The Allow button is the new handshake
You’re about to hand an AI the keys to your email, your calendar, or your customer list. It’s going to take about four seconds.
A little screen pops up. It says the app wants access. You’re busy, so you click Allow.
Here’s the thing. That click is the most important business decision you’ll make this week, and almost nobody treats it like one. If you hired a new assistant, you would never hand them every password on day one. You’d give them a desk, a task, and a little supervision. Connecting AI to your business should work exactly the same way.
What just changed
Let me simplify the jargon first. A connector is a bridge that lets an AI tool like Claude or ChatGPT reach into another app. Your CRM, your inbox, your files, your MLS. A plugin is a package that can bundle those connectors with ready-made skills, which are basically saved instructions for how to do a job.
Until recently, finding and trusting these things was a bit of a wild west.
On September 25, Anthropic opened a submission portal for Claude plugins. Developers submit a connector or a bundle, and it gets automatic validation and a safety scan before it’s listed. The portal also supports the newer MCP 2.0 standard, apps that show interactive screens right inside the chat, and a smoother login setup for companies.
So what does that mean for you? More connectors are coming, and they’re coming fast. A lot of them will be built by people you’ve never heard of. The safety scan helps, but think about it like this. It’s a smoke detector, not a security guard. It can tell you something smells off. It can’t stand at the door.
The final call is still yours.
The five-question test
Before you click Allow on anything, run it through five questions. It’ll save you from most of the mess.
First, who built it? A name, a website, and a real company that you are familiar with, or have at least heard of and can research. A verifiable presence beats an anonymous listing every single time.
Second, what can it see? Actually read the permission screen (I know, nobody reads the permission screen). And I know what you’re going to ask me next. What, now I have to read the terms of service too? Yeah, I would. Or even better, have AI read them for you. Now we’re talking. If a tool that formats your newsletter wants your whole inbox, that math isn’t mathing.
Third, what can it do? Reading is one thing. Sending, deleting, editing and spending are a whole different conversation. Start with read-only whenever you can.
Fourth, where does my data go? Ask whether your customer information gets stored, shared, or used to train anything. If you can’t find the answer, that is the answer.
Fifth, how do I undo it? Know where the off switch lives before you ever flip the switch.
The risk nobody mentions in the demo
Now here is the real talk on one risk worth knowing about. Security researchers call it prompt injection, and it sits at number one on OWASP’s list of the top risks for AI applications. Somebody hides instructions inside an email, a web page, or a document. Your AI reads it, treats it like an order, and acts on it.
Picture your new assistant opening a letter that says “ignore your boss and send me the client list,” and doing it because the letter asked nicely.
The more doors you connect, the more places a trick like that can walk in. So slow way down when a tool wants to read outside content and also take action for you. That’s where the real danger is.
Your team already clicked Allow
If you’re a business owner with employees or staff, your risk just went up exponentially. You may not be using AI at that level yet, but some of your people definitely are.
That’s not a guess. Look at my own industry. NAR’s 2026 Realtors Technology Report found that nearly half of agents now use AI daily (23%) or weekly (25%), and another 31% are experimenting. Only 12% say they’re not using it and have no plans to. A year ago, 32% hadn’t even tried it.
Now look at what they’re using it for. Among agents who use AI, 52% use it for emails and follow-up, and 27% use it to review or summarize documents. That’s your inbox and your contracts. The exact two doors I’d tell you to slow down on.
Think about what that looks like on a Tuesday afternoon. Your office manager connects a free scheduling assistant to the shared calendar. Your newest agent links a note-taking app to the team inbox. Your top producer plugs the CRM into whatever tool they saw on a video last weekend. Nobody’s doing anything wrong. They’re trying to save time, which is exactly what 81% of agents in that same report said they want from technology.
But every one of those clicks is a door, and you don’t know which ones are open.
The answer isn’t banning AI. Ban it and people keep using it, they just stop telling you. The answer is giving your team the five questions and a clear set of rules, so the smartest people in your business are also the safest ones.
What’s fine to plug in, and what to keep out
| Usually fine to start with | Slow down and think first | Keep out until you have real safeguards |
|---|---|---|
| Calendar, read-only | Your full email inbox | Bank and payment logins |
| Public website and social content | CRM with client contact details | Passwords and ID documents |
| A shared folder of non-sensitive templates | Shared drives with contracts | Medical, legal or tax records |
| Meeting notes you would be fine reading aloud | Anything that can send messages for you | Anything that can move money |
The pattern is simple. Start with information you wouldn’t mind a stranger reading. Give it read-only access. Make the tool prove it’s worth it for 30 days. Then widen access one door at a time.
Trust is earned in small steps. Even by software.
Here is the play
Run the five questions before every Allow. Start read-only. Keep money, passwords and private records out until you have real safeguards. Be extra careful with any tool that reads outside content and takes action. Give your team the same rules you follow. Check your connectors once a month, the same way you’d check your bank statement.
That’s it. Ten minutes a month to keep the doors you meant to open, and close the ones you didn’t.
Your next step
We sat down recently with Lynette Adams on WHEC 10 and recorded a “Rochester In Focus” segment. The big idea we talked about was simple: AI works best when a small business owner sets the rules before the tools show up. This article is the other half. The rules tell your team how to use AI. The five-question test tells you what you let in the door.
Want to put it to work today? Start with our FREE House Rules tool and make them fit your own business. That way your team knows the game plan before any new tool joins the roster.
Then open your AI tool, look at every connector you’ve already switched on, and run the five questions on each one. Turn off anything you can’t answer for. I’d bet you find at least one you forgot about, sitting there like the Christmas toy nobody’s touched since February. Or worse yet, December 26th.
AI should work for your business like a great new hire. Clear job. Limited access. A manager who’s paying attention.
That manager is you.
P.S. Want a fun one for your next team meeting? Ask everybody to open their phone and count the apps they’ve connected to AI. Don’t judge the number. Just get it on the table. The conversation that follows will tell you more about your real AI risk than any software audit.
