AI can improve customer service and reduce repetitive work, but a chatbot becomes a security concern as soon as it can access customer records, internal documents or business actions.
Start with a narrow use case
Choose one low-risk task first, such as answering public service questions or helping staff find approved information. Do not begin with unrestricted access to a CRM, finance system, shared drive or inbox. The first test should prove usefulness without exposing sensitive data.
Check what the vendor does with data
Review the provider’s privacy policy, data storage, retention, model-training settings, subcontractors and incident-notification process. Confirm who owns prompts, uploaded documents and conversation records. If the answers are unclear, do not connect real customer or staff information.
Keep permissions and actions small
Use a separate account or integration with the minimum access required. Limit the records the chatbot can retrieve, and separate answering questions from changing data, sending messages, approving transactions or making commitments.
Keep a human in the loop
Chatbots can produce inaccurate or manipulated answers. Require a person to review legal, medical, financial, safety, privacy and complaint-related responses. Record who is responsible for checking the output and what happens when the system behaves unexpectedly.
Test before launch and keep reviewing
- Use fake or anonymised data during testing.
- Try misleading prompts and requests for information the tool should not reveal.
- Check that access is logged and permissions can be removed quickly.
- Document the response plan if the vendor, model or integration is compromised.
Sources
- Cyber.gov.au: Artificial intelligence for small business
- Cyber.gov.au: Artificial intelligence for small business guidance
- Cyber.gov.au: Defending against AI-enabled cyber attacks
Leave a comment