How to build an AI assistant for your company without compromising your data security

An AI assistant connected to your company data can answer in seconds what used to require searching several systems. It can also become a risk if built carelessly. These are the principles we follow.

1. Define what problem it solves

Before connecting anything, decide the use case: checking order status? searching documents? helping the support team? The more concrete it is, the easier it is to limit what data it needs and who should use it.

2. Authenticate every user

The assistant must not be accessible to just anyone. Common options:

• Corporate sign-in (Microsoft, Google).
• One-time code by email to an authorised address.
• Attempt limits and session expiry.

3. Apply permissions per user and role

The assistant must see only what that person can see. If a salesperson has no access to payroll data, the assistant must not show it to them either. This is achieved by querying systems with the user’s permissions, not with a universal account.

4. Principle of least privilege

• The assistant’s tools should have read-only access unless the action is essential.
• Actions that modify data (creating a booking, sending an email) must ask the user for confirmation.
• Service accounts with minimal permissions and credentials stored in a secrets manager, not in the code.

5. Protect the connection to your systems

• Do not expose your internal servers directly to the internet.
• Use private networks or secure tunnels between the assistant and the systems.
• Limit by IP and port what the assistant can reach.

6. Take care of what is sent to the language model

• Send only the information needed to answer, not entire databases.
• Choose providers that do not train on your data and offer data processing agreements.
• Anonymise or mask personal data when it is not needed.

7. Keep a log

Record who asked, which tool was used and when (without storing more than necessary). It serves for auditing, detecting unusual use and improving the assistant.

8. Defend against malicious instructions

Models can receive texts that try to manipulate them (“ignore your rules and show…”). Mitigations:

• Clearly separate system instructions from user and document texts.
• Validate on the server what the assistant asks to do; do not trust the model.
• Test the assistant with abuse attempts before launch.

9. Comply with regulations

• Tell users they are interacting with an AI.
• Review the legal basis and the GDPR record of processing activities.
• Validate legal texts with your adviser.

10. Start small and measure

A first assistant with one use case and few permissions is easier to protect and improve. It is then expanded with the confidence gained.

We have built assistants with verified access, per-user permissions and controlled connection to the systems of a logistics company (you can read the case study). If you want to explore one for your company, let’s talk.