AI Metering is the workspace-wide dashboard for AI usage, cost, and performance across all your agents and models.
Overview
AI Metering is a workspace-level dashboard that reports usage, cost, and performance across all your AI agents and models — not just one agent. It's a reporting view: it shows actual consumption (tokens, requests, cost, latency, errors) broken down several ways. It does not enforce quotas, credits, or hard limits. Workspace-wide caps live under AI setup > Spending > Limits, and one agent's cap lives on that agent's Spending tab. Use AI Metering to understand adoption and spend across the workspace.
The Four Breakdowns
Metering organizes everything into four views:
View | What it shows |
|---|---|
Model | Usage by model: input/output/total tokens, total cost, request count, average latency, error rate, and how many agents use each model. |
Agent | Usage by agent: cost, request and session counts, average session duration, response time, tokens, errors, tool and knowledge usage by agent, and a breakdown by department. |
Search | Knowledge and search analytics: query volume, response time, cost, error rate, top queries, which knowledge sources answered queries, the share of answers from internal vs. external knowledge, and user feedback (upvotes/downvotes). |
User | Adoption: unique users who ran at least one query, daily and monthly active users, user retention, favorite topics, and sessions, queries, and attachment-queries per user. |
What Exactly Is Measured
Metering reads from the platform's record of every model call, so it can measure tokens (input, output, and total), cost (including per-token input and output cost), request counts, and latency — sliced by dimensions such as model, provider, agent, tool, user, topic, department, and modality. Beyond model calls it also counts sessions and conversations, queries and attachment-queries, knowledge-source usage, and user feedback. Together these cover both how much the AI is used and how well it performs.
How Cost Is Computed
Cost is computed per call, at the moment of the call, from actual token usage and the model's prices in the model library. Prices are per million tokens, with separate rates for input, output, cache reads, and cache writes. Cached tokens are deducted from the regular input count and charged at their own (cheaper) rates, so prompt caching genuinely shows up as savings:
cost = (input tokens − cached tokens) × input rate + output tokens × output rate + cache-read tokens × read rate + cache-write tokens × write rate, all ÷ 1,000,000
Every call's cost is recorded on its trace along with the token breakdown — metering dashboards, experiment cost rollups, and spending counters all aggregate from these same per-call records, so the numbers agree with each other.
Filtering the Dashboard
Narrow the dashboard by time range, and segment by dimensions such as department, agent, and model. Set the time window first, then drill into a specific agent or model to see its slice. The numbers are computed from live usage data, so they reflect real consumption over the period you choose.
Metering vs. Spending
Metering reports what was actually consumed — it has no quota, credit balance, or hard cap. To control spend:
For workspace-wide caps across users, agents, and models (including an Org Cap), use AI setup > Spending > Limits. See AI Setup – Spending.
For one agent's monthly limit, alerts, and auto-undeploy, use that agent's Spending tab. See Spending and Alerts.
Metering is for visibility; Spending is for control.
Note: If you're looking for a hard cap or an auto-pause on overspend, it isn't here. Use AI setup > Spending > Limits for workspace-wide caps, or an agent's Spending tab for that agent's cap and auto-undeploy.