Skip to main content
ON DEMANDEveryone builds. IT approved.Watch the replay
Sigma Computing
Fundamentals

AI KPI monitoring: How scheduled and triggered agents watch metrics & send alerts

Justine Yuan
Justine YuanSoftware Engineer
November 19, 2026
9 min read
AI KPI monitoring: How scheduled and triggered agents watch metrics & send alerts

A threshold alert tells your team that a number moved. AI KPI monitoring adds the next steps. An agent in a scheduled or webhook-triggered sequence checks the surrounding data, explains the likely cause, and can start the follow-up before anyone opens a dashboard.

This approach fits operations, RevOps, and finance teams that open the same dashboards every morning. It also serves teams that have learned to ignore alerts with no explanation attached.

Key takeaways

  • Fixed rules, such as hard limits and missing-data checks, work well as conditional alerts. Agents detect less static anomalies that depend on seasonality, launches, or several data sources.
  • Agents can run on a schedule or start when an external system calls a webhook. A condition in the sequence routes to sending notifications when the agent finds something worth reporting.
  • Approval prompts don't pause unattended runs, so approvals for automated work can run as a round trip through Slack.
  • Unattended agents run with the permissions and row-level security of the user who scheduled them, and each run is logged for review.

What AI KPI monitoring is

AI KPI monitoring uses AI agents to check key performance indicators against live warehouse data. The check runs on a set cadence or when an event arrives from another system.

A threshold alert stops at "pipeline coverage fell below target." A monitoring agent's message names the accounts that moved, offers the likely reason, and can point to a next step the agent has already started, such as a ticket or an approval request.

Monitoring differs from producing recurring reports, which go out on schedule whether or not anything changed. A monitoring agent speaks up when a metric moves.

Why threshold alerts alone create noise

A static threshold can flag expected changes as problems. Launch week can push signups past the ceiling, and a holiday can drop orders below the floor. Both alerts fire without understanding the broader context.

An unexplained alert also sends someone back to the dashboard to investigate by hand. If that happens repeatedly, people stop reading the alerts and return to checking dashboards every day.

When a conditional alert is enough

For fixed rules that need no explanation, a conditional alert is the right tool. Sigma is the AI apps and analytics platform that lets users build agentic analyses and applications directly on live cloud data warehouse data. Teams can run conditional alerts and monitoring agents against the same workbook data.

To set up one of Sigma's conditional alerts, you choose Alert when on a workbook element and pick a condition. Conditions include whether data exists, whether it's missing, and value comparisons on a column or KPI.

Conditional alerts can also catch statistical outliers. A formula column computes standard-deviation bounds and sets an isOutlier flag, and the alert fires when that flag is true. Test condition checks the rule against current data before it goes live, and Stop notifying after (beta) caps repeat notifications.

ApproachTriggerLogicOutputBest use
Conditional alert, value ruleScheduled checkData exists, data is missing, or value crosses a set comparisonNotification that the rule was metHard limits and missing-data checks
Conditional alert, outlier flagScheduled checkFormula column sets isOutlier from standard-deviation boundsNotification that a point is unusualStable metrics without strong seasonality
Monitoring agentSchedule or webhookPrompted agent reads workbook data plus outside contextExplanation, likely cause, and follow-up actionAnomalies that depend on calendars, launches, or several sources

How agents watch metrics: scheduled and triggered runs

Agents run unattended through automated action sequences, which are in public beta. A sequence can start in two ways, and neither requires anyone to open the workbook.

On a schedule trigger, a sequence runs daily, weekly, monthly, or on a custom cron schedule. On a webhook trigger, an external system such as Slack, Salesforce, ServiceNow, n8n, or Zapier sends an authenticated POST request to a webhook URL. Values the request carries become action variables that later steps can use. The schedule or the webhook request starts the sequence, and its Call agent action runs the agent.

What the agent does on each run: check, explain, act

Each run is a short agentic workflow with three steps:

  • Check: A Call agent action sends the agent a prompt that a person wrote. The agent reads workbook data and any context it can reach through tools, including warehouse agents, Model Context Protocol (MCP), or API calls. That outside context can be a campaign calendar or launch schedule, so the agent can judge whether a spike matches a planned event.
  • Explain: The agent writes a summary and a likely cause for someone to confirm. Structured output variables, such as an email subject and a body, carry that result to the next steps.
  • Act: Follow-up actions include sending a notification by email or Slack, writing the finding to an input table, calling an API, running a stored procedure, exporting attachments, and running a Python element.

With if/else logic between the explain and act steps, the sequence can notify people only when the agent finds something worth reporting. On a quiet day, nobody gets a message.

Sigma's anomaly detection example writes each finding to an input table before notifying anyone. The table keeps a running log of detected anomalies for later review and stages the message for the notification step.

Starting follow-up work: notifications, approvals, and updates

Beyond notifications, a sequence can update other systems through API calls. It can open a ticket in a service desk when a breach repeats, or update a customer relationship management (CRM) record when a deal slips.

Approvals take more care. An unattended run can’t stop and wait for someone to click an approval prompt, so the sequence hands the decision to Slack. When an automated action calls an agent, any action that needs approval waits for a person’s sign-off in three steps:

  1. The sequence posts an approval request to Slack and stops short of the action that needs approval.
  2. A person approves or rejects it there.
  3. The decision comes back through an authenticated webhook trigger, which writes it into the workbook and runs the gated action only if the request was approved.

This way, a service-level agreement (SLA) breach can escalate to a manager in Slack before a ticket opens. A reorder can route to procurement for sign-off before the order update is written to the warehouse.

AI KPI monitoring by team

Each team watches different metrics, but the check, explain, and act steps carry over. The setups below are illustrative.

  • Operations: An operations team could schedule an agent to check SLA metrics after the expected refresh window. When a breach occurs, the agent explains which queue or region caused it and alerts the owner.
  • RevOps: A RevOps agent could spot slipping deals or a shift in pipeline coverage and name the accounts that moved. It can then post the finding to Slack or update the CRM.
  • Finance: A finance team could point an agent at revenue and cost of goods sold to flag the line items driving a change. The pattern also extends to budget variance.

Start simple on the first agent with one metric with one owner, so the team can progressively evaluate its outputs before adding more checks.

Guardrails for agents that run unattended

An agent that works while nobody watches maintains limits that your team can see and adjust. Automated sequences run with these guardrails:

  • Inherited permissions: A scheduled sequence runs with the permissions and row-level security of the user who created the schedule. If another user publishes a change to the sequence's name, schedule, or steps, ownership moves to them, so check ownership after those edits.
  • Admin visibility: Admins can see each automated action sequence in the organization. The agent management page shows ownership and an execution log for each run, plus token usage once the AI usage dashboard is set up.
  • Failure handling: After repeated failures, the owner gets an email alert, and Sigma will suspend the sequence. Publishing the workbook or selecting Resume sequence restarts it, so fix the underlying error first.

When the agent's data sources are the governed metrics your data team already defined, the explanation it sends rests on numbers the team trusts.

Set up AI KPI monitoring in Sigma

Start with an agent that already knows your KPI tables, then wrap it in a sequence. Sigma runs only the sequences saved to the published workbook, and a sequence can only write to input tables that allow edits in the published version. Sigma agents are generally available as a premium paid feature. You'll need the Manage agents permission, permission to publish workbooks, and Can edit access to the workbook. A first monitoring agent then takes five steps:

  1. Build or reuse a custom AI agent with the KPI data, instructions, and tools it needs.
  2. Add an automated action sequence on the workbook page, then choose a schedule or On webhook received.
  3. Add a Call agent action with a specific prompt and output variables, such as an email subject and body.
  4. Add the follow-up actions and instruct the agent on the conditions that should trigger the notification.
  5. Publish the workbook, then confirm the sequence appears on the Automated actions page.

Replace dashboard checks with Sigma agents

Monitoring agents can replace the morning dashboard check with an alert that already says what moved and what to do next. Sigma runs these agents on live warehouse data in Databricks, Snowflake, BigQuery, or Amazon Redshift. Their findings rest on the data and permissions your data team already governs.

Get a demo or try Sigma free.

FOLLOW SIGMA

Related articles

Get started