Understand AI usage and cost totals
Documentation for Tutela v3.9.0.
Use AI Spend to see recorded requests, tokens and estimated costs, then identify missing prices or attribution. The totals cover traffic reported to Tutela; they are not a bill for every AI tool used in your organization.
Prerequisites
- Permission to view AI Spend.
- At least one connected source that reports usage, or imported billing evidence for billed costs.
- A pricing entry for each model whose usage you want to include in estimated spend.
Review the totals
- Open AI Spend and select the period you want to review. Note its start and end and the latest recorded usage time.
- Check whether the summary loaded successfully. If it is unavailable or a refresh failed, retry before exporting or making a budget decision from the figures. An older usage record is not, by itself, a collection failure.
- Compare requests, tokens, estimated spend and unpriced usage. Use the definitions below to distinguish an actual zero from missing information.
- Open the provider, model, identity or cost-group breakdown that answers your question. Keep the same period while comparing it with the summary.
- Investigate unmatched usage before assigning costs. Use View human actors and View cost groups for their corresponding breakdowns.
What the numbers mean
| Measure | Meaning | Common mistake |
|---|---|---|
| Requests | Recorded usage events within the selected period | Treating this as the number of inspection decisions or tool invocations |
| Tokens | Recorded input and output tokens | Assuming an unreported token count is zero |
| Estimated spend | Priced usage valued using its recorded pricing information | Treating this estimate as the provider's bill |
| Billed spend | Cost supported by imported billing evidence | Expecting a billed value before an import exists |
| Pricing coverage | Priced requests divided by recorded requests | Assuming excluded unpriced usage costs nothing |
| Identity attribution | Requests with a recorded human or service identity divided by recorded requests | Assuming every service identity identifies a person |
| Cost-group allocation | Requests assigned to a cost group divided by recorded requests | Treating an identified request as automatically assigned to a cost group |
For example, all requests can have an identity while none have a cost-group assignment: 100% identity attribution and 0% allocation are compatible. Neither percentage measures how much of your organization is protected.
The selected period includes its start and excludes its end. Repeated delivery of the same organization, event time and event identifier is counted once. The same identifier recorded at a different event time is a separate usage record.
Latest usage in this period is the time of a recorded request, not the last successful collection check. Older records are expected when reviewing a past period or when no new usage has been reported. If you expected newer activity, check the period first, then ask your administrator to trace the missing request through the connected source. An empty period does not prove that every source is working.
Read spend and token history
Use the Spend / Tokens selector to compare one measurement at a time. Short periods use hourly observations; longer periods use daily observations, with UTC bucket labels. View trend data exposes exact values and availability. Amber points show known cost with incomplete pricing. Unpriced usage can still have recorded tokens. Missing observations are not interpolated as zero and are not joined into a continuous line.
Provider and model bars rank recorded estimated spend while the tables retain exact cost, tokens and requests. Partial pricing is labeled; a zero-cost record has no artificial minimum bar. The server's Other row remains visible where the returned groups omit smaller categories. A larger bar indicates more known cost, not greater risk or more complete protection.
In the 3.9 stable line, small spend-axis labels use enough precision to distinguish adjacent ticks. Use View trend data for retained values; display precision does not improve missing pricing evidence. Verify the installed build before expecting the corrected labels and pricing guidance.
Read breakdowns and exports
Each breakdown groups the same usage population by one dimension. Do not add provider, model, project, user, department and cost-group totals together. Other combines omitted rows rather than discarding their usage.
Human actor not supplied and Human identity unresolved describe different missing identity states. Expand a recorded identifier when available; do not assign the traffic to yourself merely because you are viewing it.
Exports contain the visible breakdown rows, aggregated remainders, selected period and scope. They are not individual request evidence. Older recorded values can be exported when the summary and breakdowns loaded successfully. A failed read or refresh must be resolved before exporting the affected data.
Resolve missing information
- Unpriced usage: identify the provider and routed deployment, then check their explicit pricing configuration with the administrator. A returned model name alone does not establish a compatible endpoint's billing rate. Recheck coverage with a new request after correction; historical gaps require retained pricing evidence.
- Unattributed or unallocated usage: inspect the source's identity and organizational assignment. Ask the operator to distinguish an allocation lookup failure from a successful lookup with no membership. Keep unresolved usage unallocated; current memberships alone cannot establish historical assignments.
- No recent events: check whether the source was active. Usage telemetry uses the last 24 hours, while usage totals use your selected period.
- Unavailable count: retry the failed query. A missing denominator cannot establish zero usage or complete coverage.
Validate
- The summary and breakdown use the intended period and loaded successfully.
- Estimated, billed and unpriced amounts are distinguished.
- Missing identity and allocation are documented rather than guessed.
- Exported totals match the visible period and breakdown.
Next steps
- Check a single request's usage reporting.
- Investigate missing or unexpected costs.
- Test a budget before enforcing it.
Read the summaries
In v3.9, the workspace keeps spend, tokens or imported billed cost, and pricing coverage in three compact summaries. Identity attribution and cost-group allocation remain separately labeled below them, including unavailable counts and unallocated requests. Use Period and latest usage for exact time bounds and the source timestamp. Expand View trend data or a breakdown's Exact values for recorded numbers; gaps are unavailable observations, not zeroes. Help opens this guide and Usage telemetry keeps the selected date range.
Estimates and optional provider-reported costs
Token-based estimates remain the default. They use the usage Tutela observed and the applicable recorded model price, not an assumption that all account spending passed through Tutela. Missing usage or pricing is a limitation, not confirmed zero.
An administrator may supply an OpenAI or Anthropic billing admin key for a one-time provider cost check in AI Provider settings. The credential is not persisted. The report covers the last seven complete UTC days for the accessible provider account, including activity outside this workspace; it does not import invoices or reconcile the workspace ledger. Reporting can lag and differ from the final invoice. A failed check does not stop token estimates or AI Chat. Azure, AWS and Google billing connectors are not included in this optional check.