AI consumption

Track the company's AI usage (tokens/cost) by period, user and feature. The plan defines the monthly cap; with BYOK (your own key), consumption runs on your provider account.

The period's consumption, against the plan's cap.

In the platform view, the "In-house cost (provider)" and "Month's margin" cards separate what IS our cost from what runs on the customer's own key (BYOK): the gross spend of those using their own key comes out of THEIR account and therefore stays out of the in-house cost. The "Platform — internal use" line gathers translation, public help, evaluation and onboarding — consumption that doesn't belong to any company.

The "Consumption by company" table is the master: click a row and, right below, you see that company's consumption by person and the entries one by one — date/time, user, screen, model, tokens, cost and call status. It's the way to investigate an out-of-line month: from the card's total down to the exact call that produced it. The period comes from the Month/Year filter in the side panel, and combinations you repeat can be saved as Views, per user.

Each interaction with Tenkai is worth a fixed number of credits based on the type of request — 1 to choose a tool, 2 for a chat question, 5 for an analysis or summary, 10 for an agent step, 20 to generate a document. A call that didn't answer (block, error) is worth 0. That's how the market charges (Copilot, GitHub, QuickSight), and it's what lets you compare months and people without talking about tokens. The cost in reais is still in the column next to it, for those allowed to see amounts.

Tenkai is measured in credits. Each interaction is worth a fixed amount: question and answer 2, fill-in suggestion 2, summary 5, analysis 5, document extraction 5, agent step 10, generated document 20. Your plan has a monthly credit allowance that refills every month and does not accumulate. Beyond the allowance, each extra credit uses the account wallet, which you top up by buying packs under Settings ▸ AI; purchased credit arrives when the invoice is paid and does not expire while the contract is active. When allowance and wallet run out, Tenkai warns and stops until the month turns or a new purchase. Ask Tenkai "how many credits do I have?" to see allowance, usage and balance right away.

In the By answer list, the Model column tells which AI model wrote that answer. “No AI” means the answer came straight from the database, without writing — that is information too: that question cost no AI.

Once the plan's AI cap is reached, calls get blocked until the next cycle (or an upgrade). Consumption is measured per call, with estimated cost per model.

When you select a company, the "By person" table shows who used AI in the period. A daily watcher compares each person's consumption with their own 14-day average: if it exceeds the factor and the floor (parameters IA.ANOMALIA_*), an operational warning goes out. Legitimate spikes happen — closing, data loads — so the warning asks you to check here before drawing conclusions.

With automatic routing turned on (Settings ▸ AI ▸ Parameters), short questions go to the cheapest model among those the account has enabled and document writing goes to the strongest; the rest stays on the default model. The "routing savings" scoreboard, in Settings ▸ AI, shows how much the account avoided spending in the month because of it — the same tokens, at the default model's price, minus what they actually cost.

The last-30-days block shows questions answered, how many had no data, positive and negative ratings, median and p95 response time, cost charged and the amount recovered by the digital employees. Zeros are declared: a month without use shows zero, not blank. Ask Tenkai "how much did Tenkai cost and earn me this month?".

Para o que consome IA de verdade (perguntas, análises), sim, até o próximo ciclo ou até ampliar o teto; o que é regra ou busca continua. O vigia avisa quem administra antes de chegar lá, e o consumo por pessoa mostra onde o custo está.