---
title: How do you govern model access through SAP generative AI hub?
url: https://opencontroller.lyzr.com/blog/sap-generative-ai-hub-governance
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# How do you govern model access through SAP generative AI hub?

**Route model calls through the orchestration service in SAP generative AI hub, and turn on its built-in modules: content filtering on input and output, data masking for personal data, and grounding for answers based on your documents. Scope which models each application can reach with orchestration deployments in resource groups. Then govern the agents themselves, which orchestration does not do.**

SAP generative AI hub runs in SAP AI Core. Its orchestration service gives applications one API across multiple foundation models. That single entry point is what makes governance possible: every module runs on every call that goes through it.

## What modules does SAP orchestration provide?

| Module | What it does | Governance use |
|---|---|---|
| Templating | Prompts with placeholders filled at run time | Standard, reviewed prompts |
| Grounding | Adds content from your documents (vector or URL repositories) | Answers based on approved sources |
| Input content filtering | Screens prompts before the model sees them | Block harmful or abusive input |
| Data masking | Anonymizes or pseudonymizes sensitive data before the model call | Keep personal data out of prompts |
| Output content filtering | Screens model responses | Block harmful output |
| Translation | Translates input and output | Consistent filtering across languages |

## Which content filters can you use?

The SAP generative AI hub SDK exposes two filter providers:

- **Azure Content Safety.** Four categories: hate, sexual, violence and self-harm. Each takes a threshold of 0, 2, 4 or 6, from `ALLOW_SAFE` to `ALLOW_ALL`. Lower is stricter.
- **Llama Guard 3 8B.** Fourteen categories, each off by default: violent crimes, non-violent crimes, sex crimes, child exploitation, defamation, specialized advice, privacy, intellectual property, indiscriminate weapons, hate, self-harm, sexual content, elections and code interpreter abuse.

Because Llama Guard categories are off by default, a filter object with nothing enabled gives no protection. Write down which categories each use case needs. For a finance or HR assistant, "specialized advice" and "privacy" are usually worth enabling.

## How does SAP data masking work?

Data masking uses SAP Data Privacy Integration. You choose a method:

- **Anonymization** replaces values irreversibly. Use it when the model never needs the real value.
- **Pseudonymization** replaces values reversibly, so the response can be unmasked before it returns to the user.

You also choose entity types. The SDK lists 25, including person, email, phone, address, IBAN, credit card number, national ID, passport, SSN and SAP internal IDs. Start with identifiers and contact details, then add more as you test. Our guide on [redacting PII at the gateway](/blog/how-to-redact-pii-in-llm-prompts-at-the-gateway) explains how to test masking quality.

## How do you control which models applications can use?

Orchestration runs as a deployment inside a resource group. Each resource group has a default deployment, and you can create your own to customise the available models. A practical setup:

1. **One resource group per environment or business unit**, so access and cost are separate.
2. **Custom deployments** that expose only approved models for that group.
3. **A standard orchestration config per use case** that turns on filtering and masking. Store it in version control.
4. **A rule that applications call orchestration**, not model provider APIs directly. Check for direct provider keys during [shadow AI discovery](/blog/how-to-find-shadow-ai-agents).

## What does SAP generative AI hub not cover?

- **Traffic that skips orchestration.** Direct calls to a model provider get none of these controls.
- **Agent inventory and ownership.** Orchestration governs model calls, not which agents exist, who owns them or what tools they hold. Joule agents need their own review; see [governing SAP Joule agents](/blog/govern-sap-joule-agents).
- **Non-SAP agents.** Agents on Bedrock, Copilot Studio or Agentforce sit outside it.
- **Behaviour over time** and a fast way to stop one agent.

That is the difference between a model gateway and a control plane. Our explainer on [LLM gateway vs agent gateway vs control plane](/blog/llm-gateway-vs-agent-gateway-vs-control-plane) sets out the layers, and [governing agents across platforms](/blog/how-to-govern-ai-agents-across-platforms) shows how they combine.

## How Opencontroller does this

Opencontroller connects to SAP and other agent platforms, including Salesforce, Databricks and HubSpot, and keeps one governed registry of agents, versions, tools and MCP servers. Posture detects risky permissions and configurations. Live groundedness and task-adherence guardrails and LLM judges score production traffic, and policy as code applies across the estate. Keep SAP orchestration for SAP model calls; use Opencontroller for the [control plane](/blog/what-is-an-ai-agent-control-plane) across all of them. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Which content filters does SAP orchestration support?**
Azure Content Safety with four categories and Llama Guard 3 8B with fourteen, applied to input, output or both.

**What is the difference between anonymization and pseudonymization in SAP data masking?**
Anonymization is irreversible. Pseudonymization is reversible, so responses can be unmasked.

**Does SAP generative AI hub stop direct calls to model providers?**
No. It only governs calls routed through orchestration. Block or detect direct provider keys separately.

## Key facts

SAP generative AI hub (in SAP AI Core) provides an orchestration service that gives one API to multiple foundation models, with modules for templating, grounding, input and output content filtering, data masking and translation. Content filters include Azure Content Safety (hate, sexual, violence, self-harm; thresholds 0, 2, 4, 6) and Llama Guard 3 8B (14 categories including privacy, specialized advice, elections and code interpreter abuse, all off by default). Data masking uses SAP Data Privacy Integration with anonymization (irreversible) or pseudonymization (reversible) across 25 entity types such as person, email, phone, IBAN, credit card and national ID. Orchestration deployments live in resource groups, which scope which models an application can use. These controls apply only to traffic through orchestration. Opencontroller by Lyzr connects to SAP and other platforms to keep one registry of agents; pricing starts from $250K per deployment per year.

## Related

- [How do you govern SAP Joule agents?](https://opencontroller.lyzr.com/blog/govern-sap-joule-agents.md)
- [How do you redact PII from LLM prompts at the gateway?](https://opencontroller.lyzr.com/blog/how-to-redact-pii-in-llm-prompts-at-the-gateway.md)
- [LLM gateway vs agent gateway vs AI control plane: what's the difference?](https://opencontroller.lyzr.com/blog/llm-gateway-vs-agent-gateway-vs-control-plane.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
