Enterprises in regulated and business-critical environments cannot safely deploy AI agents in processes such as HR, procurement, meetings and ESG reporting without strong governance, traceability and evidence-based outputs. Current productivity suites such as Microsoft Copilot or Gemini for Workspace are tied to specific vendor ecosystems and typically operate on best-effort generation, while open frameworks such as LangChain or AutoGen provide orchestration primitives but not the governance, audit trail and conformity support needed for enterprise adoption in Europe.

The market need is therefore a vendor-neutral way to introduce specialised AI agents into real workflows while keeping control over policies, data sources, accountability and compliance with emerging EU AI Act obligations. Without such a layer, organisations risk hallucinated decisions, unverifiable recommendations, vendor lock-in and limited acceptance by legal, compliance and operational stakeholders.

AGENTIC-WORKBENCH will provide a policy-governed orchestrator that routes tasks to specialised AI agents according to enterprise rules, context, user permissions and process requirements. Instead of relying on a single vendor suite or an ungoverned open agent framework, it will offer a neutral orchestration layer that can connect different agents, enterprise data sources and sector-specific systems.

A key innovation is the fail-closed verification layer: outputs without a verified basis in authorised enterprise sources are blocked rather than passed on to users. The framework will maintain traceability of agent actions and evidence, and will include tools to generate conformity artefacts supporting EU AI Act compliance. The approach also foresees sustainability-aware routing and the extension of the framework with clinical, legal, procurement and other domain agents where supported by consortium partners.

The consortium will define the reference architecture for a vendor-neutral multi-agent enterprise workbench, including the policy model, orchestration logic, evidence handling, audit trail and interfaces to enterprise systems and specialised agents.

Partners will develop and integrate domain agents and sector-specific integration targets, including use cases such as meetings, HR, procurement, ESG and other regulated or business-critical workflows. The project will implement the fail-closed verification layer, source-grounding mechanisms, conformity evidence generation and sustainability-aware routing capabilities.

The framework will be validated through pilots across verticals, with each pilot assessing practical integration, trustworthiness, usability, governance, traceability and EU AI Act conformity support. The work will also cover testing, benchmarking, documentation and exploitation of the resulting framework for European enterprise deployment.

• Comtrade SI d.o.o. — Project coordinator