AI Agents and Your Customer Data: A Practical Security Checklist

AI agents are different from the chatbots most of us have tried. A chatbot answers a question and forgets you. An agent is given access to your tools, such as your email, calendar, CRM or accounting software, and then acts on your behalf, often while you are not watching. That is exactly what makes them useful, and exactly why they deserve a proper security conversation before you switch one on.

This checklist is written for business owners and managers, not engineers. You do not need to understand how the technology works internally. You need to know what questions to ask and what settings to insist on. Work through it before any agent is connected to customer data, and revisit it whenever the agent gains a new capability.

1. Map what the agent can actually see

Start with a simple inventory. When you connect an agent to a tool, what does it get access to? Often the answer is far broader than the task requires. Connecting an agent to your mailbox so it can sort supplier invoices may also expose every customer complaint, contract and password reset email in that mailbox.

Write down, in plain language, each system the agent will touch and the types of customer data inside it. Typical categories include:

If you cannot say what the agent can see, you cannot say what it can leak. Treat this map as the foundation for everything else.

2. Apply least privilege

Least privilege is an old security principle with a simple meaning: give a person, or a program, only the access needed to do the job and nothing more. It matters even more for agents because they act quickly and at scale.

In practice, that means:

A useful test: if this agent were a new part-time employee on their first week, would you hand them this level of access? If the answer is no, narrow it.

3. Decide where a human must approve

Not every action carries the same risk. Drafting a reply is low risk. Sending it to a customer is higher. Issuing a refund, deleting records or sharing a file outside the company is higher still. Decide in advance which actions the agent can take on its own and which need a person to say yes first.

A sensible default is to require approval for anything that is hard to undo, involves money, or leaves your organisation. Look for tools that make approval a built-in step rather than a polite suggestion the agent can skip. Be wary of any setup where approvals are so frequent that your team starts clicking yes without reading. That is approval fatigue, and it quietly turns a safeguard into a rubber stamp.

4. Insist on logs you can read

If something goes wrong, you will want to answer three questions quickly: what did the agent do, what information did it use, and who or what told it to? That requires logs.

Check that the tool records each action with a timestamp, the data it touched and the instruction behind it. Then check that the logs are readable by a non-engineer, that you can export them, and that you know how long they are kept. Logs that exist but that nobody reviews do not help.

5. Questions to ask any vendor

You do not need to be a security specialist to ask good questions. Put these to any vendor, and note whether the answers are clear, vague or missing:

That last point deserves a moment. Agents read content from outside your business, such as inbound emails and web pages. A malicious message can contain instructions aimed at the agent rather than at you. Good vendors have thought about this and can explain their defences. A vendor who has never heard the question is telling you something.

6. Retention and deletion

Agents often keep a memory of past work so they can be more helpful. That memory is also stored customer data. Find out what is remembered, where it lives, how long it is kept and how to delete it. Ideally you can set a retention period and wipe specific items on request.

Also think about your own side. If a customer asks you to delete their information, can you be sure it is gone from the agent’s memory, its logs and any connected tools? Build that check into your process now rather than discovering the gap during a request.

7. Data protection obligations in plain terms

Many countries have data protection rules, with the GDPR in Europe being the best-known example, and similar ideas appear in laws around the world. This is general information, not legal advice, and the details depend on where you and your customers are. But the common themes are easy to remember:

Adding an agent does not remove any of these duties. It adds a new supplier and a new place where data flows, so your privacy notice, supplier records and internal procedures may need a small update. If you handle sensitive categories of data or operate in several jurisdictions, take advice from a qualified professional.

Start small and review

The safest way to adopt an agent is to give it one narrow job, with read-only access, human approval and active logging, and then expand slowly.

If you would like a second pair of eyes on how an agent fits into your systems before you connect it, an experienced digital partner can help you map the data, set sensible limits and build the review habit from day one.

← Back to Blog
Share on LinkedIn Share on X