Core Concepts¶

This page explains the key ideas behind Trust3 AI Governance so you can get oriented before diving into the product. If you are new to Trust3, read this before the Quick Start.

The governance problem¶

Enterprises run AI agents across multiple platforms — Databricks, Microsoft Copilot Studio, Azure Foundry, custom applications, and more. Each platform has its own admin console, its own logs, and its own way of describing what agents exist and who can use them.

The result: no one has a complete picture. IT teams piece together spreadsheets. Compliance teams ask questions that take days to answer. New agents go to production without documented owners or approvals. Agents that were registered last quarter have changed scope without anyone noticing.

Trust3 AI Governance closes this gap. It gives your organization one place to see every AI agent, evaluate it against your policies, and act on what you find.

Core concepts¶

AI asset inventory¶

The inventory is the central catalog of every AI agent and related asset Trust3 knows about. Each entry records where the agent lives (platform, workspace), who owns it, what it is for, which identities can invoke it, and how it is performing against your policies.

Inventory is built two ways:

• Automated discovery — Trust3 connects to your platforms (Databricks, Microsoft Copilot Studio, Azure Foundry) and pulls agent metadata on a regular schedule. New agents appear in the inventory automatically.
• Manual registration — For platforms without an automated connector, teams register agents through a structured form. GIA assists with filling the form from a plain-language description.

Environments¶

Agents and MCP servers record the environment they were detected in, taken from the source platform — for example, a project or workspace. Use the environment to tell test and production assets apart at a glance, and to filter the inventory. See Filter AI Assets.

An agent that exists in several environments or accounts — for example, the same agent in DEV, QA, and PROD — appears as separate inventory entries, one per environment. Trust3 identifies each entry by the unique ID the source platform assigns, so agents with the same name never overwrite one another.

Editing inventory records¶

Inventory records are not read-only. From an asset's detail page you can correct its Purpose, Environment, and Owner, including assigning more than one owner. Each edited field offers Revert to System Default, which restores the value from the latest scan.

Trust Score¶

Every agent in the inventory has a Trust Score on a 1.0–10.0 scale (shown in the UI as X.X / 10) plus a trust level band and a derived risk level.

The Trust Score is not a fixed label. It moves as the agent's governance state changes: it drops when violations open, unapproved models are detected, or required information goes missing, and it recovers when those issues are resolved. See Trust Score for how scores are calculated.

Policies¶

Policies are the rules Trust3 evaluates every agent against. They encode what "governed" means for your organization — whether agents have assigned owners, whether they use approved models, whether they are documented appropriately for your compliance framework.

Trust3 ships with pre-built policy packs for common frameworks including NERC CIP / Utility, EU AI Act, FERC, and Internal AI Governance. You can also create custom policies manually or by uploading a compliance document and letting GIA extract the rules.

When an agent fails a policy rule, Trust3 records a violation — what failed, the severity, the affected agent, and what remediation looks like — and raises it as a risk to track to resolution. See Policies for examples and how violations are handled.

Policy types¶

See Policies for evaluation modes, rule fields, examples, and how to create policies.

Risks¶

A risk is a governance problem Trust3 has found on an asset — a policy violation, or an agent drifting outside its declared scope. Earlier releases called these issues.

The Open Risks page lists each risk with its type, the affected asset and platform, the policies it violates, its severity, and when it was detected. Use the All Risks, Critical, and Assigned to Me tabs, or search, to focus the list.

From a risk you can link the ticket your team is tracking it in, and Mark as Resolved with a note describing how it was fixed. A risk you mark as resolved is not reopened by later evaluations, so resolve it once the underlying problem is actually fixed.

Shadow AI¶

Shadow AI refers to agents operating in your environment without governance oversight — agents that were never registered, or agents using organizational credentials that no one is tracking.

Trust3 surfaces two types of shadow AI that organizations can act on:

• Agents running on approved infrastructure (Databricks, Microsoft Copilot Studio, Azure Foundry) that were never registered in the governance inventory
• Agents calling model APIs using credentials issued by your organization but not tracked in your governance program

When Trust3 discovers an unregistered agent, it appears in the inventory so your team can evaluate it and decide what to do.

Pre-development approval¶

The pre-development approval workflow is how new AI agents get documented sign-off before they reach production. A developer submits a registration describing the agent, its purpose, its data scope, and its intended users. The submission goes through a review chain — Legal, then Compliance — before the agent is approved to build. Reviewers work through submissions on the Approval Requests page, where the Pending Legal and Pending Compliance tabs show what is waiting at each step.

Every step in the chain is recorded: who submitted, who reviewed, who approved, and when. That record stays attached to the agent's inventory entry permanently and is available in compliance evidence exports.

GIA — Governance Intelligence Agent¶

GIA is the AI assistant built into Trust3. It works across inventory, policies, and workflows and is available from the header on any page.

You can ask GIA governance questions in plain language and get answers grounded in your actual inventory:

• "What needs attention now?"
• "Show me agents without owners"
• "Which agents have low trust scores on Databricks?"
• "Why did this agent's trust score drop?"

GIA also assists with registration: describe a new agent in one sentence and GIA fills in the registration form. For policies, GIA can read a compliance document and extract enforceable rules ready for your review.

GIA remembers your conversation as you navigate between pages, so follow-up questions build on prior ones naturally.

Identities and ownership¶

Every agent should have a documented owner — the person or team accountable for it in production. Trust3 also maps the identities that can invoke each agent (users, service principals, automated accounts) from the platform metadata it discovers.

For short-lived or delegated credentials, Trust3 links them to a parent identity so compliance reviews show who ultimately acted, not just which temporary token appeared in a log.

Identity and access lineage¶

The Identity tab on an asset shows everyone who can use that agent — individual people, teams, and groups — in one place. It helps you answer access questions quickly: Who can run this agent? Did we grant access to the right people? Is access broader than we intended?

Compliance and security teams use this view for access reviews and incident follow-up, without hunting through each platform’s admin console separately.

Agent security posture¶

Trust3 also records each agent's security posture — how exposed it is and how users and tools authenticate to it:

• Microsoft Copilot Studio — whether the agent is shared broadly, its sharing level, how many people and groups it is shared with, how many tools and knowledge sources it can reach, its access-control setting, and how end users and tools sign in.
• Azure AI Foundry, Microsoft 365, and Databricks Genie — how end users sign in and, where the platform reports it, whether the agent is shared broadly.

You can use these attributes as conditions in your policies. Trust3 also includes ready-made Compliance policies for common agent security risks, including risky combinations — for example, an agent that is shared broadly and does not require end users to sign in. Violations of these policies lower the agent's Trust Score.

Asset lineage (data dependencies)¶

The Lineage tab shows what an agent is built on and connected to — datasets, tables, other agents, models, MCP servers, tools, and knowledge sources it relies on to run. It helps you answer: What does this agent actually use? What information can it reach? Is any of it sensitive? Is it using an allowed model? What tools does it have access to? Which knowledge sources is it using?

Data governance and risk teams use this view to understand the full picture behind an agent, spot unexpected dependencies, explain exposure to auditors, and prioritize agents that need closer review.

Together, Identity and Lineage show how each agent fits your estate: who can use it and what it can reach.

The AI Lineage page extends this to your whole estate. Assets appear as cards you can group by platform or trust score, search, and filter; open any asset to explore its lineage graph and the agents, data assets, and user identities it connects to.

Overview¶

The Overview page is the starting point for your estate: governance metrics, agents by trust score and by provider, the risks that need attention, and your recent conversations with GIA.

Monitoring¶

The Monitoring view gives your team a live operational picture of agent behavior across your entire estate. It surfaces two classes of signal:

• Scope drift — agents queried or acting outside their declared purpose, detected from trace data.
• Token anomalies — agents consuming tokens significantly above their baseline, which can indicate runaway behavior or misuse.

The summary tiles show Unique Users, Total Sessions, Avg Tokens/Day, and Drift Rate at a glance, alongside the count of active Token Anomalies. The agent behavior table lists every agent with its trace count, scope drift status, token consumption, and an overall risk level — so your operations or security team can spot outliers and act without navigating each platform separately.

Finding your way around¶

The sidebar groups the pages by what you are doing:

Open GIA from the button in the top bar on any page.

How the pieces connect¶

An agent is discovered (automatically through the AI Assets Collector, or by registration) and added to the inventory. Trust3 maps identities that can access it and builds lineage graphs of its downstream data dependencies. It evaluates the agent against your active policies — including violations propagated from sensitive dependencies — and assigns a Trust Score. Violations are raised as risks and tracked to resolution. If the agent was new, it went through pre-development approval before reaching production. GIA gives your team natural-language access to everything in the inventory so you can ask questions, act on findings, and assist with governance tasks without navigating multiple screens.

The Overview page brings it all together: one view of your entire AI estate, scored, evaluated, and ready to act on.

Next steps¶