An AI tool is safe enough to connect to your business when it signs in through a proper Google or Microsoft screen, lets you limit it to what one job needs, keeps a human check before anything is sent, lets you export and delete your data, and has an off switch you can find before you connect it. If a tool fails any of those, it waits until it can pass.
I teach this in the third week of my AI for Ops cohort, because trust is the real blocker for most of the people I teach. They want AI to help with their inbox and their calendar, and they are nervous about handing over the keys. That nervousness is healthy, and it deserves a method.
Start with least access
The way most people start is "connect everything and see what happens," because every tool makes it one click. The trouble is that you can't tell which tool did what, and you often can't take it back.
Least access means giving a tool only what one job needs. That might be one email label instead of the whole inbox, or permission to read and draft instead of permission to send. It also means adding one new tool at a time, so if something goes wrong, you know exactly where to look.
First, pick the lane
Before you judge a tool, decide whose job it is:
- Build: the job is yours alone, like a weekly update in your own voice. You build it inside your own AI project.
- Hire: it is a common job lots of businesses share, like booking or note taking, and a product already does it well.
- Tame: a tool you already use has AI built in, so you switch it on carefully and turn the extras off.
All three lanes go through the same safety check. Building something yourself does not make it safe by default, because it still connects to your email.
Step by step: the 10-question safety check
Answer each question from the tool's own permissions screen, settings and privacy policy, not from its homepage.
- How does it sign in?
- What can it see?
- What can it do: read, draft, send or delete?
- Can I limit its access to what the job needs?
- Does it train its AI on my content?
- Who else sees my data?
- Can I export my data?
- Can I delete my data and disconnect?
- Is there a human checkpoint before it sends anything?
- What is the result: pass, limited use or fail?
The rule I use is simple. Any red flag, or more than two "not sure" answers, and the tool waits. Waiting is not a ban forever, and it means the tool has to earn its way in.
Question three deserves extra attention now. Connectors used to mostly read your data, and many of them can now write and send as well.
Four red flags that end the conversation
- It asks for your password. A real tool sends you to a sign-in screen from Google or Microsoft that lists what it wants.
- There is no off switch. If you can't find how to disconnect it before you connect it, don't connect it.
- You can't export your data. If your data can't leave, it has effectively become theirs.
- It wants full access on day one. A tool that won't work without everything was not built for least access.
You don't have to read privacy policies for fun. Ask Claude or ChatGPT to read the policy for you and answer questions five and six, then ask a second AI the same thing for another opinion.
Frequently asked questions
Is it safe to connect Claude or ChatGPT to my Gmail? It can be, if you run it through the same check. Start with read and draft access, keep a human check before anything is sent, and limit it to the label or folder the job needs.
What if a tool I already pay for fails the check? Tame it. Turn off the AI features you don't need, tighten its permissions, and keep using the parts that pass.
How often should I review what my AI tools can access? Once a month works well. Tomorrow's post walks through a simple monthly access audit.
Learn it live
My free AI 101 session is tonight, Wed Nov 18, 6 to 7 PM PT at SHACK15 in San Francisco, with online audit seats. It is full right now, and you can still join the waitlist. This week's YouTube lesson shows what ChatGPT says about my own business, which is a useful companion to this check. Watch it on the Build with AI Club YouTube channel