SAP AI Agent Hub Explained: A Register for Digital Workers

In 30 seconds

What it is One register for every agent, model and tool, each with an owner and a status.

You use it when You are about to put a second or third AI agent into production and nobody can yet list them all with an owner.

Guide covers AI Agent Hub

Every agent gets a register card: who owns it, what it does, what it touches, how risky it is, and whether someone has verified it.

40-second version of this explainer, no sound.

Part 6 of 26 in SAP Business AI, Explained.

Once a company runs more than a handful of AI agents, audit asks a plain question: which agents do we have, and who answers for each one? Most teams can’t reply straight away. SAP AI Agent Hub is SAP’s answer.

It’s the register where every AI agent, model and tool connection in your company gets a name, an owner, a risk rating and a yes-or-no approval.

The short version

It’s one inventory and governance place for AI agents, large language models (LLMs) and Model Context Protocol (MCP) servers. It sits inside SAP LeanIX, and SAP News describes it as mapping them to capabilities and owners.
It isn’t limited to SAP. The product page calls it a “vendor-agnostic command center”, with automatic discovery from SAP, Microsoft, Google, AWS (Amazon Web Services), Databricks and ServiceNow.
Every discovered agent starts as unverified until a reviewer marks it verified or rejected. SAP says verification can be enforced at runtime.

Think of it as HR for digital staff

Every new joiner at a company gets an HR file: who they are, who their manager is, what job they do, which buildings their badge opens. The badge comes after the file, not before.

The hub does the same for digital employees. It doesn’t do the work or hold the badge reader. In the hospital picture from post 02 it’s the board’s licence register. The board, meaning your governance people, makes the decisions, and the register records them so the security desk knows whom to let in.

From discovery to retirement: what the hub tracks

Discover. You connect the platforms where agents run. The hub surfaces agents from SAP and third-party systems, and anything discovered waits in a review inbox before it enters your inventory.
Describe. Each agent, tool server and model gets its own record, which SAP LeanIX calls a fact sheet. It says which business job the agent supports, where in the process it works, which team owns it and who supplies it.
Verify. A reviewer opens the verification flow, checks the attributes your administrators picked as criteria, and records verified or rejected. The status shows on the fact sheet for everyone.
Enforce and observe. According to SAP’s product page, only verified MCP servers can be called in production workflows, and access is revoked when the status changes. The same page lists observability such as session health, goal completion rate and tool call correctness.
Retire. SAP’s documentation describes a structured lifecycle from planning through decommissioning, so an agent’s record ends with a deliberate retirement rather than an old login left running.

The register card for one agent

When I review an agent, I want eight answers on one card. These are my plain-English labels, not SAP’s screen labels.

Card item	The question it answers	Where the idea comes from
Owner	Who is accountable when it goes wrong?	Organisation fact sheets carry business and technical ownership (SAP Help)
Capability	Which business job does it support?	Agents link to business capabilities (SAP Help, SAP News)
Systems touched	Which applications and data can it reach?	Agents are a subtype of application and link to the landscape (SAP Help)
Model	Which large language model does the reasoning?	AI model fact sheets, linked to providers (SAP Help)
MCP servers	Which tool connections does it use?	MCP server fact sheets, which can be verified too (SAP Help)
Risk tier	How bad is a wrong action?	AI risk rating on the product page, automated risk classification in SAP News
Autonomy level	Does it answer, propose or act?	Autonomy level is named as a fact sheet attribute in the SAP Community Agent Mesh post
Status	Has someone verified it?	Unverified, verified or rejected (SAP Help)

For autonomy I use the ladder from post 02: L0 Inform, L1 Answer, L2 Propose, L3 Act with approval, L4 Act then report, L5 Fully autonomous.

Getting the payment agent verified at NextSAP

The example company is NextSAP Utilities, the fictional Dubai utility used across this site’s SAP content. It serves about 5 million customers across electricity, water, gas and district cooling, has a shared services team in India, and runs SAP S/4HANA with SAP Utilities (IS-U) and contract accounting, which SAP calls Contract Accounts Receivable and Payable (FI-CA), plus SAP SuccessFactors and SAP Ariba.

Omar, the finance controller, has a problem with incoming payments. The shared-services team clears about 140 unclear customer payments every working day, so about 3,000 a month worth around AED 50 million (UAE dirham). These are the ones the bank statement couldn’t match on its own, so they wait on clarification account 1190100. Marina Bay Café LLC pays several bills in one transfer; Al Noor Property Management pays for many premises in one go, often with a missing or wrong reference. So Omar’s team built a payment clarification agent in Joule Studio for FI-CA. It reads the payment, searches open items and proposes a match. A person still posts.

Priya, the IT (information technology) Platform Owner, connects the agent platforms to the hub, and the payment clarification agent shows up in the discovery inbox. Hana, the AI Governance Lead, accepts it into the inventory, where it sits as unverified. Its card looks like this:

Owner: Omar, finance controller. Technical owner: Priya.
Capability: clarify and match incoming payments.
Systems touched: SAP S/4HANA, FI-CA contract accounting, read only.
Model: one approved large language model, named on the card.
MCP servers: two tool connections, open-item search and payment lookup.
Risk tier: medium. Money is involved, but the agent can’t post.
Autonomy level: L2 Propose.
Status: unverified.

Hana’s criteria ask for an owner, risk tier, autonomy level, systems touched and test evidence. Omar’s team runs 200 historical payments through the agent, and 184 proposals match what a clerk chose, a 92 percent hit rate. Then Daniel from Internal Audit asks the question I’d ask too: can the agent post? It can’t, and the tool list proves it. Hana marks it verified.

Two weeks later the team adds a third MCP server that could create a clearing document. That changes what the agent can do, so Hana sets the status back for re-review. My advice is to treat any new tool, new model or higher autonomy level as a fresh verification.

This isn’t only a NextSAP pattern. SAP’s Joule Agents page describes Uniper building a custom procurement agent in Joule Studio, and Bosch using a classification agent. Each one deserves a card.

Where the hub fits next to Joule Studio, the gateway and Cloud ALM

The hub is one piece of a bigger picture, alongside Joule Studio, the agent gateway and SAP Cloud ALM (Application Lifecycle Management), SAP’s tool for running and monitoring cloud solutions.

Your question	Where it is answered	Source
How do we build a custom agent?	Joule Studio	SAP Joule Agents page
Where do people use agents day to day?	Joule Work	SAP Business AI page
Which agents exist, who owns them, are they approved?	SAP AI Agent Hub, part of SAP LeanIX	SAP Help, SAP News
Who may call what, at runtime?	The enterprise agent gateway: identity, access and policy	SAP Business AI page
Where could an agent add value in a process?	SAP Signavio Process Consulting Agent	SAP News
What did each agent actually do after go-live?	SAP Cloud ALM agent monitoring	SAP News

The limits of the hub

Beyond the inventory and the verification record, SAP’s documentation mentions dashboards on portfolio, approval coverage and risk exposure. The limits matter just as much:

It won’t choose your owners or your criteria. The SAP Community Agent Mesh post says the agent owner, the verification criteria and the autonomy threshold for human approval stay with the enterprise.
It doesn’t build agents. That’s Joule Studio or another platform.
It doesn’t replace your authorisation concept in SAP Cloud ERP. Roles still decide what an agent’s identity can do.
And not every feature is available everywhere today. SAP says some arrive in a phased approach.

Explain it to your CFO

It’s one list of every digital worker we run, with the manager who answers for each, and none of them touches money until a named person has checked and approved it.
When something goes wrong, we can see in minutes which worker did it, who owns it and what it was allowed to do.

Three calls your governance lead has to make

Who verifies an agent? I’d use a named AI governance lead, with the business owner signing off. SAP’s documentation says admins and members with a responsible or accountable subscription can verify by default, so narrow that if you need to.
What are your verification criteria? A sensible default is named business owner, risk tier, autonomy level, systems touched, tool list and test evidence. An agent without an owner doesn’t get verified, and it doesn’t go to production.
Do you enforce status at runtime? For production my default is yes. Unverified agents and MCP servers stay in sandbox.

Go deeper

Technical detail for architects and developers

Meta model. SAP’s Key Concepts page explains that the hub is based on the SAP LeanIX meta model. AI agents are a subtype of the application fact sheet, MCP servers a subtype of the interface fact sheet, and AI models a subtype of the IT component fact sheet. Business capability, business context, organisation and provider fact sheets add the ownership context, and as subtypes they inherit the normal relationship and governance model.

Verification. SAP Help describes verification as a formal review for agents, MCP servers and AI models. The flow only appears when the verification badge is enabled for that asset type. For agents discovered from an SAP environment, verification applies to the logical agent rather than each deployment, so production and sandbox instances share a status. It’s also separate from the SAP LeanIX quality seal, which tracks data freshness. An asset can be verified while its quality seal is broken.

Runtime enforcement. SAP’s product page says verified status creates a trust layer that can be enforced at runtime, and that only verified MCP servers can be called in production workflows. The SAP Community Agent Mesh post adds that Joule Studio, SAP Integration Suite and the agent gateway need to enforce the status, and that the agent identity in SAP Cloud Identity Services is created during registration. Confirm these points in your own landscape.

Risk and compliance. SAP News describes a new AI Governance Assistant (check availability) that brings EU AI Act and NIST (US National Institute of Standards and Technology) compliance intelligence into the inventory, with automated risk classification and reassessment when a system changes. The Agent Mesh post mentions an EU AI Act extension in SAP LeanIX that classifies the risk tier.

Prerequisites and status. SAP Help says workspaces created before 11 May 2026 need the AI governance extension activated, and the documentation notes phased availability. The Agent Mesh post reports that the hub documentation still carried an Early Adopter Care notice in October 2026. Check the current status in the Joule capability watch.

Words you’ll hear

AI agent: software that uses a model to decide its next step and then takes it, such as searching or proposing a posting.
Large language model (LLM): the text model doing the reasoning behind an agent.
Model Context Protocol (MCP): an open standard for connecting tools to agents. An MCP server is the plug that exposes tools.
Fact sheet: the basic record in SAP LeanIX. One fact sheet describes one thing, whether that’s an agent, a capability or an organisation.
Verification: a recorded review that marks an agent, MCP server or model as verified or rejected. Unverified is the default.
Agent gateway: the runtime checkpoint for identity, access and policy. Think of it as the building’s security desk.
Risk tier: how much harm a wrong action could cause, usually low, medium or high.

Keep going

Independent explainer, not affiliated with or endorsed by SAP SE. SAP, Joule and other SAP product names are trademarks of SAP SE.

Frequently asked questions

Is SAP AI Agent Hub only for agents built by SAP?

No. SAP describes it as vendor-agnostic, and its product page lists automatic discovery from SAP, Microsoft, Google, AWS (Amazon Web Services), Databricks and ServiceNow.

Where does SAP AI Agent Hub live?

It is part of SAP LeanIX. SAP's help documentation lists it under SAP LeanIX, and SAP News describes it as inventorying agents, large language models and MCP servers mapped to capabilities and owners.

What happens to an agent when it is first discovered?

It starts as unverified. A reviewer checks it against the criteria your organisation has configured and marks it verified or rejected. The status is visible to everyone in the workspace.

Does the hub decide who owns an agent or how much autonomy it gets?

No. The hub records those decisions. Choosing owners, setting verification criteria and deciding when a human must approve stay with your organisation.

Is every feature available today?

SAP's documentation says some features become generally available in a phased approach and may not yet be in your region. Check the current status in the Joule capability watch on this site.