---
title: How do SAP AI Core resource groups isolate AI agent workloads?
url: https://opencontroller.lyzr.com/blog/sap-ai-core-resource-groups-for-agent-isolation
published: 2026-11-17
updated: 2026-11-17
author: Lyzr
---

# How do SAP AI Core resource groups isolate AI agent workloads?

**In SAP AI Core, a resource group is the isolation boundary for AI workloads: deployments and executions run inside a resource group, backed by a Kubernetes namespace, and every API call names its group in the `AI-Resource-Group` header. A tenant can have at most 50, so map them deliberately to environments and teams rather than creating one per agent.**

SAP AI Core hosts the generative AI hub, which many SAP-centric agents use to reach foundation models. How you lay out resource groups decides who can see which deployments, how cleanly you can split costs, and how easy it is to retire an agent later.

## What does a resource group isolate?

Facts from SAP's documentation and practitioner guides:

| Aspect | What happens |
|---|---|
| Workload scope | Models, deployments and executions run within a resource group context |
| Technical isolation | Resource group isolation is based on a Kubernetes namespace |
| API scoping | Calls carry an `AI-Resource-Group` header naming the group |
| Creation | Through the AI Core API (`/v2/admin/resourceGroups`), not the BTP Cockpit |
| Visibility | Without a resource group, model deployments are not visible in the hub |
| Limit | Maximum of 50 resource groups per tenant |
| Metering | Usage reports include resource-group-level consumption |
| Secrets | Secrets can be scoped to a resource group; generic secrets can also be created tenant-wide |

Many setup guides simply create a group called `default`. That works for a proof of concept, but it puts every team and every agent in one boundary.

## How should you map resource groups to agents?

With a 50-group ceiling, one group per agent does not scale. A practical layout:

1. **Separate by environment first.** Development, test and production in different groups, so a test deployment can never be called by a production agent's credentials.
2. **Then by team or business domain** for production, for example finance agents and procurement agents. This lines up with resource-group-level metering for chargeback.
3. **Keep a shared group only for genuinely shared services,** such as one orchestration deployment used by many agents, and document who owns it.
4. **Name groups consistently,** for example `prd-finance` and `dev-finance`, so inventory and cost reports are readable.
5. **Scope secrets to the group** that needs them; reserve tenant-wide generic secrets for things every group really uses.

## What do resource groups not give you?

A resource group is a container. It does not by itself tell you:

- **Which agent made a call.** Several agents in one group look alike unless each passes its own identity and you log it.
- **Whether the agent behaved.** Groundedness, task adherence and multi-turn behaviour need evaluation and runtime checks.
- **What else the agent touches.** A Joule or custom agent may call SAP systems and non-SAP tools outside AI Core.
- **Anything about agents on other platforms** that share the same business process.

For SAP's model-level controls such as orchestration filtering and data masking, see [SAP generative AI hub governance](/blog/sap-generative-ai-hub-governance). For agents built in Joule, see [governing SAP Joule agents](/blog/govern-sap-joule-agents).

## A resource group checklist for agent teams

1. Inventory existing groups and what is deployed in each; look for everything sitting in `default`.
2. Decide an environment-by-domain layout that stays well under 50.
3. Move production agents out of shared or default groups.
4. Scope secrets per group and review any tenant-wide generic secrets.
5. Record, for each agent, the group and deployment it calls.
6. Use resource-group consumption in usage reports to allocate cost. Our guide to [allocating shared AI platform costs to divisions](/blog/allocate-shared-ai-platform-costs-to-divisions) covers the method.
7. When an agent is retired, remove its deployments and secrets, and the group if it is now empty.

For how SAP fits alongside agents on other clouds and SaaS platforms, see the [platform guides pillar](/blog/how-to-govern-ai-agents-across-platforms).

## How Opencontroller does this

Opencontroller connects to SAP as a third-party agent platform and puts SAP agents in the same governed registry as agents on AWS, Azure, Google Cloud, Salesforce, Databricks and HubSpot, with versions, tools and MCP servers. Posture flags risky permissions and configurations, LLM judges score production traffic, and policies with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001 apply across the estate. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**How many resource groups can an SAP AI Core tenant have?**
SAP documents a maximum of 50 resource groups per tenant.

**Do I have to send a resource group on every call?**
Yes. Calls to SAP AI Core, including inference against generative AI hub deployments, include the `AI-Resource-Group` header naming the group.

**Can resource groups separate costs by team?**
Yes, partly. SAP's usage reports include resource-group-level consumption, so a team-per-group layout supports chargeback. Per-agent cost still needs per-agent attribution.

## Key facts

- SAP AI Core resource groups are the isolation boundary for AI workloads; models, deployments and executions run within a resource group context; isolation is based on a Kubernetes namespace.
- Calls specify the group via the AI-Resource-Group header; groups are created through the AI Core API (/v2/admin/resourceGroups), not the BTP Cockpit; deployments are not visible in generative AI hub until a resource group exists.
- SAP limits resource groups to 50 per tenant; usage reports include resource-group-level consumption; generic secrets can be tenant-wide as well as resource-group scoped.
- Opencontroller by Lyzr connects to SAP and governs SAP agents in a cross-platform registry; pricing starts from $250K per deployment per year.

## Related

- [How do you govern model access through SAP generative AI hub?](https://opencontroller.lyzr.com/blog/sap-generative-ai-hub-governance.md)
- [How do you govern SAP Joule agents?](https://opencontroller.lyzr.com/blog/govern-sap-joule-agents.md)
- [How do you allocate shared AI platform costs to divisions?](https://opencontroller.lyzr.com/blog/allocate-shared-ai-platform-costs-to-divisions.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
