zen does with the parent-reviewed option. That keeps the AI tier's margin predictable and lets the premium tier carry its own cost.
Top comments (0)