Real Per-Generation AI Cost Tracking Automation: A Call Is Not a Cost

A usage counter answers "how many times did this happen" — a genuinely useful number for enforcing a simple limit, and one that is comparatively easy to build correctly the first time. A cost figure answers a different question entirely: "what did this actually consume," which for a system built on a variable-cost resource (an AI model, billed by the specific number of tokens processed rather than by the number of requests made) can vary by several multiples between two individual instances of the identical counted event. Building the second question's answer out of the first question's counter — assuming, implicitly, that every call costs roughly the same — produces a number that looks precise and specific while actually being closer to a rough guess, discoverable as wrong at precisely the moment someone checks it against a real bill.

A raw count of events (calls, requests, generations) is the simplest, cheapest signal to capture and enforce a usage limit with, and it is entirely reasonable to build that first, especially under real time pressure to ship a working quota system at all. The gap this specific finding closes is between "a working quota" and "an accurate cost figure" — two related but genuinely different capabilities that can look, from the outside, like the same feature already built. The AI provider's own usage metadata sitting unread on every response is a specific, concrete symptom of exactly this gap: the real, precise information needed to close it was already arriving, for free, the entire time — the missing piece was never data availability, only the decision to read and record what was already there.

Centriu Helix's own usage-tracking table counted AI generation CALLS, and the module's usage quota — a fixed ceiling of 100 calls per day and 2,000 per month — was built entirely on that same call count. A call, however, is not a cost: measured directly the same day this was fixed, one AI generation averages roughly 1,783 input tokens and 2,021 output tokens, with the gap between a short, simple brief and a context-heavy, single-topic deep-dive spanning several multiples of that average. Asked what an AI-active client genuinely costs per month, the honest answer, before this fix, was an estimate — a number whose accuracy nobody could actually verify against real usage.

The underlying AI provider returns detailed usage metadata — precise input and output token counts — on every single response it sends, and nothing in Centriu Helix had ever read that information. The fix captures those real per-generation token counts and records them directly on the same database row that already stores the report each generation produced, so the cost figure lives exactly where the content it paid for lives. A dedicated function then sums real cost by organization and by month, taking price-per-million-tokens as an explicit PARAMETER passed into the calculation rather than a fixed number hardcoded inside it — deliberately, because a price baked directly into a database view becomes a silently stale, hidden number the moment the real price changes, with nothing prompting anyone to notice or update it.

Generations that happened before this change correctly show their token counts as NULL, meaning "before real measurement began," and are deliberately never treated as zero-cost: automatically zeroing out real historical usage that was simply never measured would make every calculated average across time lie systematically low, understating true cost rather than honestly acknowledging a measurement gap in the historical record.

Two closely related, independently real defects were found and fixed in the same change. First: a rate-limit response from the AI provider itself (a condition Google's own API reports as "RESOURCE_EXHAUSTED," which clears on its own within minutes and needs nobody's attention) had produced the exact same generic failure message as a genuine error requiring a person to investigate — two structurally opposite situations, indistinguishable to whoever encountered either one, and the rate-limit case is specifically the signal that appears exactly when the product is being used successfully by multiple organizations at the same moment. The fix responds to a rate-limit specifically, stating explicitly that no usage quota was actually deducted for the attempt — without that explicit statement, someone hitting this condition would reasonably believe they had lost a generation from their quota and might simply stop trying rather than retry moments later.

Second: the generation log had always recorded the PRIMARY AI model as the source of every single answer, even on the runs where the system's own fallback model had actually produced the response after the primary either failed or was unavailable. That exact history — which model actually wrote which report — is the same history later used to judge both output quality and real cost per model, meaning a silently incorrect model attribution was quietly corrupting the accuracy of both of those downstream judgments.

Measured directly in production, after this fix deployed: four simultaneous AI generations, run using the identical organization's own access credentials, produced four successes in 25 seconds of total wall-clock time — the same latency as a single generation running by itself, direct confirmation that concurrent generations are not silently queued and processed one at a time behind the scenes.

Before the fix, asked what a specific client's AI usage cost over the past month, the only honest answer available was a rough estimate derived from a raw call count, with no way to distinguish a month of short briefs from a month of context-heavy deep-dives. After the fix, the same question is answered from real, per-generation token counts already captured and summed, at whatever current price-per-million-tokens is passed into the calculation — a number that could actually be checked against a real invoice.

Centriu Helix now records real, per-generation AI token usage and calculates genuine, parameterized cost by organization and by month, closing a gap where usage quota (based on a raw call count) was the only signal available and a real cost figure required guessing — alongside two related fixes distinguishing a self-clearing rate-limit from a genuine failure, and correctly crediting whichever AI model actually produced each response.

This automation covers specifically how Centriu Helix measures and calculates the real cost of its own AI generation activity — it does not change the module's usage QUOTA logic itself (still based on call count, which remains a reasonable, simple fairness limit even though it is not a cost measure), and it does not change what any individual AI-generated report actually contains. A team looking specifically for how AI-generated content itself is produced, rather than how its cost is measured afterward, should see this pillar's separate companion pages on the per-topic deep-dive generation architecture and the shared report shell.

Using a raw call count as a stand-in for cost is simple to build and reasonably fair as a USAGE LIMIT, since it treats every action a person takes as roughly equivalent regardless of what it specifically asked for. The moment the actual underlying resource varies significantly in cost per call — which token-based AI generation genuinely does, by several real multiples between a short and a long request — that same simplicity becomes actively misleading the moment anyone asks a genuine cost question rather than a fairness question. Capturing the real, already-available usage metadata the provider sends on every response, and calculating cost from that instead of from the call count, costs almost nothing extra to build (the data was already arriving) and turns a permanent estimate into a number that can be checked against reality.