In a previous post in this series, we covered the difference between a large language model (LLM) and an agent, or agentic AI. To recap: with an LLM, you ask a question and get an answer. With an agent, you give it a high-level task, and it figures out the steps and does the work for you.
That is what makes agentic AI so powerful, and it's also what makes it risky. With a chatbot, the worst case is usually a wrong answer. With an agent, the failure mode that concerns us most is a bad action: a deleted folder, a changed setting, or an email sent to a client that never should have gone out.
If you are planning to use an agentic AI tool such as Claude Desktop, OpenClaw, or ChatGPT's agent, here is the checklist we recommend.
1. Use a Dedicated Workstation
Do not install an AI agent on your regular workstation. Your everyday computer has access to your email, file shares, line-of-business apps, saved passwords, and more. If you install an agent there, it inherits all of that access, which is almost certainly far more than it needs.
Follow the principle of least privilege: give the agent just enough access to do the job you need done, and nothing more. A dedicated AI workstation makes that possible.
2. Sandbox the Workstation on Its Own Network Segment
The dedicated workstation should live in its own silo, a practice often called sandboxing. In practice, that means:
- A separate subnet and VLAN for your AI workstations
- A stateful inspection firewall controlling all traffic in and out of that segment
This lets you limit what the agent can and cannot see on your network. Just as important, it gives you a kill switch. If the agent starts behaving unexpectedly, you can cut off its access immediately.
3. Create a Dedicated Login with Minimal Rights
Don't let the agent run under your regular user account. Your login probably has more rights than the agent needs for any given task.
Instead, create a dedicated account for the AI workstation. Start it with the minimum level of rights, then increase them gradually as you give the agent more responsibility.
Even on that dedicated login, require multi-factor authentication (MFA) to access the workstation, using an app such as Duo or Okta.
4. Install the Full Cybersecurity Stack
An AI workstation deserves at least the same protection as any other endpoint, and arguably more. Make sure it has:
- Antivirus
- EDR / MDR (endpoint detection and response / managed detection and response)
- Application whitelisting
- Change tracking, so you can compare how the system looked when the agent was behaving well against how it looks after something went wrong
- Centralized patch management for the operating system, the agentic AI application, and the browser
AI tools are updating constantly. Staying current on patches isn't optional.
5. Build In Human Approval Gates
This is where many people get into trouble. It is tempting to say, "Clean up my inbox and automatically reply to my clients." Don't.
Put a human checkpoint in front of any action that matters. For example:
- Have the agent filter and organize your inbox, but don't let it delete anything.
- Ask it to propose a plan and wait for your approval before it acts.
- If it drafts replies, have it save them to your Drafts folder so you can review them.
- Never let the agent send email directly. Ever.
AI can do a great job helping you write those emails. You should still be the one who clicks Send.
6. Monitor Access Closely
Watch who logs into the AI workstation and what happens on it. Review access logs and activity regularly so you can catch problems early.
7. Start Small and Ramp Up
Take baby steps. Inbox management and spam filtering are a great first project. They save real time quickly and carry relatively low exposure, as long as the agent isn't allowed to delete anything. Once you're comfortable with how the agent behaves, expand its responsibilities gradually.
The Bottom Line
Agentic AI is fantastic. It can save you a ridiculous amount of time. But it needs to be deployed safely and securely:
- Dedicated workstation
- Sandboxed network segment behind a stateful firewall
- Dedicated, least-privilege login with MFA
- Full security stack, change tracking, and patch management
- Human approval gates and no direct sending of email
- Close monitoring
- Start small
And one more time, because it's the most important point: please do not install an AI agent on your regular production workstation. You are asking for trouble.
Need Help Setting Up a Secure AI Workstation?
ADS Consulting Group can help you design and deploy an agentic AI workstation that's isolated, locked down, and monitored, so you get the productivity gains without the risk. Email us at info@adscon.com or book a free discovery call.
Prefer video? Watch the full discussion on our YouTube channel.