Skip to main content

Cost Attribution & Intelligence

The Cost Attribution page (/analytics/costs) helps security, platform, and CIO/FinOps stakeholders understand where AI spend is concentrating across the fleet. The Cost Intelligence tab adds budgets, forecasting, and optimization recommendations.

Current readiness

AreaStatusNotes
Cost views by agent/team/model/providerLiveThe shipped page supports the core “who is expensive?” workflow.
Trend and optimization reviewLiveUse the page to find concentration by model or provider before changing policies or routing.
Multi-scope budgetsLiveSession, agent daily, team daily, and org monthly budgets with auto-reset at period boundaries.
Cost forecastingLive30-day linear regression forecast with trend direction and confidence level.
Optimization recommendationsLiveModel downgrade candidates with estimated savings.
Department rollups / chargeback-grade exportPlannedThe current page is analytical, not a full billing/chargeback workspace.

What the page is for

Use Cost Attribution and Intelligence to answer questions such as:

  • Which agents are driving the highest LLM spend?
  • Which teams are relying most heavily on premium models?
  • Which provider/model combinations are costing more than expected?
  • Am I on track to stay within my monthly budget?
  • Which agents could save money by switching to cheaper models?

Typical workflow

  1. Group or filter the table by the dimension you care about.
  2. Identify the highest-cost agents, teams, models, or providers.
  3. Review budget utilization bars to check budget health.
  4. Check the forecast to see projected month-end spend.
  5. Review optimization recommendations for model downgrade opportunities.
  6. Pivot into Agents, Policies, or operational reviews to decide whether the spend is justified.

Cost Intelligence features

Budgets

Set budgets at four scopes: session, agent daily, team daily, and org monthly. Budgets auto-reset at their period boundary (midnight UTC for daily, first of month for monthly). Configure an alert threshold percentage to get early warnings.

Forecasting

The backend runs ordinary least squares (OLS) linear regression on daily cost data to project 30-day spend. The dashboard shows trend direction (increasing / stable / decreasing) and confidence level.

Optimization recommendations

The system identifies agents using expensive models and suggests cheaper alternatives with estimated savings percentages. Recommendations are advisory only.

Internal vs customer cost

The customer_cost_usd() metric excludes Pavri's internal inference costs (judge evaluations, session summaries) so operators see only agent-attributable spend.

For full configuration details, see the Operator Intelligence guide.