The AI Assistant is the one feature that sends your data outside the application, and only to the provider you configured with your own key. Everything else in MTC Skopos runs locally. To limit what leaves, the assistant substitutes placeholders for sensitive fields before any request is sent.
How it works
- Sensitive field values are replaced with placeholders of the form
«u_1»,«r_2»and so on - The model reasons and answers using the placeholders
- Placeholders are substituted back to real values on your machine before the answer is displayed
The same value always maps to the same placeholder within a session, so the model can still tell that two rows concern the same user without knowing who that user is. It never receives the original value for an anonymized field.
Anonymizable fields
| Field | Placeholder | Anonymized by default |
|---|---|---|
| User IDs | u_1, u_2, … | Yes |
| Role Names | r_1, r_2, … | Yes |
| Composite Roles | c_1, c_2, … | Yes |
| System Names | s_1, s_2, … | Yes |
| Full Names | fn_1, fn_2, … | Yes |
| Departments | dp_1, dp_2, … | Yes |
| Locations | lo_1, lo_2, … | Yes |
| User Groups | ug_1, ug_2, … | Yes |
| HR Functions | hf_1, hf_2, … | No |
| Permission Values | v_1, v_2, … | No |
Eight of the ten fields are anonymized out of the box. The two left visible are the ones the model needs in clear text to be useful: permission values are technical identifiers such as transaction codes and authorization objects, and HR function is what lets the model reason about whether an access pattern fits a job.
Configuring anonymization
Go to Settings > AI Assistant > Anonymization and toggle fields individually.
- Preview Anonymization shows the exact payload before each request is sent. Turn it on the first time you use the assistant on production data, and leave it on if your policy requires review before egress
- If the preview shows a value you expected to be masked, select it there and apply anonymization from the menu that appears
Choosing a policy
- Against a cloud provider, keep the defaults at minimum. User IDs, role names, composite roles, and system names are the fields that identify your organization and its people
- Anonymizing everything, including permission values, is supported and costs you some answer quality: the model can no longer tell that
v_7isSM30, so it cannot reason about what an action does - Against a self-hosted or local model with no egress, you may disable anonymization entirely
- Anonymization applies to the assistant only. It has no bearing on exported reports or analysis logs, which stay on your machine in clear text