AI Governance brings the AI you already run elsewhere under control from UnifyApps — the agents in your own SDK-instrumented apps and the LLMs your systems call on connected platforms — so you can discover, register, observe, and govern them without rebuilding them here.
Overview
AI Governance is about external AI: agents running in your own apps (AWS Bedrock, Anthropic, OpenAI, Snowflake, Databricks, Salesforce, ServiceNow, n8n, LangChain/LangGraph, LiveKit, or custom self-hosted code) and the LLMs those systems call. You connect the platforms they run on, register the agents and models you want governed, watch their traffic — sessions, traces, spans — and apply governance policies that can flag or block.
Note: This is distinct from AI Agents, where you build agents on the platform. AI Governance is about observing and governing AI that already exists elsewhere, not building new agents.
How the Pillar is Organized
Select AI Governance in the side-nav to open its launch pad. Its modules are grouped into five pillars that follow the workflow:
AI Assets — Discovery (connect the platforms your AI runs on, and sync their agents and models in), the Agent Registry (the external agents you're governing), and the Model Registry (the external LLMs you're governing).
AI Observability — Monitoring (a fleet-wide dashboard across your registered agents) and Incidents (the hits your guardrails log).
AI Compliance — Policies and Controls, the guardrails you apply.
AI Evaluation — Experiments, Datasets, and Metrics from the shared AI evaluation workspace (the same evaluators an asset's Evaluation tab draws on), plus Agent Optimizer, which ranks fixes for agent quality and cost, scored against real sessions.
AI Cost & ROI — Cost Explorer, which breaks down AI spend by agent, model, team, and use case, and Outcomes & ROI, which ties AI cost to business outcomes and measured value.
Note: Model Router isn't part of AI Governance. It's in the AI Agents section of the side-nav, next to Models.
Agent Registry vs. Model Registry
They're two halves of the same idea — bringing external AI under governance — split by what you're governing:
The Agent Registry holds agents (things that reason and act), which you register manually or sync from Discovery, and instrument to send their activity in.
The Model Registry holds models (the LLMs your systems call), which you import from a connection.
Both have Overview, Observability, Evaluation, Incidents, and Policies tabs. Agents also have a Setup tab for connecting them, a Monitoring tab, and an Outcomes tab.
Note: The Model Registry is not the same as the AI Agents Model Library, which is the on-platform catalog of models you build with.
The Discover → Register → Observe → Evaluate → Govern Flow
The pieces are meant to be used in order:
Discover — connect a platform in Discovery so its agents and models are found.
Register — bring the ones you care about under governance (register an agent or import a model; or turn on the Register toggle on a discovered asset).
Observe — instrument the asset so its traffic flows in, then watch its sessions, traces, spans, and activity.
Evaluate — add online evaluators that score the live sessions and traces for quality (an agent's or model's Evaluation tab).
Govern — apply policies and controls, and review the incidents they raise.
You can also skip Discovery and register an agent manually, or import models directly — Discovery just automates the first step.
Observe vs. Control
Each registered agent shows a Connection type describing its governance tier:
Tier | What it does |
|---|---|
Observe · Advisory | Governance watches and can flag, but can't stop a call in real time. |
Control · Enforced | Governance sits inline and can block, redact, mask, or require approval before a call completes. |
Note: Control · Enforced can be chosen only for platforms that support it — today the Anthropic Agent SDK, LangChain, LangGraph, n8n, and the OpenAI Agent SDK. On every other platform, Control is greyed out with the tooltip "This platform can't route calls through the gateway yet, so it can only be observed." Those assets run in Observe mode, where a control's Block action logs a "Blocked" incident without actually intercepting the call.
Roles and Permissions
Access follows the platform's standard roles and permissions: viewing a registry needs view access to it, and registering/deleting assets, connecting platforms, authoring controls and policies, and triaging incidents each need the matching create/edit/delete permissions on those asset types, granted through your workspace's roles.
There's no AI-Governance-specific permission screen — if a button errors with a permission message, ask your workspace admin for the relevant role.
Relationship to Enterprise Search Governance
The Controls, Policies, and Incidents in AI Governance are the same engine as Enterprise Search governance — built and triaged the same way. The differences are all about scope: AI Governance applies them across your whole fleet of registered agents and models (targeting agents and models, not chat users or knowledge sources), and only on the Chat surface.