An operations dashboard for your AI estate gives you a real-time view of what is running, what it is costing, and whether anything has changed in a way that warrants attention. The value of a single aggregated view is not convenience — it is the ability to correlate signals across agents, spend, members, and security events that would otherwise sit in separate tabs and never be compared.

This post covers what belongs on an AI operations dashboard, how to read the signals it surfaces, and what to do when the numbers look wrong.

What "AI estate" means in practice

An AI estate is the set of agents, connections, workflows, and users that your organization has put under management. Unlike a microservices estate — where the inventory is relatively stable and defined by deployment — an AI estate can grow quickly as teams add new agents, connect new data sources, and build new workflows. The estate is also more heterogeneous: some agents run on a fixed schedule, some respond to webhooks, some are invoked interactively.

A dashboard that only shows infrastructure health misses most of what matters here. You need agent-level signals (how many are online, what is their trust posture), task-level signals (throughput, failure rate, latency trends), and financial signals (cost over time, pacing against budget) — all in one place, scoped to your organization.

The KPIs that matter at a glance

The top of an AI operations dashboard should answer a set of questions you ask every time you open it:

How many agents are active and online? Distinguishing between total registered agents, agents that have run recently, and agents that are currently connected tells you whether your fleet is healthy. A gap between active and online counts — agents that have run recently but are not currently reachable — is worth investigating before it turns into a missed task.

How many members does the organization have, and are the right people in? This is as much a security question as an operational one. A count that does not match your expectations — too high after an offboarding, too low after an onboarding — surfaces access hygiene issues early.

How many guardrails are active? Active guardrail count is a proxy for your governance posture. If the number drops unexpectedly, a policy was disabled or deleted. If it climbs sharply, someone is adding rules — which may be good, or may indicate a reactive response to an incident that should be tracked more formally.

What is the average trust score across agents? Trust scores aggregate signals about agent behavior, attestation status, and policy compliance. A declining average trust score across the fleet is a leading indicator: something is shifting in how your agents are configured or behaving, even if no individual agent has tripped an alert yet. The mechanics of how these scores are computed are explained in trust scores and attestations: deciding which agents to trust.

What security events happened recently? A count of recent audit events — permission denials, authentication failures, administrative actions — gives you a quick read on whether the security surface is quiet or active. You dig into the detail elsewhere; the dashboard tells you whether to look.

The two time-series that belong on every AI operations dashboard are task volume over time and cost over time. Together they answer the question that matters most for day-to-day operations: is the platform doing the work it is supposed to do, and is it doing it at the rate you expect?

Task volume trending down mid-week when it should be flat suggests an integration broke, a trigger was misconfigured, or an agent is failing before work is recorded. Task volume trending sharply up suggests a new workflow launched, a loop is running uncontrolled, or someone connected a new high-frequency trigger. Both directions warrant attention; neither is obviously bad without context.

Cost over time tells the same story through a financial lens. Costs that track task volume proportionally are healthy. Costs that rise faster than task volume suggest efficiency is declining — longer runs, more tool calls per task, or a model switch that costs more. Costs that rise while task volume stays flat suggest background activity you have not accounted for. For guidance on acting on these signals, see visualizing AI usage and cost.

A seven-day view is generally the right default window for these charts. It is long enough to show weekly patterns — usage is typically lower on weekends, for example — and short enough to keep recent changes visible without being buried by older data.

Reading the onboarding checklist as an operational signal

Most dashboard designs include an onboarding checklist for new organizations. For operators, this checklist serves a secondary purpose beyond setup guidance: it is a quick audit of whether the foundational controls are in place.

If your organization has been running for months and the checklist still shows items unchecked — no MFA configured, no guardrails active, no API key rotation policy set — those are operational gaps, not onboarding oversights. Reviewing the checklist periodically is a low-effort way to catch configuration drift before it becomes a finding in a security review.

What the dashboard does not replace

An operations dashboard is designed for breadth, not depth. It tells you that something changed; it does not tell you why. When a signal on the dashboard warrants investigation, the next step is always to move into a deeper view: the agent detail page, the task log, the audit trail, or the advanced analytics surface. For the deeper analytics layer, see Advanced Analytics for AI Operations.

This is intentional. A dashboard that surfaces every possible signal at full detail becomes unreadable. The job of the top-level view is to narrow the search space, not to close it.

Specifically:

  • If agent online count drops, move to the agent list to see which agents are offline and when they last connected.
  • If cost over time spikes, move to cost analytics to see which agents or workflows drove the increase.
  • If audit event count rises, move to the audit log filtered to the relevant time window to see what actions occurred.
  • If trust score average declines, move to the agent trust view to see which agents changed and what signals contributed.

The dashboard is the starting point for each of those investigations, not the ending point.

Authentication and API access

The dashboard supports both browser session authentication and API key authentication, which makes it practical to pull these stats into external monitoring systems or executive reporting workflows without requiring an interactive login.

If you are integrating dashboard stats into a Slack digest, a scheduled report, or an external status page, an API key with read-only scope is the right credential. Avoid using a full-privilege key for a read-only integration — scope your key to what the integration actually needs. For a broader look at what the full API surface exposes, see organization API keys and scopes.

Praesidia's dashboard design

Praesidia's home dashboard surfaces the core KPIs — total agents, active agents, online agents, member count, active guardrails, average trust score, and recent audit events — alongside seven-day task and cost time-series, all scoped to the authenticated organization. Every read is organization-scoped, so a user in multiple organizations sees only the data for the organization they are currently viewing.

The dashboard is designed to stay responsive under load, so opening it after a period of inactivity does not degrade the platform for other users. KPI reads are scoped and served quickly even for organizations with large agent fleets or high task volumes.

For teams that want real-time alerting on top of the dashboard view, see Slack and Multi-Channel Alerting for how to route threshold breaches to the right channels.

Common questions

How often does the dashboard update?

Dashboard KPI cards reflect the state of the platform within a few minutes of a change; they are optimized for low-latency reads rather than sub-second precision. If you need real-time precision on a specific metric — for example, to monitor a workflow that is running right now — use the task detail view or the real-time event stream, which update continuously.

Can I embed dashboard stats in an external tool or report?

Yes. The stats endpoint accepts API key authentication in addition to session tokens, so you can fetch the same data programmatically. Scope the API key to read access only, store it in your secrets manager rather than in the integration config directly, and rotate it on your standard key rotation schedule. The response format is covered in API-First: The Praesidia API Surface.

Why does average trust score show zero for some organizations?

The average trust score is computed only across agents that have a non-zero trust score. An organization that has registered agents but has not configured trust attestations or allowed enough task history to accumulate will show zero — which is technically correct but can be surprising. It is a signal to configure trust scoring for your agents, not a data error.

What should I check first when task volume drops unexpectedly?

Start with the agent list to confirm which agents are online and when they last ran. A widespread drop usually points to a trigger misconfiguration, a connectivity issue between the platform and external systems, or a policy change that inadvertently blocked task acceptance. A drop isolated to one agent points to that agent's configuration or the specific workflow it serves. The dashboard narrows it to an agent; the agent detail view and task log tell you the root cause.