For AI governance, security and model risk teams
Govern your existing agents.
Keep the agents your teams already run. Route the actions that matter through AAES Govern in your own environment.
What Govern can enforce
Govern can block an action only when the call goes through Govern with a key Govern holds. An agent that keeps its own keys can still use them. Govern then sees only the activity you report to it, and it can’t block that activity.
| Your agents | Govern enforces | Outside Govern |
|---|---|---|
| Agents where you set the model address and key, such as Claude Code, Codex, LangGraph, CrewAI or your own runtime | Tool calls sent to its MCP endpoint, and model calls sent to its model gateway | Calls made with keys the agent still holds |
| Microsoft Copilot Studio | Actions an administrator connects to Govern as tools | The agent’s own connectors and credentials, and its model calls |
| Salesforce Agentforce, Amazon Bedrock Agents, Vertex AI Agent Engine | Actions you point at Govern | Model calls, and tools wired straight to your systems |
| ServiceNow AI Agents | Actions an administrator adds to the agent as Govern tools | The agent’s other tools, and ServiceNow’s model calls and orchestrator |
Moving an agent’s keys into Govern is a change each platform team makes and tests.
What you can show your board
- A named, accountable person for every registered agent.
- Approval by an authorized person before an irreversible action, bound to that exact request.
- Money set aside from a budget before an action runs.
- Records your auditors can check offline, without AAES.
Start with one agent.
Pick one existing agent and one action that matters. A paid evaluation runs it in your environment with test data, your identity provider and your own model account. Pricing follows the size of the business unit or entity you license, so you can start with one unit.
AAES-hosted production runs across zones for high availability. Customer-hosted high availability requires a cluster that can power machines off through an API. Both deployment types use backup and restore. There is no SOC 2 report or penetration-test report. See the trust center.
What Govern enforces.
Govern applies the same checks to every agent that sends its calls through it, wherever that agent runs. It can block a call only if it controls the key that call needs.
What it stops
- Actions outside the agent’s permissions. Every agent has a manager, and the chain of managers ends at a person.
- Actions marked irreversible, and any others your rules select, until an authorized person approves that exact request. Approvals expire, a changed request needs a new one, and an agent can’t approve itself.
- Spending past a budget. Govern sets the money aside first and refuses what doesn’t fit.
- New requests while Govern is down or can’t save its record.
- The emergency stop revokes issued access at Govern’s checks and cancels AI responses in progress. Actions already sent to another system can’t be recalled, and offline workers stop within 30 seconds.
What it records
- Which agent asked, what it asked for, and whether Govern allowed it, refused it or held it for approval.
- Who approved or denied the request.
- The money set aside for it.
- Whether the change held, for actions with an outcome check.
- Activity reported in watch-only mode, labeled as observed.
Where it stops
- Calls that go around it. An agent that keeps its own key can still use that key.
- Watch-only actions, which are recorded but not blocked.
- Permissions already issued, which stay valid until they expire, at most 15 minutes after issue.
- What an agent says. Govern doesn’t verify its reasons or open its documents.
How Govern checks an amount
Before allowing a request, Govern sets aside the action’s registered cost ceiling. If there is no ceiling, a registered request contract must check the amount against what will be sent before Govern sets it aside and uses it for approvals or the task budget. If neither is available, Govern refuses any action registered as a payment and any request stating a nonzero amount. Other actions that omit the amount or state zero count as unpriced: every money threshold in their rule applies. This limits commitments made through Govern. It does not cap the provider’s final bill or spending outside Govern.
| Where the agent runs | What goes through Govern |
|---|---|
| Claude and Codex agents in Operate | Model calls and connected tool calls, from a sandbox with no provider key. |
| HTTP agents in Operate | Tool calls, and model calls once an administrator connects them. |
| Agents you run elsewhere, such as Microsoft Copilot Studio, Azure AI Foundry, LangGraph, CrewAI or your own runtime | Only what they send through Govern. |
| Agents on vendor platforms | Agents on Amazon Bedrock Agents, Bedrock AgentCore, Vertex AI Agent Engine and Salesforce Agentforce can also be registered. Govern covers their tool calls only, because their model calls stay with the vendor. |
Govern model calls and tool calls.
Send your agents’ AI model calls to Govern instead of straight to the provider, and connect their tools through Govern too (as an MCP server). Each call gets the same checks and a sealed record, like any other action. If your agent framework lets you set the address and key it uses, it needs no extra AAES software.
Model gateway
Govern accepts calls in the Anthropic and OpenAI formats. It checks each call, sets its cost aside in a model budget, and passes it on with a provider key the agent never sees. When a call reaches the budget limit, the budget can ask a person to approve it.
Model-source support depends on the API and tools your agent uses. Check protocol compatibility.
AAES Operate and AAES Govern use your organization’s model-provider accounts. Govern picks the provider from the model name. Live calls through the gateway have passed for 8 sources, some only for certain call formats; the rest are tested against recorded responses or not verified by AAES. The Models page shows each result. Test each source with your own agents before you rely on it. See all 29 model sources.
MCP bridge
Add Govern to your agent as a tool server (remote MCP server). Each tool matches an action registered in Govern, so permissions, spending limits and approvals apply to every tool call. When a call needs approval, a person decides and the agent tries again.
Capability packs
The capability library has 95 reviewed application packages, and 13 more are in review. Each pack defines the actions an agent can request.
Metering
Govern counts governed decisions, not model calls, in a signed usage report you can check offline. The report doesn’t set your price.
Bring your agent identities.
Agent identities from your directory
AAES Govern can import your existing agent accounts from Microsoft Entra ID (service principals and managed identities) and from Okta service apps. An administrator reviews a preview before anything is applied. Each imported identity becomes a registered agent with one accountable person as its sponsor. If Govern cannot match exactly one owner to a registered person, the administrator names the person in the preview; Okta identities always need this.
An import grants no roles, credentials, tasks or capabilities. When a directory sync reports an identity as disabled, Govern deactivates the agent. Imports are previewed before application; test the configured directory connection before relying on it.
Keys, records and outcomes.
Keys
A permission from Govern covers one task and doesn’t change the key to the target system. With brokered execution, Govern makes the allowed call itself, so the agent never gets that key. Review the control boundaries.
The dated capability matrix shows which ways of handling agent keys are built, tested only in the lab, or refused.
Records and offline verification
Check exported AAES Govern records for integrity without connecting to an AAES service. The offline check uses a public key you trust separately to confirm that exported records haven’t been changed. It can’t prove that every action was captured or that an action actually ran. A production install requires an independent witness or timestamp service, which confirms the signed summary of the record log.
The verification guide explains what an exported record does and does not establish.
Outcome checks
For actions with an outcome check, AAES Govern reads the target back after the action runs and seals the result as confirmed, mismatch or unknown. It stores field names and digests, not the values. A mismatch raises an alert, and an unknown result never counts as success.
The first checks cover ServiceNow incidents and Zendesk tickets that an agent creates or updates.
Deploy and evaluate on your stack.
Deployment
You can run AAES Govern as a single-tenant deployment on your infrastructure, controlling keys, storage and availability, or have AAES host it for you.
- One server. Docker Compose.
- Kubernetes. A standard install, tested on Google Kubernetes Engine and in CI.
- Offline or on-premises. Your own on-premises network, including networks cut off from the internet.
AAES-hosted production runs across zones for high availability. Customer-hosted high availability requires a cluster that can power machines off through an API. Both deployment types use backup and restore.
Production also needs HashiCorp Vault, an external witness or timestamp authority, and your identity provider. See what a production install needs.
Pricing
Start with a free workflow assessment. A paid evaluation is then scoped and priced for your workflow. Operate and Govern are priced in headcount bands, from under 500 employees to 300,000, and we quote larger organizations individually. Users and agents are unlimited. AI model charges are separate. Production pricing is shared on request. Operate and Govern are sold separately or together. Installation is assisted. Training and enhanced support are optional and agreed separately.
Headcount bands: under 500 employees, 500 to 2,499, 2,500 to 9,999, 10,000 to 24,999, 25,000 to 49,999, 50,000 to 99,999, 100,000 to 199,999, and 200,000 to 300,000. Above 300,000 employees, we quote by scope.
Contact us to request production pricing.
Assurance status
AAES has no independent certification or assessment. No SOC 2 report or penetration-test report exists. AAES has a documented scope and test plan for an independent penetration test.
The evaluation package runs in your environment and comes only with a paid, scoped evaluation.
Read the evaluation guide, or browse the evidence library.
Build with AAES Govern.
Read the Govern developer documentation for this product's integration scope, interface contracts and current release status. See the combined integration boundary when you use both products.
Frequently asked questions
Do we create agents in AAES or connect our existing agents?
Both. Create and run agents in AAES Operate. Connect an existing agent to Operate to give it work, and to Govern to control its actions and model calls. The agent keeps running where it runs today and can route its model calls through Govern with its own gateway credential. The Connected agents guide, supplied with the evaluation package, lists the connection steps. Each agent registered in Govern has an accountable manager whose chain of managers ends at a person. Govern governs actions routed through its controlled paths.
AAES Operate and AAES Govern use your organization’s model-provider accounts. Agents you create in Operate use those accounts. Existing agents connected to Govern can include Microsoft Copilot Studio, Azure AI Foundry, frameworks such as LangGraph, CrewAI and the OpenAI Agents SDK, and your own hosted runtime. Copilot Studio agents stay in your Microsoft environment.
Does this replace our identity provider?
No. Keep employee access in Microsoft Entra, Okta, or your existing identity provider. Name the agent’s first approver in AAES Govern. A directory integration can supply additional approvers where required; live tenant compatibility is validated during a paid evaluation.
Which actions require approval by an authorized person?
Actions registered as irreversible always require an authorized person’s approval on enforced paths. Your policies can require approval for other actions too. An action approval covers the exact request and expires; changing the request means asking again.
An access approval instead grants a specific permission for a limited scope and lifetime. Eligible manager agents can approve where policy permits, but cannot provide a required human approval. Agents cannot approve themselves.
Can we observe before we enforce?
Yes. Set an action to watch only, and Govern records the activity reported to it, labeled as observed. If the agent keeps keys it can use directly, Govern can’t stop those calls.
Switching to enforcement means setting up and testing how the agent’s keys go through Govern. Changing a setting isn’t enough.
What happens if AAES Govern is unavailable?
If AAES Govern goes down, or can’t save its record of a decision, requests that need a new decision are refused wherever Govern enforces. Govern issues no new permission and makes no call for the agent on that path. This is called fail closed.
Permissions Govern already issued stay valid until they expire, at most 15 minutes after issue. An outage doesn’t cancel work that was already approved or undo its effects, and it doesn’t stop actions that are watch-only or that don’t go through Govern. Employee access stays with your identity provider. Deploy on your infrastructure or have AAES host Operate and Govern. 99.9% monthly uptime applies to deployments hosted by AAES under our service level agreement. For a deployment you host, the agreement covers our response times.
Does AAES Govern see everything my agents do?
No. Govern checks only supported actions sent through it. Agents that can use their own keys directly can go around it, and actions set to watch only aren’t stopped.
Do I need AAES Operate to use AAES Govern?
No. Govern connects agents you already run, such as Microsoft Copilot Studio, Azure AI Foundry, an agent built with LangGraph, CrewAI or the OpenAI Agents SDK, or your own runtime. AAES Operate is a separate product for creating, running, and coordinating agents on your infrastructure or hosted by AAES.
How is AAES Govern priced?
Start with a free workflow assessment. A paid evaluation is then scoped and priced for your workflow. Operate and Govern are priced in headcount bands, from under 500 employees to 300,000, and we quote larger organizations individually. Users and agents are unlimited. AI model charges are separate. Production pricing is shared on request. Operate and Govern are sold separately or together. Installation is assisted. Training and enhanced support are optional and agreed separately.
Is AAES Govern certified, and does it have a SOC 2 report?
No. AAES has no independent certification or assessment. No SOC 2 report or penetration-test report exists. AAES has a documented scope and test plan for an independent penetration test. The evaluation package runs in your environment and comes only with a paid, scoped evaluation.