AI Agent Governance 2026: AWS, Microsoft and Anthropic Establish Enterprise Controls
AWS Bedrock AgentCore, Microsoft Foundry and Anthropic Enterprise offer production-ready controls for autonomous AI agents for the first time. The clock is ticking: the EU AI Act's high-risk obligations take effect in December 2027.
AI agents increasingly act autonomously in enterprises, yet governance infrastructure is missing in most organisations. Only 43 percent have a centralised AI data gateway. AWS, Microsoft and Anthropic released new governance controls in March 2026: AWS uses policy gateways with the Cedar language, Microsoft integrates agents into identity management via Entra Agent IDs, and Anthropic delivers a Compliance API with real-time monitoring. Implementation is urgent for European businesses, as only 5 percent use AI for structural transformation, and the EU AI Act's high-risk obligations for Annex III systems take effect on 2 December 2027.
Why AI Agent Governance Is Now Mandatory
AI agents increasingly act autonomously in enterprises: they call tools, access data and execute workflows without a human approving every step. This autonomy is the core of their value, but also the core of the problem.
A red-team study from February 2026, involving researchers from Harvard, MIT, Stanford and Carnegie Mellon, demonstrated how serious the situation is: AI agents independently deleted emails, exfiltrated social security numbers and triggered unauthorised actions because no effective stop mechanism existed.
The EU AI Act 's high-risk obligations for Annex III systems take effect on 2 December 2027. Organisations without documented control mechanisms face fines of up to 15 million euros or three percent of global annual turnover. The three largest AI providers responded in March 2026, delivering production-ready governance tools for the first time. Google followed in April with governance as one of four pillars of the Gemini Enterprise Agent Platform.
AWS Bedrock AgentCore Policy: Control Outside the Code
On 3 March 2026, AWS made policy controls for Amazon Bedrock AgentCore generally available across 13 AWS regions. The core principle: security rules are defined outside the agent code and checked by the AgentCore Gateway before tool execution.
An agent may not call a tool without explicit permission. The default behaviour is denial. Security teams can adjust rules without touching the agent code. Developers can describe rules in plain English, and AgentCore automatically translates them into Cedar and checks for logical consistency.
AWS separates agent logic from security rules: developers write code, security teams write policies. The Cedar language enables thousands of policy decisions per second at millisecond latency.
AgentCore is also available as an MCP server, enabling policy authoring directly in AI-powered code editors. Every policy decision is logged via CloudWatch for compliance and audit purposes. The AgentCore SDK has exceeded two million downloads in five months since its preview.
Microsoft Foundry: Agents in the Identity System
Microsoft Foundry takes a fundamentally different approach from AWS: instead of a separate policy gateway, Microsoft treats AI agents as manageable entities in the existing enterprise identity system. Every Foundry agent receives its own Microsoft Entra Agent ID, the same mechanism used for human users.
This allows IT administrators to manage agents with familiar tools: Conditional Access, Identity Protection, Identity Governance and network policies. KPMG already has the system in production. KPMG's Global Head of Audit Innovation and AI stated that governance and observability in Azure AI Foundry deliver exactly what KPMG needs in a regulated industry.
Foundry Control Plane Details
Limitation for EU organisations: Hosted Agents in Microsoft Foundry are currently only available in the North Central US region. European organisations with data residency requirements need to plan their own solution architectures.
Anthropic: Compliance API and Admin Controls
Anthropic significantly expanded Claude's enterprise governance capabilities with its 24 March 2026 updates. The new Compliance API gives compliance teams programmatic real-time access to usage data and customer content, without manual exports or periodic reviews. The Enterprise Analytics API delivers usage and engagement data at organisation and daily level.
Compliance API
Programmatic real-time access to usage data for continuous monitoring and audit trails
Enterprise Analytics
Usage and engagement data at organisation and daily level for data-driven decisions
Admin API (25 Endpoints)
Programmatic organisation management with six granular roles
Managed Policy Settings
Tool permissions, file access and MCP server configurations centrally deployable
For developer teams, the Admin API with 25 endpoints enables full automation of organisation management. Six roles (user, developer, billing, admin, claude_code_user, managed) provide role-based access control. Granular spending controls at organisation and user level prevent uncontrolled costs. Bring Your Own Key (BYOK) for custom encryption keys is planned for H1 2026.
The Three Approaches Compared
All three providers address the same problem but with different philosophies. AWS separates policy from code, Microsoft integrates agents into the identity system, and Anthropic delivers APIs for monitoring and compliance.
| Criterion | AWS AgentCore | Microsoft Foundry | Anthropic Enterprise |
|---|---|---|---|
| Governance Approach | Policy gateway before tool execution | Identity-based (Entra Agent ID) | API-based monitoring and compliance |
| Policy Language | Cedar (open, human-readable) | Azure Policy + Conditional Access | Managed Policy Settings (JSON) |
| Default Behaviour | Deny by default | Identity-dependent | Role-based |
| Audit Trail | CloudWatch logging | Azure Monitor + Purview | Compliance API + Analytics API |
| EU Regions | 13 regions (incl. EU) | Hosted Agents US only (North Central) | EU hosting available |
| Status | GA since 3 March 2026 | GA (Control Plane) | GA since 24 March 2026 |
For European organisations already using Microsoft 365, Foundry offers the lowest barrier to entry. Multi-cloud organisations benefit from AWS AgentCore as a model-agnostic solution. Anthropic is particularly relevant for teams using Claude as their primary LLM and looking to automate compliance evidence for the EU AI Act.
The AI Adoption Paradox
According to Deloitte, some European markets show the highest AI adoption rates among the 14 countries studied while simultaneously ranking last in structural transformation. This is not a contradiction but a diagnosis: organisations use AI for efficiency, not for change.
Deloitte's finding is clear: adoption alone does not create value. Only 2 percent of CEOs anchor AI responsibility at board level, the last place among all 14 countries studied. 84 percent have not yet adapted roles and processes to AI. 28 percent invest more than 20 percent of their technology budget in AI, but 41 percent remain below ten percent.
High AI adoption without structural transformation creates the illusion of progress without delivering real business value.
Deloitte, "The ROI of AI", March 2026EU AI Act: What to Implement by December 2027
Updated October 2026. The first version of 26 March 2026 named 2 August 2026 as the deadline for the high-risk obligations and treated the postponement as a proposal. That is out of date. Regulation (EU) 2026/1744 has applied since 27 July 2026 and moves the obligations under Annex III to 2 December 2027, and those for AI in products under Annex I to 2 August 2028. The German KI-MIG has been in force since 29 July 2026. We have corrected the heading, the dates and the FAQ, but not updated the survey figures.
The EU AI Act's high-risk obligations take effect on 2 December 2027 under Regulation (EU) 2026/1744. Organisations deploying AI systems in areas such as hiring decisions, credit scoring or safety-critical applications must demonstrate comprehensive compliance by that date.
Obligations by December 2027
- Establish and document quality management systems
- Create risk frameworks for all AI systems
- Complete technical documentation
- Finalise conformity assessments
- Apply CE marking
- Complete EU database registration for high-risk systems
- Conduct data protection impact assessments under GDPR Article 35
The Digital Omnibus has moved the dates: Regulation (EU) 2026/1744 has applied since 27 July 2026, the Annex III obligations apply from 2 December 2027, and those for AI in Annex I products from 2 August 2028. The German Federal Network Agency has established an AI Service Desk as a first point of contact for SMEs. The AI Market Surveillance and Innovation Act (KI-MIG) has been in force since 29 July 2026, with the Federal Network Agency as market surveillance authority.
Challenges and Risks
The new governance tools from cloud providers do not solve the problem alone. Governance cannot reside solely in IT. It is a cross-functional task involving legal, compliance, HR, data science and executive leadership.
For organisations that have not yet conducted an AI inventory, even the December 2027 date is tight. According to Deloitte, pulling the investment after six months means losing the entire stake. Real ROI requires 24 to 36 months of continuous investment.
Further Reading
Frequently Asked Questions
AI agent governance encompasses the technical and organisational controls ensuring autonomous AI agents only perform authorised actions. From 2 December 2027, the EU AI Act's high-risk obligations for Annex III systems require documented control mechanisms. AWS, Microsoft and Anthropic have released new governance tools that make these requirements technically implementable.
AWS Bedrock AgentCore Policy defines security rules outside the agent code. The AgentCore Gateway checks every tool request against policies before execution. Policies are written in Cedar, an open, human-readable and machine-evaluable language. The default behaviour is denial: an agent may only call a tool when explicit permission exists.
Microsoft Foundry takes an identity-based approach: every AI agent receives its own Entra Agent ID and is managed like a user in the identity system. IT teams use familiar tools such as Conditional Access and Identity Governance. AWS AgentCore uses a policy gateway that checks requests before tool execution. Both approaches are complementary and address different governance aspects.
Anthropic offers a Compliance API for real-time access to usage data, an Enterprise Analytics API for organisation-level metrics, and an Admin API with 25 endpoints for programmatic management. These are complemented by six organisational roles, granular spending controls and centrally deployable Managed Policy Settings for tool permissions and file access.
Organisations deploying AI systems in high-risk areas must demonstrate quality management systems, risk frameworks, technical documentation and conformity assessments by 2 December 2027 (Annex III, Regulation (EU) 2026/1744). CE marking must be applied and EU database registration completed for high-risk systems. National regulators have established AI service desks as first points of contact for SMEs.
Deloitte found that some European markets have the highest AI adoption rates among the 14 countries studied, yet rank last in structural transformation. Only 5 percent of organisations use AI for business model evolution, and just 2 percent of CEOs anchor AI responsibility at board level. Organisations use AI for efficiency, not for change.
Let's collaborate
One contact, a team behind it
Thirty minutes on what you're working on, whether grid, metering, market processes, IT architecture or AI. Afterwards you'll know how we'd approach it and who would sit at the table.
Energy and measurement technology since 2010 · innobu GmbH since 2014 · Based in Schleswig-Holstein, Germany
Whoever pitches also delivers: whoever leads the first call stays your contact right through to the steering committee, with established partners in the project. Direct: info@innobu.com