Skip to main content
A report is a written analysis of your AI agent operation — the kind of thing a business analyst would put together in a real BPO. Where a finding is the durable record of one recurring problem and a change is the record of a fix, a report answers an open-ended question about how the operation is running: “Why did our conversion go up in May?”, “How are we handling the after-hours scenario?”, or “What’s different between agent A and agent B in the A/B test?” Reports live under Reports in the sidebar, scoped to your organization (not to a single agent). Written reports are organization-wide snapshots. Their list and content are hidden while a company is selected; return to All agents to read them.
Every plan includes one report, counted in total rather than per month. A report that fails to generate is not counted. The Observability and Improve plans include more. On the Free plan you can keep working on your first report, and the Reports page shows which plan unlocks the rest. See Plans, limits and retention.

Saved monitor reports

Open Monitors → Reports to group up to 20 monitors. Owners and admins choose a name, company, metadata filters and timezone. All filters intersect each monitor’s configured scope. A saved report uses its own scope, independent of the sidebar company filter; it does not change which conversations a monitor evaluates. The default period is the previous complete Monday–Sunday in the report timezone. Choose up to 31 calendar days and download all selected monitors together as CSV or JSON. The table separates conversations in scope, evaluated conversations, unscored results and conversations without a result. Affected rate uses evaluated conversations as its denominator, without extrapolating sampling. No evaluations means “No data”. Only the current monitor version counts. Enable Send this report every week and enter a public HTTPS webhook URL and a signing secret of at least 16 characters. Reports run Mondays at 09:00 in the saved timezone and arrive within the scheduler’s 15-minute interval. Delivery history shows retries, failures and downloadable snapshots. Retries preserve the snapshot and delivery ID; your receiver should deduplicate that ID. The webhook includes Idempotency-Key, X-Zelto-Delivery-Id, X-Zelto-Timestamp and X-Zelto-Signature. Verify v1= plus the hexadecimal HMAC-SHA256 of timestamp + "." + rawBody using your signing secret. Acknowledge with 2xx. The body is { "type": "monitor_report.weekly", "deliveryId": "...", "report": { ... } } and contains aggregate counts, not transcripts or conversation evidence. You can also pull GET https://api.zelto.ai/v1/reports/monitors/REPORT_ID?from=2026-09-07&to=2026-09-13 with your organization’s API key in Authorization: Bearer YOUR_API_KEY. Find REPORT_ID in the saved view’s id URL parameter. Add format=csv for a spreadsheet export or deliveryId=DELIVERY_ID for a frozen historical snapshot. Omit both dates to use the previous complete week. The Reports feature must be included in your plan.

Requesting a report

Press Request report, describe what you want to understand — “Compare how agent A and agent B handle objections, and explain why conversion went up” — and press Generate report (or ⌘/Ctrl + Enter). The request can be up to 4000 characters. The report opens on its own page, split in two:
  • The chat, on the left, sends your request to an AI assistant. It researches your organization’s data with read-only tools — agents, conversations, reviews, findings, and more — and each lookup appears in the chat as it runs.
  • The report, on the right, fills in while the assistant writes it. When it finishes, the report is saved, named, and marked Ready.
The assistant can query your organization’s data to calculate totals, compare periods, and prepare chart data. Queries are read-only and restricted to your organization. This research capability is unavailable in HIPAA (PHI) mode. If a query fails or the available data is incomplete, the assistant uses its other research tools and explains any gaps in the report. Keep talking to the assistant to refine the report: “make it shorter”, “add a recommendations section”, “break the numbers down by week”. Each change rewrites the report in place, starting from what’s there — including any edits you made by hand. You can also ask questions about the report without changing it. The chat is saved with the report and shared with everyone in your organization who opens it. The report keeps being written if you close the page or navigate away — come back later to find it finished. One generation runs per report at a time: if a teammate (or another tab) is already writing it, the page shows that it’s in progress and updates when it’s done. Report generation has a monthly usage limit per organization; once it’s reached, new requests resume on the 1st of the next month.

Charts

When a trend, comparison, or breakdown reads better as a picture, the assistant draws a chart. It writes a short Python (matplotlib) program over the numbers it gathered and runs it in an isolated sandbox — a throwaway virtual machine with no network access, so the data can’t leave it. Each chart appears in the chat as it’s made and lands in the report as an image block with a caption, which you can edit, replace, or delete like any other image. Ask for charts directly (“add a chart of weekly volume”) or for changes to one (“make it a line chart”). A turn draws at most four charts. Charts are not available to organizations with HIPAA (PHI) mode enabled. The assistant only reads your data — it can’t change agents, findings, or anything else. It reports only figures it found; when your data can’t answer part of the request, the report says so.

Generation status

Every report carries a generation status shown as a badge on the list and detail pages: A failed rewrite keeps the previous version and its Ready status. The body is fully editable regardless of status.

Creating and editing from chat or the API

You can also file a report without the assistant writing it. Ask Zelto in chat — “create a report titled … about …” — and it proposes the report for you to confirm before it’s saved. The same actions are available to scripts and external agents over the MCP server as the create_report and update_report tools (with list_reports / get_report to find and read existing ones), so an automated workflow can file a report or update its title, request, body, or status programmatically.

Anatomy of a report

Data widgets

A data widget turns a report from a static write-up into a live dashboard. It runs a query against your analytics store and renders the result as a table or chart that re-runs every time the report is opened, so the numbers stay current as new conversations come in. In the report editor, type / and choose Data widget, then fill in: Press Run to execute it. Use the refresh button in the widget header to re-run on demand, the pencil to change the query, and the chart/table toggle to switch views. Results are capped at the first 1,000 rows. Data widgets are read-only and scoped to your organization: every query is automatically restricted to your org’s rows on the analytics tables, and only SELECT / WITH / DESCRIBE / EXPLAIN / SHOW statements are allowed — writes and cross-organization access are rejected. The same querying capability is available conversationally through Ask Zelto and over the MCP server.
  • Findings — the recurring problems a report often explains.
  • Changes — the fixes that follow from what a report uncovers.
  • Agents — the agents whose behavior a report analyzes.
Monitor reports also support Download PDF, including historical deliveries. PDFs include the company, period, timezone, saved filters, metrics and evaluation coverage. Use format=pdf&locale=en-us on the same API endpoint; supported PDF locales are en-us, es and pt. Weekly webhooks send JSON; use the delivery ID and your API key to download that delivery as a PDF.

Send monitor results to your system

Use outbound webhooks to send each new positive monitor result to your database or CRM. Configure destinations under Settings → Integrations → Webhooks, verify a test, and deduplicate received events by their event ID. These per-conversation monitor.matched events are separate from weekly aggregate report webhooks.

Share outside your organization

Owners and admins can select Share on a written report, benchmark, or experiment to create a read-only /shareable/TOKEN link on the current app host. Anyone with the link can open the shared view without signing in; it does not grant membership or access to the organization workspace. Choose the content mode when creating a link:
  • Snapshot stores the saved content and results at creation time. Later report edits and new evaluations do not change that copy. Images retain their saved URLs.
  • Live reads the latest saved content and results when the link is opened. For benchmarks it follows the latest completed run.
Choose an expiration of 7, 30, or 90 days (30 the first time; after that, your last expiry is remembered; the type always starts on Snapshot). The form shows the date it ends. Optionally add a label naming who the link is for (for example “Acme QBR”), then select Create link. The link is copied for you when the browser allows it. On a benchmark, a snapshot freezes the run you are viewing, or the latest completed run, and the form names it (“Freezes Run #12”) before you create the link. If the selected run has not finished, only Live is offered; if nothing can be shared yet (a report with no saved body, a benchmark with no completed run), the popover says why instead of showing the form. Once a resource has links, the Share button shows how many are active, Active links comes first, and New link opens the form. Each link is titled by its label, or by its mode and creation time, and shows when it expires (and who created it, if someone else did). Copy link is on each link; its More actions menu (…) holds the rest:
  • Open shared view to review what recipients see.
  • Add label or Edit label. Saving an empty label removes it.
  • Extend to 7, 30, or 90 days from today without changing the URL. Only periods that lengthen the link are offered, and no link stays public more than one year after it was created; past that, create a new link.
  • Revoke. Anyone using the link loses access immediately.
With two or more links, Revoke all (next to New link) revokes every link to the item at once after a confirmation. Members who can’t manage links still see a Public · N links indicator while any link is active, because their edits (and new benchmark runs) show up in Live links. Expired and revoked links are unavailable; deleting the source also removes its links. Written reports need a saved body before sharing. The public view includes rich text and HTTPS images. Data widgets are materialized as tables with up to 500 rows each; their SQL is not shared. Diagrams appear as code. Call examples, prompt embeds, private hyperlinks, and the report chat are excluded. Review the saved body and widget results before sending the link. Benchmarks share aggregate candidate results, while Experiments share monitor comparisons and statistical evidence. These links do not expose individual calls, transcripts, recordings, or experiment configuration.