Security and Governance

Enterprise AI, Under Your Control

AI makes data a user is already allowed to see in the ERP visible faster; it does not widen permissions, it does not move data out, and it does not make the decision. This page keeps the controls in place today apart from the architecture we are aiming at.

Let’s go through your security requirements

Security discussions in AI projects usually start with the model itself. On the enterprise side the real question is simpler: whose permissions is this system using, which data is it looking at, and who can see that?

Controls in place today, and the target architecture

In place
  • Queries run read-only and parameterised; the model never reaches the database directly
  • The user’s ERP permissions apply; the AI creates no rights of its own
  • Role-based and row-level authorisation in ErpwareBI
  • The question, the generated query and the result are written to an audit log
  • By default the model runs inside the company; data does not leave the building
  • Integration transfers are logged with a timestamp, the source record and the outcome; failed records go to an error queue
  • Nothing is written directly to the IFS schema; transactions go through the system’s own business rules
Target architecture
  • A separate identity for the agent, with narrow and time-limited rights
  • Tool-level permission: an explicit list of the functions an agent may call
  • Data-changing operations opened only inside an approval flow
  • End-to-end tracing of an agent’s steps (which model, which tool, which data)
  • Standard, auditable connections to outside systems through MCP

The target architecture items are not products today. If a security capability is marked “target architecture” here, it is not running in the field yet.

The chain of permission and audit

  1. User
  2. Identity
  3. Role
  4. Data permission
  5. Read-only query
  6. Audit log
  7. Tool permission target
  8. Transaction target

The dashed links in the chain arrive with the agent layer; today the AI reads data, it does not act on it.

Where does the data stay?

Running in-house (default)

When the model runs on your own hardware, the question, the schema and the result never leave your network.

  • Local model (Ollama / TensorRT)
  • No internet connection required
  • Logs stay on your own systems

If a cloud model is used

Where a more capable model is needed, the scope is agreed in writing beforehand.

  • Which fields are sent is defined one by one
  • Personal data and commercially sensitive fields are kept out of scope
  • The provider’s data processing terms are reviewed together

Frequently asked

Will our data be used to train a model?
In an in-house installation the question does not arise, because the data never leaves. If a cloud model is to be used, the provider’s data processing terms are read together before the project starts and the scope is written into the contract.
How are user permissions enforced?
The permissions in the source system apply: data a user cannot see in the ERP does not arrive through the AI either. In ErpwareBI, role-based and row-level restrictions can be defined on top of that.
What happens if a wrong or harmful query is generated?
A generated query is not run before it is validated, and it only reads. Operations that change data are out of scope today — the simplest way to limit the impact of an error from the start.
How long are the logs kept?
Audit records are kept on your own systems; the retention period and who may read them follow your organisation’s own policy.
Are you GDPR/KVKK compliant?
Compliance is a legal state assessed together with an organisation’s own processes; it would be wrong to claim that a piece of software is “compliant” on its own. What we can do is prepare the technical ground: keeping data inside the company, separating permissions, logging for audit, and putting the scope in writing where a cloud model is used.

Read on

Every control on this page is written to the extent that it is applied in the field. If you have a security questionnaire, send it over and we will answer it line by line.