Usage

What your organization is spending on AI, who is spending it, and what the rest of the month is likely to cost.

The Usage page lives at /admin/usage and requires the View usage & billing permission (canViewUsage) — a sensitive permission carrying the "Owner-reserved" badge (Strata's escalation ceiling; there is no separate Owner role — built-in roles are admin and user). The built-in admin role holds it by default, so admins see this page out of the box. Drill-downs into individual users open at /admin/usage/<userId>, and into a single model at /admin/usage/model/<modelId>.

Note

Usage covers AI consumption — tokens, requests, cost attribution. Your subscription itself (plan, seats, invoices, payment method) lives on the separate Billing page, which the same canViewUsage permission unlocks read-only; changing the plan additionally requires canManageBilling.

The control bar

A sticky bar at the top of the page sets the window for every figure below it — the summary, both charts and all the tables. It was previously inside one chart's header, which read as though it only drove that chart.

• 7 days / 30 days / 90 days pills, plus an explicit date range picker. The pill row shows whichever window is active rather than sitting with nothing selected. Longer windows — 180 and 365 days — are reachable from the Key insights dropdown.
• Forecast toggle. Projects the rest of the month at the window-average daily rate, anchored on real month-to-date spend rather than a pure extrapolation. The window average divides by the days that actually have data, not the full range.

The summary

Total spend is the dominant figure, with Questions answered, People using Strata and Agent runs as supporting numbers.

A further figure appears only while someone is actually being throttled — it is absent rather than zero the rest of the time, which for most organizations is most of the time. It reads "People are at their hourly limit", or "Person is at their hourly limit" when it is exactly one, above a Review link that jumps to the All users table. Throttled people are shown in red there: "Users in red are currently at their hourly limit."

Every other figure carries a period-over-period comparison that names the prior absolute value — for example "↑ 2130% from $5.20 in the previous 30 days".

Note

These comparisons are deliberately colourless. Higher spend is not a failure and lower spend is not a win: a drop can mean the team stopped using what you are paying for. The number tells you what changed; only you know whether that is good.

Where the money goes

Spend over time is a single trend chart with two views:

• By surface — splits spend across chat, scheduled agents, the Outlook add-in and the support assistant.
• By model — splits by the model that produced the spend. Bar and line rendering apply to this view.

The time axis plots every day in the window, including days with no activity, so a quiet week is visible as a gap rather than being closed up. When the last day in the window is today, the chart says "Today is still counting" — today is a part day, and undecorated it reads as a collapse in spend.

Below the chart, the Spend by model table breaks the same window down per model.

Note

A charge below one cent renders as <$0.01, not $0.00. A model that billed millions of tokens and a model nobody called should not print the same string.

Spend by model (detail)

Below the user table is a model-level breakdown — one row per model your organization has actually used in the window, ordered by cost. Each row carries the model's name, provider, and logo, with Uses, Tokens, and Cost columns. The rows are whichever models from your Model Catalog generated activity, so the list reflects the models you have enabled rather than a fixed lineup.

Click any model row to open /admin/usage/model/<modelId> — a per-model drill-down with the model's totals, a per-user split, a daily series, and its input/output rate card.

Note: agent model spend rolls up under each agent's owner in the All users table, not into this table. The Spend by model section covers per-model platform AI spend excluding scheduled-agent runs — it includes regular chat plus non-chat auxiliary AI such as Knowledge Base embeddings, the vision document-review chain (PDF/DOCX/XLSX/PPTX), image OCR, the in-app support assistant, the Strata for Outlook add-in, and the assist helpers.

Feature labels

Auxiliary AI calls are labeled by feature with a model-id suffix, so platform helpers are distinguishable from regular chat:

The assist helpers under these labels include chat-title generation, agent-capability suggestions, writing-voice distill/rewrite, announcement draft/edit, memory import, and analytics categorization. They run on the cheapest small chat model your organization has enabled (for example gpt-5.4-mini or another economical model from your catalog), so the row they appear under reflects your org's own model set.

The Strata for Outlook add-in spans several base models. Rather than scattering them, all -outlook rows are collapsed into a single Strata for Outlook row whose drill-down breaks the spend down by underlying model.

People and seats

All users table

Ranked by total spend in the window. Columns:

• User — display name + email.
• Chat queries — count of chat requests.
• Agent runs — count of scheduled agent runs.
• Chat cost — dollars spent on chat tokens.
• Agent cost — dollars spent on agent tokens.
• Total — sum, with a sort arrow on this column by default.

Every column header is sortable. Users currently at their hourly limit have their name rendered in red (hover for how much of their cap they have used). A search field above the table filters by name or email; an export button to the right of search downloads the current view as CSV.

Click any row to open /admin/usage/<userId> — the per-user billing profile.

Per-user billing profile

The profile page shows every metric for one user in detail:

• Their cumulative spend over the window.
• A daily breakdown of their queries and runs.
• A per-agent table showing how each of their scheduled agents contributes to spend.
• Their effective rate limit (from the cascade) and current usage against it.

Use this when investigating a heavy user, building a chargeback report, or troubleshooting why someone's agent is more expensive than expected.

What you're getting for it

Time saved, compute credits and key insights sit together at the foot of the page.

Time saved is modeled, not measured, and the card says so: the deliverable counts are real records, but the minutes attached to each kind are your organization's own assumptions. The subtitle lists whichever assumptions are in play for the rows shown, so it reads differently for different organizations. Treat the hours as "what our stated assumptions imply", not as an observation — it is the one figure on this page that is not a direct record of something that happened.

Compute credits

Below the model breakdown is a compute-credits card that tracks credit consumption for the selected month. Credits are Strata's internal unit for metering finished work — agent runs, generated documents, Analysis runs — scaled by the size of the job rather than counted by token.

A month pager at the top of the card moves between months; the card always reflects one calendar month at a time. The card shows:

• Credits consumed this month — the headline total for the selected month.
• By type — the total split across the kinds of activity that draw credits.
• By user — the same total attributed to people in your org, so you can see who generated the most credit activity.
• Recent events — a spot-check list of the latest individual credit events, for confirming the totals against real activity.
• Reconciliation — a view comparing the recorded credit ledger against the actual completed activity for the month, so any drift between what was metered and what ran is visible.

How usage is counted

This describes how credit activity is counted, not what you are charged:

Note

Size-weighted counting is how credits are recorded once weighted metering is switched on for your environment. Until then, credit events record flat — one per event — and the card labels the total indicative rather than billable ("metering off — counts, not weights"). If your totals read like plain event counts rather than the weights below, that's why.

Once weighted metering is on, credits are fractional and scale with the size of what Strata actually produced:

The rule underneath the table: you are billed for output, not for effort. A question that takes several tool calls and minutes of reading to answer well draws nothing; only a deliverable draws credit. Batching works in your favour either way — flagging forty emails in one request is a single credit event, not forty.

These do not draw credit:

• Work that failed, was cut off by a spend cap, or was rejected in review.
• Strata's own quality-control pass. Generated PDFs, Word files and complex workbooks are rendered and inspected, and the regeneration that fixes a broken one draws no credits. The inspection is itself model work: it runs in your own Azure AI Foundry like any other model call, so Microsoft bills you for that inference directly.
• Re-running the same task after a failure.
• A SQL question answered inline in chat. Persisting it — a saved dashboard, a large export — is a deliverable and does draw credit.

Note

The compute-credits card is a usage and reconciliation view — it measures credit activity, it is not a bill. Your subscription lives on the separate Billing page.

Key insights

A summary block underneath the model table. Computed from the same dataset as the rest of the page — picking a longer range from the dropdown there reloads the page at the new window.

How costs are calculated

Strata uses a per-model cost reference to translate token counts to dollars. Rates come live from your organization's Model Catalog — the input/output price per 1M tokens recorded on each model — so when a platform owner edits a model's pricing, cost reporting follows. A static table mirroring the catalog is used only during the brief cold-boot window before the live catalog loads.

The rates are visible from the i button next to "Rate limit" inputs anywhere in the admin console. The AI model costs modal lists the models your organization has enabled with their input and output token prices.

The displayed totals are estimates. Real bills can vary because:

• Prompt-cache discounts — up to ~90% off for cached prompt tokens.
• Batch-API discounts — 50% off for batched requests.
• Long-context tier surcharges — extra cost for queries in the larger context windows.
• Negotiated EA/MCA rates — your Microsoft subscription may carry custom rates.

The footnote at the bottom of the page reads: "Costs are estimates. Token totals come from the AI provider; per-token rates are pulled from Microsoft Azure AI Foundry's published pricing and don't account for prompt-cache discounts (~90% off cached tokens), batch-API discounts (50% off), longer-context-tier surcharges, or any negotiated EA/MCA rates on your subscription."

Relationship to your Foundry billing

AI inference runs in your organization's own Azure AI Foundry, in your own Azure tenant.

The token spend shown here is what you can expect on your own Azure subscription's Foundry bill (see Bring your own Azure AI Foundry). Use the per-user table to attribute that spend to people in your org. The exact dollar figure on your Azure invoice may differ from the estimate here for the discount reasons listed above.

What can go wrong

See Rate limits and quotas for the cap that produces the "at limit" indicator, and Bring your own Azure AI Foundry for the billing relationship with your own Foundry.

Related

• Rate limits and quotas
• Bring your own Azure AI Foundry
• Roles and permissions