P90 is AI Pricing Operations for commercial AI products. It shows how retries, larger contexts, retrieval, concurrency, and long-running agent workflows create cost variance beyond the average, then detects, simulates, and recommends a response for human approval.
A human approves every recommendation
Observe the cost distribution. Decide what to change. Govern the decision after it leaves the dashboard.
P90 fits commercial AI products where cost behavior materially affects margin, pricing, or enterprise deal economics.
Connect usage behavior to plans, cohorts, customer value, and the trade-offs behind a pricing decision.
Trace cost attribution through retries, context, retrieval, model routes, orchestration, and infrastructure causes.
Work from fully loaded COGS, gross margin, pricing floors, forecasts, and contract economics using the same evidence.
“AI systems don't fail on averages. They fail when real usage creates variance.”
A continuous four-node loop that watches your real cost economics and acts a response when variance threatens your margins.
Watches 6 variance signals on a continuous loop. Monitors p90 and behavioral concentration, not averages. Advances the pipeline when thresholds are crossed.
Determines which of the 7 cost layers is responsible. Not just 'cost is up': it traces the source layer and severity before proceeding.
Auto-selects and runs the relevant stress scenario. Projects 30-day margin impact under worst-case conditions, not typical behavior.
Reviews the anomaly and simulation context, then returns 3 ranked pricing options grounded in the 7-layer cost framework, each with its own evidence.
Cost emerges across all seven layers. Each layer can add direct work or increase the request size, repetition, capacity, or validation required to serve the product.
Per-seat and flat-rate pricing assume predictable, uniform usage. AI products are neither. These are the principles p90 operates on.
AI systems fail at the p90 and p95, not the average. The gap between your median user's cost and your most expensive user's cost determines whether pricing survives.
Most teams obsess over average cost per request. That's the wrong metric.
Two users paying the same price can cost very differently to serve. Pricing that treats them identically can subsidize expensive usage at the cost of lighter usage.
Usage correlates with cost volatility, not guaranteed value.
Tokens are the visible fraction of AI cost. Retrieval, orchestration, retries, and concurrency are not. Some layers add direct work while others amplify request size, repeated execution, capacity pressure, or downstream validation.
This is how AI costs move through the stack without any single decision looking wrong.
If pricing only works when users behave nicely, it does not work. The agent simulates adversarial scenarios by default: not the median, the tail.
Price for your worst reasonable day, not your average day.
The same feature that was cheap at launch gets more expensive as users learn to push it harder. Pricing set once and left alone drifts underwater without anyone touching it.
Costs move after launch. Most pricing never does.
The lowest-paying and free users are often the ones generating the highest variance. A generous tier set without modeling worst-case behavior subsidizes exactly the usage you can least afford.
Your cheapest plan can carry your most expensive users.
P90 keeps an append-only record of the evidence behind a pricing decision: the economic trigger, alternatives considered, human approval, implementation history, and measured outcome when available.
Every recommendation carries structured Decision Evidence, the alternatives it weighed, and the economic anomaly that triggered it, not just a conclusion.
No downstream pricing change is claimed from an approval alone. Every approval is timestamped and attributed to the person or policy that recorded it.
An append-only decision log with exportable approvals, alternatives, timestamps, and supporting evidence for internal review, board diligence, and investor scrutiny.
Every loop produces the same chain of work (sensing, attributing, stress-testing, and recommending) so margins never drift in the dark.
Tracks p90 cost, retry rates, and behavioral concentration in real time, so no one has to open a tab.
Pinpoints which of the seven cost layers is moving, with a severity score, before anything escalates.
Projects 30-day margin impact under adversarial usage: the tail, not the comfortable median.
Generates three grounded pricing options with rationale, ready for a human to weigh and approve.
Nothing ships automatically. A person signs off on each recommendation before it counts.
A dashboard waits for a human. p90 closes the loop: detection, simulation, and evidence preparation happen without anyone opening a tab.
Launch the dashboard, activate a simulator scenario, and watch the full loop run, from anomaly detection to pricing recommendation to approval.
View live demoNo setup required in demo mode · Data is simulated