Skip to content
SundAIFree call20 min
Menu
Browse all guides

How to deploy a language model securely in your business

Before connecting an LLM to customer records, use a returns desk pilot to test API security, prevent data leaks and measure the value of staff time recovered.

AI governance starts with one decision

Imagine a customer asks, “Can I return this order?” A support assistant needs the purchase date, product and relevant returns policy. It can draft a reply. An employee checks eligibility, sends the message and approves any refund. ISO/IEC 42001 gives organisations a framework for managing AI throughout its life. Here is how a leadership team could apply it to this pilot.

  • CEO or sponsor — set a pilot budget and define the outcome worth paying for, such as shorter handling time at the same answer quality.
  • CTO — map each system the assistant can reach. Record who owns the integration and who can disconnect it.
  • CISO — agree what would stop the launch, who investigates and what evidence is needed to restart. Reopen the decision when the model, supplier or data source changes.

What GDPR and the EU AI Act mean for the pilot

Where GDPR applies, document the purpose and lawful basis for processing personal data. Send only the fields needed for the return. Check how the supplier uses prompts, which subcontractors can see them and when copies are deleted. If the supplier acts as your processor, cover the Article 28 terms. Carry out a data protection impact assessment before processing likely to pose a high risk to people’s rights and freedoms. For the EU AI Act, identify your role and intended use. One way to support staff AI literacy under Article 4 is to practise spotting a draft that promises a refund outside the policy.

Protect the LLM API and limit the damage a mistake can cause

Put the application server between the employee and the model provider. Give the pilot a read-only connection to the returns system and keep permission to issue refunds with staff.

  • Keep the provider key in a server-side secrets manager. Use an encrypted connection and set usage limits per user.
  • Check the signed-in employee’s access to each case and document before retrieving text. A shared service account with broad access needs this extra check.
  • Log the request ID, access decision, model version and cost. Keep messages, credentials and customer details out of routine logs. Agree who can read diagnostics and when they expire.

A data leakage test before launch

Create two test accounts and a fictional warranty note visible to only one. Ask both accounts the same question. Check the reply, quoted text, source link and logs for restricted details. Revoke the first account’s permission, open a fresh session and time how long the change takes effect. If restricted content reaches either account after its access should have ended, disconnect the source, fix the control and repeat the test.

Calculate the value per completed case

Illustrative monthly pilot: 300 comparable cases fall from eight minutes each to five, including checking and corrections. That frees 15 staff hours. At €40 an hour, the time is worth €600 if the team can put it to use. With €180 in monthly running costs, €420 remains in estimated value before one-off implementation costs. Compare answer quality and reopened cases alongside time. Cash savings depend on which expense actually falls.

What SundAI can do before you sign or connect

Bring one decision to a free 20-minute call. For example, should the assistant have access to the returns mailbox? Our €490 Decision Review gives you a written recommendation with the evidence behind it and the gaps to resolve. If you move to implementation, we can map data flows with your IT provider, test permissions and measure handling time with your team. Finding an overbroad connector at this stage can spare a second integration and another round of supplier review.

Discuss your AI decision in a free 20-minute call

Sources