Onsite AI The honest version
Is my data safe?
This is the question that stops most owners, and most people selling AI answer it with a shrug and the word “encrypted.” Here’s the actual answer, including the parts that don’t help me sell anything.
Everything on this page comes from Anthropic’s own published documentation and terms. I’ve linked the sources. Things change fast in this industry, so this page is dated, and if you find something that’s gone stale, tell me and I’ll fix it.
Last checked: July 29, 2026.
What it can and can’t actually do
People imagine an AI with the run of their whole company. What it actually has is a specific, limited set of functions. Here they are.
| Service | It can | It cannot |
|---|---|---|
| Gmail | Search your mail, read messages and threads, write drafts, update drafts, manage labels | Send email. Delete email. |
| Calendar | Read, search, create events, change events, delete events, RSVP, suggest times | — this one has full access, see below |
| Drive | Search files, read contents, download, create new files, copy files | Delete files. Overwrite your existing files. |
The email one matters most, so let me be specific. There is no send function. Not “we promise not to,” not “it’s disabled by default.” The capability doesn’t exist in what it’s allowed to do. Every draft it writes sits in your drafts folder until a human opens Gmail and hits send.
One heads-up: the Google permission screen you’ll see during setup does mention email sending, because Google’s permission categories are broad. I’ll warn you before it appears rather than after.
Calendar is the honest asterisk
Calendar genuinely can book, move, and cancel things. That’s the point of connecting it, and I’m not going to soften it.
What I do about it: when we set you up, I teach one rule and I repeat it until it’s annoying. Allow always for anything that only reads. Approve one at a time for anything that writes, sends, books, or deletes. It costs you an extra half-second and it means nothing moves on your schedule without you seeing it.
Does it train on my stuff?
Anthropic’s documentation says this directly: “We do not train our models on your Gmail, Drive, or Calendar connector data.”
But there’s a real exception and I’m going to tell you about it, because you’d be right to be annoyed if you found it later.
On a personal plan with model improvement turned on, content that you copy and paste into the chat yourself can be used to improve the models. The exclusion covers what the AI goes and fetches through the connector, not what you paste in by hand.
So two things happen at setup. We check that setting together on your screen and you decide, and I get you in the habit of letting it go fetch things rather than pasting them.
If you handle anything confidential, use a business plan
On Team and Enterprise plans, your data isn’t used for training at all by default. Those are the commercial products, and they’re also the ones covered by the security certifications: SOC 2 Type I and II, ISO 27001, ISO 42001, and a HIPAA-ready configuration with a business associate agreement available.
Worth knowing: those certifications cover the business plans. They do not cover the free and personal plans.
Team now starts at two seats, so a small office is looking at roughly $20 to $25 per person per month. If you’re a law office, an accounting practice, an insurance agency, a medical practice, or anyone else holding client-confidential material, that’s where I’ll point you. Not because it’s a bigger sale for me, since you buy your own subscription either way and I don’t make a dime on it.
How you turn it off
Two clicks, and I show you this before we connect anything, not after.
- In Claude: Connectors → the Google connector → Disconnect
- In Google: myaccount.google.com/connections → find Claude → Delete all connections
Do both. Disconnecting inside Claude leaves the permission sitting on the Google side, which is the kind of thing nobody mentions.
Questions I get
Can it read all of my email?
It can search and read your mail, yes, but only while it’s actively working on something you asked for. It isn’t sitting there indexing your inbox in the background. You ask it to find the Henderson thread, it goes and gets the Henderson thread.
Could it email a customer something wrong?
No. It can write email, it cannot send it. Everything goes to drafts and waits for you. Nothing leaves with your name on it that you didn’t read first.
What about my IT guy? He’s going to hate this.
Get him on the call, genuinely. If your Google Workspace has third-party app access restricted, an admin has to approve it before any of this works, so he’s involved either way. He’ll want to see the security certifications, and they exist. I’d rather have him on my side than be the guy who went around him.
Where does my data physically go?
Anthropic’s servers, encrypted in transit and at rest. Connector content is pulled while you’re actively asking for it. Retention follows the conversation, so deleting the chat removes it. If you need something stricter, like US-only processing or a custom retention window, that exists on Enterprise and it’s a real conversation we can have with them.
Is this going to replace my people?
Nothing I do touches your headcount, and I say this in front of your whole team, not just you. What we’re doing is taking the parts of the day nobody likes, the follow-ups and the retyping and the where-did-I-put-that, off their plate. If somebody ends up with an extra hour a day to do the work you actually hired them for, that’s the whole win.
Do I have to buy the software from you?
You can’t, actually, and I want to be clear about it. Reselling access isn’t allowed under Anthropic’s terms, and sharing accounts is explicitly prohibited. You create your own account, in your name, on your card, and you own it. I never hold your login. I charge for my time and nothing else.
One thing worth knowing that most people selling this won’t tell you: connecting Gmail, Calendar, and Drive works on the free plan. You don’t have to pay for a subscription to do the thing I’m setting up. There are good reasons to upgrade, and I’ll walk you through them honestly, but “you need to pay for it to connect Google” isn’t one of them.
What if my Google Workspace blocks it?
Fairly common, especially if someone set up your IT years ago and locked things down. You’d see a message about your admin needing to review the app. The fix is on the admin side in the Google Admin console, and Google says changes can take up to 24 hours to take effect, though it’s usually faster. This is exactly why I ask about it on the call beforehand instead of discovering it while you’re paying me to stand in your office.
Sources
Everything above comes from these. I’d rather you check than take my word for it.
- Anthropic: Use Google Workspace connectors — connector capabilities, training exclusion, admin fix
- Anthropic Privacy Policy — consumer data use and the opt-out
- Anthropic: Is my data used for model training? — commercial plans
- Anthropic Consumer Terms — account sharing and resale prohibitions
- Anthropic: Certifications — SOC 2, ISO 27001, ISO 42001, HIPAA
Still got a question I didn’t answer?
Ask me on a call. If I don’t know, I’ll tell you I don’t know and then go find out, which is more than you’ll get from most people in this business.