Usage & ROI Reports

Bifrost ships two complementary reporting surfaces:

• Usage Reports (/usage) — what the platform actually consumed: AI tokens, AI dollars, CPU seconds, peak RSS, knowledge storage. Useful for cost conversations.
• ROI Reports (/reports/roi) — how much human time and dollar value the platform saved. Useful for value conversations.

Both surfaces are visible to platform admins for any org; org admins see only their own org’s slice.

Usage Reports

Navigate to Reports → Usage Reports, or go directly to /usage.

Filters

The Report Period card at the top controls everything below it:

• Date range — defaults to last 30 days. The trends chart respects this range; the workflow / conversation / agent tables sum across it.
• Source tabs — All, Executions, Chat, Agents. Toggling source hides irrelevant tables (e.g. Chat hides the workflow table) so you can focus on one cost driver at a time.
• Organization (platform admins) — filter to a single org or look across all orgs at once. Choosing All organizations also surfaces a By Organization table at the bottom.

Summary cards

Across the top: Total AI cost, Input tokens, Output tokens, AI calls, Total CPU seconds, Peak memory. The cards reflect the active filters.

Trends chart

A daily line chart of AI cost and token usage across the date range. Use it to spot:

• A sudden cost spike on a single day (usually a runaway agent or a misconfigured workflow firing on every webhook)
• A weekly weekend dip (normal — the chart is in UTC, so weekend = expected)
• A gradual climb (capacity planning signal)

By-workflow / By-conversation / By-agent tables

Each table sorts by AI cost descending. The AI cost column is broken out per model — you’ll see Claude Opus, Sonnet, GPT-4o, etc. listed separately for the same workflow if it called more than one provider during the date range. This makes it obvious when a workflow is paying Opus prices for a Haiku-class job.

Provider pricing is auto-synced — Bifrost periodically refreshes its model price table from a central source, so cost numbers reflect current per-token rates without operator intervention. There is no manual step to keep pricing accurate.

By-organization table

Visible only when Organization is set to All organizations. Sorts orgs by their AI spend so you can identify the heavy hitters across a multi-tenant deployment.

Knowledge storage table

Shows S3 / vector-DB bytes consumed per knowledge collection. Useful when storage costs become a line item.

Demo Mode

Platform admins see a Show Demo Data toggle in the header. Flipping it on replaces the real data with deterministic sample data — handy for screenshots or demos when your real data is sparse or sensitive. The amber banner makes it obvious you’re in demo mode.

ROI Reports

Navigate to Reports → ROI, or go directly to /reports/roi.

ROI is configured per workflow at authoring time:

@workflow(
    roi_minutes_saved=15,
    roi_dollar_value=25,
)
async def triage_ticket():
    ...

Each successful execution multiplies its workflow’s roi_minutes_saved and roi_dollar_value into the totals you see on the report. The page surfaces:

• Total time saved (hours)
• Total dollar value
• Trend chart — daily roll-up
• By workflow table — sortable by executions, time, or value
• By organization table — when scoped to all orgs
• Export — download the current view as CSV

Note

ROI numbers are only as accurate as the per-workflow estimates. Treat them as a direction-of-travel signal, not a forensic accounting record. A workflow that runs but produces a no-op result still bumps the counters — you’re responsible for marking it success or failure correctly inside the workflow body.

Where agent costs land

Autonomous agent runs generate AIUsage rows that roll into both reports:

• Usage Reports → the By Agent table groups by agent_id, with per-model breakdown.
• The Spend (7d) card on each agent’s detail page (/agents/:id) reads from AgentStats.total_cost_7d, which is the same AIUsage data summed over the trailing 7 days. Backfilling summaries on N runs (via POST /api/agent-runs/backfill-summaries) will move that card by roughly N × (avg summarizer cost per run) — the API surfaces a cost estimate before you confirm.

Common questions

“Why is this workflow’s AI cost so high?” Open Usage Reports → set Source to Executions → expand the workflow row to see the per-model breakdown. If 90% of the cost is on Opus calls, the workflow is over-modeling — refactor it to route easy paths to a cheaper model.

“How much did we spend last month?” Set the date range to the previous month, leave Source on All, organization on All organizations. The Total AI cost summary card is your number.

“Is the platform paying for itself?” Compare Total AI cost in Usage Reports against Total dollar value in ROI Reports for the same date range. A 10× or better ratio is the typical signal that automation is earning its keep.

See also

• Diagnostics — CPU and memory at the worker level
• Audit Log — control-plane events
• Decorators reference — roi_minutes_saved, roi_dollar_value, category