What to Let an AI Assistant Access, and What to Approve Yourself
How I use an AI assistant every day, and the access I still keep behind my own approval
I use an AI assistant almost every day in my work at RealWis. It helps me draft emails, think through a webpage, and organize a messy set of notes. I still send the email myself. I still approve anything that would be deleted or posted. And I give the assistant only the access that particular task needs.
That last part is the one I want to talk about.
An older kind of software waits for you to click a button. An AI assistant, sometimes called an agent, can be connected to your email, files, calendar, or website, and then take steps on its own. It can look something up, draft a message, or work in another app you've allowed it to use. That's useful for a small office. It also means the assistant can only be as careful as the access you give it.
Why this matters in a small office
In a large company, a security team may review every new tool. In a small business or a nonprofit, that job often falls to the same person who answers the phone, updates the website, and keeps the books. One shared login, or one "just connect it so it's easier" decision, and a tool can see more than anyone meant it to.
Security people have a name for giving a tool more reach than the job requires: excessive permissions. They also talk about shadow AI, which is just a formal way of saying someone on the team signed up for a tool that the rest of the office doesn't know about. Neither idea is complicated. Both show up in a three-person shop just as easily as they do in a big one.
This is not a reason to avoid these tools. It is a reason to be deliberate, the same way you'd be deliberate about who has a key to the office.
Six plain rules
1. Give only the access the task needs
If the assistant is helping you rewrite a page of website copy, it does not need your bank login or your entire email inbox. Start with the smallest access that gets the job done. You can add more later if you have to. You cannot always take back what was already seen or sent.
2. You press Send
An assistant can draft a reply, a newsletter, or a social post. You should be the one who reads it and hits Send, Publish, or Delete. That short pause catches a wrong recipient, a tone that doesn't sound like you, or a fact that isn't quite right. It also means a tool can't speak for your organization without you in the loop.
3. Use separate accounts and logins
Don't hand an AI tool the same all-purpose login you use for everything else. If a service lets you connect an assistant with a limited login, use that. Keep personal accounts separate from the business ones. If someone on your team uses a tool, they should use their own sign-in, not a shared password written on a sticky note.
4. Don't paste passwords, codes, or card numbers into a chat
A chat window is not a safe place for a password, a two-factor code, a Social Security number, or a credit card. If a tool needs to sign in somewhere, use the official connect screen, or do that step yourself. If you already pasted something sensitive, change that password.
5. Review connected apps on a regular schedule
Many tools stay connected long after the project that needed them is over. Once a month, or at the end of a busy season, look at which apps still have access to your email, cloud files, or calendar. Remove the ones you no longer use. This is the same habit as cleaning leftover backup files off a web server. The leftover access is often the part that causes trouble later.
6. Know what data it can see
Before you connect a tool, ask a simple question: if this assistant can read this inbox, this folder, or this donor or customer list, am I comfortable with that? Names, private records, employee information, and anything you'd hesitate to leave on a public webpage should stay off-limits unless you have a clear reason and a clear limit.
Questions worth asking
You don't need a long policy document. A short conversation, and a list you can revisit, is enough.
- Which AI tools does our office actually use, including ones people signed up for on their own?
- What can each tool see: email, files, calendars, the website, or customer or donor records?
- What can it do on its own, and what still needs a person to approve it?
- Who is allowed to connect a new tool, and who reviews those connections later?
- Have we removed access for tools we no longer use?
- Are passwords, financial details, and private records kept out of the chat?
FAQ
Is it unsafe to use an AI assistant at work?
No. I use one every day. The usual problem isn't the tool itself. It's giving it more access than the job needs, or letting it take an action you haven't reviewed. Treat it like a helpful intern: useful, but not handed the master keys on the first day.
What's the difference between a chatbot and an AI agent?
A chatbot answers questions in a window. An agent is a tool that can also take steps, such as reading a file, drafting a message, or working in another app you've connected. The more it can do, the more carefully you should decide what it's allowed to touch.
Can RealWis help us look at this?
Yes. This is the same kind of practical review we already do for website security. In The Two-Minute Server Fix That Keeps Your Website Backup Private, I wrote about closing a door on the server that didn't need to be open. We can look at your site, and at the tools connected to your work, with the same question in mind: what should be allowed, and what should stay behind your approval?
Jim Murphy is a U.S. Army veteran and the founder of Real Wisconsin Website Design, a veteran-owned and woman-owned web design company in Green Bay, building accessible, mobile-responsive websites for small businesses and nonprofits since 2007.
