Blog · Security
Giving AI agents access to your systems? Start with the least they need
AI agents are moving from answering questions to doing things: sending emails, updating your CRM, filing tickets, moving files. That's where the real time savings are. It's also where a mistake, or a cleverly worded email, can cause real damage.
The safest approach is an old one from security: give each agent the least access it needs to do its job, and nothing more.
1. Write down the job first
Before connecting anything, describe the agent's job in one sentence. "Drafts replies to support emails for a person to review" needs very different access from "manages the support inbox".
2. Read before write
Start agents with read-only access. Add the ability to change things one action at a time, once you've watched how the agent behaves.
3. Give each agent its own account
Don't let an agent borrow a person's login. A separate account or key means you can see exactly what it did, limit it, and switch it off without locking anyone out.
4. Keep a person on the risky steps
Sending money, deleting records, emailing clients or changing permissions should need a human click. Agents can prepare the work; people approve it.
5. Treat incoming content as untrusted
An agent that reads emails, web pages or documents can be tricked by instructions hidden inside them. That's known as prompt injection. Limiting what the agent can do is the best defence, because it caps the damage even when it is fooled.
6. Log everything and review it
Keep a record of every action each agent takes. Look through it weekly at first. You'll quickly spot what it handles well and where it needs tighter limits.
A simple test
Ask: "If this agent did the worst thing its access allows, how bad would it be?" If the answer worries you, reduce the access until it doesn't.
Want a second pair of eyes on how your firm uses AI? Book a free 20-minute call.