Insights

Insights calculates Organization metrics from stored runtime events. Use the available date range and Assistant filters to change the data set.

Charts in Insights, Eval, Assistant Overview, and Usage use a grayscale palette that adapts to light and dark themes. Positive and negative outcomes remain green and red. Changes use each measure's direction: lower costs, latency, and failures are positive. Labels, legends, and exact values identify each measure. The latency timeline uses a dashed line for p95.

Chart and stat cards place their visualization in a rounded inset. The title and description sit below it, with period changes and available actions on the right. Costs, Observability, platform Admin, and individual Eval runs use the same card layout.

Motion follows Transitions.dev recipes. Updated numbers enter with a brief digit stagger, and captions swap with a light blur. Loading cards fade into their content, and data tables expand smoothly. The active tab indicator slides between selections. Reduced-motion settings remove the digit movement and blur from these transitions.

Current overview

The overview can include these measures:

• Conversations and unique Visitors
• AI answers
• AI Resolution Rate
• answer ratings and customer satisfaction data
• escalations
• answers per Conversation
• questions per Conversation: the number of messages that the Visitor sent
• average Conversation time: the time from the start of a Conversation to its last message
• languages and engagement ratios
• feedback and grading measures
• Notifications Sent

The exact cards depend on available data for the selected period.

The Conversation depth chart shows the average Conversation time and the questions per Conversation for each day, week, or month. Ciele puts each Conversation in the period when it started. A Conversation that continues after the end of the range keeps its full time.

The Usage chart separates counts, percentages, ratios, and seconds so different units do not share one axis. Select a metric to show or hide it. The Assistants and Channels tabs show Conversation breakdowns.

Use the Explore conversations over time window to zoom the charts within the selected date range. Drag the window or its handles, or use the arrow keys after focusing a handle. Summary cards, the data table, and exports still cover the full date range.

Costs and Observability

Two dashboards show what the AI of your Organization costs and how well it works. Costs shows spend and usage. Observability shows reliability, speed, and answer quality.

Both dashboards use the same filters. Select a date range, then select All, Assistants, Teammates, or Internal.

• Assistants shows widget and Preview traffic, and inbound HTTP Flows.
• Teammates shows AI Teammate chats, channel turns, and Routines.
• Internal shows platform work: indexing, scheduled jobs, and calls from API keys and MCP.

With All or Assistants selected, you can also select one Assistant.

Each dashboard starts with stat cards. A card shows a value for the range, a line of its daily values, and the change against the prior period. The prior period has the same length and ends the day before the range starts. Move the pointer along the line to read the value for each day.

Costs

• estimated spend, cost per turn, model calls, and tokens
• total estimated spend per day
• a streamgraph of daily estimated spend by surface. Band thickness represents an amount, not a percentage
• a calendar of token use per day
• a monochrome gauge for input tokens as a share of all tokens, with exact input and output counts
• spend per model and per pipeline stage
• a treemap of spend by provider and model. Select a provider to inspect its models

The surface filter applies to the daily charts and model breakdown. The surface totals below the streamgraph always compare all surfaces.

Spend is an estimate in euros. Ciele calculates it from token counts and list prices. It is not an invoice.

Observability

• finished turns, success rate, failed turns, and tool calls per turn
• turn latency: median, 95th percentile, a histogram, and a daily trend
• Answer accuracy: the share of Assistant answers that the verifier passed
• Autonomy: the share of Visitor Conversations that did not go to a help desk
• monochrome gauges for reliability, accuracy, and autonomy, with exact counts beside the charts
• a comparison of reliability, accuracy, and autonomy against the prior period, in percentage points
• the most frequent error classes
• the five most used Flows, ranked against each other per day or per week

Answer accuracy and Autonomy apply to Assistants only. For Teammates and Internal, these measures do not show.

Notes

Days are UTC days. A model call recorded before surface attribution has no surface. It counts in All as Unattributed, and in no other surface.

Use Export CSV to download one row for each day in the range. For a report file with more options, use Exports.

Platform reporting

In managed deployments, authorized Owners and Admins of the designated platform Organization see Admin as the last Insights tab. This tab shows usage and estimated costs across all Organizations, including the platform Organization. Select a date range to review model calls, tokens, crawl pages, and estimated costs for each Organization. Platform-funded cost excludes customer API keys and personal subscriptions. Other Organizations and Members do not have access to this tab.

Data freshness

Insights stores the calculated overview and dashboards for up to five minutes for each Organization and filter set. Every Member of the Organization reads the same stored data.

A change a Member makes in the console (to an Assistant, a Conversation, or an Improvement) clears the stored overview at once. Widget traffic (new Conversations, answer ratings, satisfaction scores, escalations) reaches the numbers within five minutes.

To force a new calculation before the five minutes end, send DELETE /api/insights from a signed-in browser session. The response is 204 and the next read calculates the overview again. Any Member of the Organization can send this request.

Notification counting

A Notification is not an AI answer. A Notification-only session does not increase Conversation or answer KPIs.

The Notifications Sent measure and its chart series record proactive delivery separately.

Investigate a change

1. Select the required date range and Assistant.
2. Identify the metric that changed.
3. Open the Inbox with the same scope.
4. Read representative Conversations.
5. Create Improvements for actionable answer problems.

Use Exports for structured Conversation records.

Receive scheduled reports

Use Exports to schedule weekly or monthly Insights files. Choose Member recipients and keep the same report filters and file formats.

Read Exports for rolling date windows, notifications, and schedule controls.