AI
·
Integration
Content yes, names no. Helpdesk e-mail triage with SAP Generative AI Hub
Patrik Majer

Recently I was looking at a very common support problem: a shared helpdesk inbox with dozens of new e-mails every morning. Before anyone can help, someone has to read each e-mail and decide what it is, how urgent it is and which system it is about. That's a perfect job for an LLM.
The catch: every one of those e-mails contains personal data. Names, e-mail addresses, phone numbers. Sending them to an external model as they are is not something a data protection officer will sign off.
My first idea was a small app that calls the model provider's API directly. In this post I will show why that is a dead end, and how to do it properly with the orchestration service in SAP AI Core and Generative AI Hub, where personal data is masked before the model ever sees it.
I will not go into detail on how to provision SAP AI Core and SAP AI Launchpad. You can refer to this SAP tutorial to do so.
1. Why calling the model directly is a dead end
There are two traps.
Trap 1: every app holds a provider key. Each application that calls OpenAI, Anthropic or Google directly needs its own API key, its own contract and its own security review. There is no central place to see who calls which model, and when the model is deprecated, every app has to change.
Trap 2: personal data goes to the model as-is. The e-mail text is the prompt. Unless you build your own detection and masking of names, e-mail addresses and phone numbers, and then put the values back into the answer, all of it leaves the company with every call.
So with a direct call you get a working classifier, but not one you can run on real helpdesk data.
2. The solution: Orchestration in Generative AI Hub
Generative AI Hub in SAP AI Core gives you one BTP-authenticated endpoint to models from several providers. The orchestration service wraps the model call in a pipeline of modules that you configure, not code:

The key point is Data Masking. It runs inside SAP AI Core, before the model call. In pseudonymize mode the original values are put back into the response afterwards, so the model works with placeholders and the user still gets a reply with real names.
The idea is:
the app knows only the AI Core service key, no provider keys
the prompt lives in Generative AI Hub as a versioned template
personal data is masked before the model call and restored after it
input and output filters block inappropriate content in both directions
3. SAP AI Core and AI Launchpad setup
Step 3a: Entitlements
In the BTP cockpit, add these service plans to your subaccount:
SAP AI Core, plan
extendedSAP AI Launchpad, plan
standard
Note: Generative AI Hub is only available on the
extendedplan of SAP AI Core. Withfreeorstandardyou will not see any foundation models.
Step 3b: Connection and resource group
Create an SAP AI Core instance and a service key, then subscribe to AI Launchpad. In AI Launchpad go to Workspaces → Add and create a connection with the values from the service key. Select the default resource group.
Note: the orchestration deployment is created automatically only in the
defaultresource group. If you create your own resource group, you have to create a configuration with scenarioorchestrationand executableorchestrationand deploy it yourself.
4. Orchestration configuration
In AI Launchpad go to Generative AI Hub → Orchestration and create a new configuration, for example helpdesk-triage.
Prompt Template
The prompt asks for a strict JSON output. The e-mail is a variable:
Variables use two curly brackets and a question mark: {{?email_body}}. Launchpad picks the variable up automatically under Variable Definitions, where you can also set a default value for testing.

Model Configuration
Select the model and its parameters. For classification and extraction a small model is usually enough, and temperature 0 keeps the output stable. You can browse all available models in Model Library.

Data Masking
Switch Advanced on to see the optional modules, then activate Data Masking:
Mode: Pseudonymize
Masked Data Types: Person Names, Email Addresses, Phone Numbers

Note: Anonymize replaces the values permanently. The model's reply would then start with "Hi MASKED_PERSON_1". For a suggested reply you need Pseudonymize.
Note: I deliberately do not mask Organizations. The company name helps with routing, and it is not personal data. Values that must never be masked, such as your own product names, go to the Allow List.
Input and Output Filtering
Activate both and keep the default thresholds for hate, violence, self-harm and sexual content. Helpdesk inboxes get some colourful e-mails, and you don't want them passed to the model or a generated reply in the same tone.
Save the configuration. The prompt is stored in Prompt Management, versioned and grouped by scenario, so the app refers to it by name and a prompt change doesn't need a redeployment.

Calling it from the app
With the SAP Cloud SDK for AI, the call from a Node.js or CAP app is short. The SDK handles authentication against AI Core:
The SDK can also reference a configuration saved in Generative AI Hub instead of defining it inline. That way masking and filtering stay in one place and are not duplicated in code.
5. Testing
Run the configuration in Launchpad with a typical helpdesk e-mail:
Hi, this is Jana Nováková from ACME (jana.novakova@acme-example.com, +420 601 123 456). Since this morning I can't post invoice 2026-10045 in FB60 – it says the period is closed. Could you help urgently? We have a payment run tomorrow.
This is what the model actually receives:
Hi, this is MASKED_PERSON_1 from ACME (MASKED_EMAIL_1, MASKED_PHONE_NUMBER_1). Since this morning I can't post invoice 2026-10045 in FB60 – it says the period is closed. …
And this is what comes back to the app:
The model never saw the name, yet the reply says "Hi Jana". Orchestration put it back.
Negative tests
These show that the protection works, and help to understand the principle.
Anonymize instead of Pseudonymize. Switch the mode and run again. The reply now contains the placeholder instead of the name, because anonymized values are not restored.
Offensive e-mail. Send an e-mail with abusive language. The input filter stops it, and the model is never called.
From here it is just a matter of where the ticket goes: a ticketing tool, an SAP workflow, a Teams channel, or a draft reply in the mailbox.
Two more things I recommend:
Don't put real data into the template. Default values in Variable Definitions are stored with the configuration. Use made-up test e-mails only.
Keep a human in the loop. The suggested reply is a draft. Somebody from the team approves it before it is sent.
Summary
With orchestration in Generative AI Hub, the helpdesk e-mail is masked inside SAP AI Core before the model call, and the real values are restored in the response. The app holds only the AI Core service key, the prompt is a versioned template, and switching to another model is a configuration change.
With a direct call to a model provider, every app needs its own key, and personal data goes to the model as it is, unless you build masking yourself.
The same configuration works beyond the helpdesk: supplier e-mails to Accounts Payable, order confirmations or approval requests. Only the prompt template changes.
To sum up the best practice: if an LLM needs to read e-mails with personal data, call it through orchestration in Generative AI Hub with Data Masking in Pseudonymize mode. Keep direct provider calls for prototypes on made-up data.
Screenshots from the SAP AI Launchpad documentation © SAP SE, used under the CC BY 4.0 license.
AI
Content yes, names no. Helpdesk e-mail triage with SAP Generative AI Hub

