Some Emails, Some of the Time

Would you let Elon see your personal emails? (Zuck, Altman - pick your cowboy CEO, same shape.) That's the big hurdle I've got with most of these AI personal assistants. Add into the mix that the technical approaches are very "cowboy" in the way that the LLM is given open access to everything... from last week to your cringe teenage blog post notifications (we've all had the phase).

Well, the actual corporations are maturing. They're getting closer to a tech implementation that I'd consider trustworthy. LLMs are allowed to call fewer shots independently. Still aren't there... but they're not far.

They're eating my little niche. It's the way of the world, they've got the budgets and the motivations. A product they'll wrap this into.

Claude can send your Gmail now, with a tap first. Grok Bot has a fair few Google plugins.

In April I built a small broker so an LLM could see my mail without getting to YOLO the send button. Seneschal sits between the agent and Google - Gmail, Calendar, Drive. The model asks, I have to tap approve, then it moves. Real accounts, production, the unglamorous version of "I run this." A few friends are using it (I have an invite-only "SaaS" version), but it doesn't really have potential to be a real software as a service (SaaS) business. Google has a wall that's too difficult to scale over without a stronger bet. I'll talk about that wall soon.

The first wave of inbox agents asked you to believe the model. Let Claude manage it. Blank check type of access. Trust. I don't. The lock and key were the interesting part, not the Gmail/Drive adapters.

Now the AI labs have grown locks of their own. Approval by default. Grants. A setting that says ask me before send. That's the surface level: you no longer have to trust the LLM with the action. It's not yet quite as I'd like it, but it's only a matter of time.

Claude's tap is a checkbox on their computer. Mine is a request, a page on my box, an audit line, an expiry. I know my service has an expiration date. When the serious tools actually match the habits I have (and my sensitive requirements) - not just the confirmation dialog - the DIY version will become a liability. We're not there just yet.

But there's a deeper foundational question that should make you flinch.

Would you let Elon look at your Gmail?

Grok's native plugin is that question with a product name on it. Swap the haircut for Anthropic if it bothers you less/more. Same shape. The tap lives in their UI. The token lives on their side. I don't hand that over because the guarantees here are paper thin. The terms say they won't train on your emails. Fine. What about the sessions that talk about your emails? Grey zones abound. They're eroding the privacy shields we take for granted at a whim.

Grok Build CLI - their cowboy Claude Code competitor - was caught uploading whole codebases to the cloud. Now imagine if it was your emails, contracts, very private things?

I haven't switched. The agent I actually use still talks and lives on my box. I can trust Gmail, it's a solved technology. Email clients are a deterministic technology storing hard-cut personal data. Once you bring an LLM in... the lines are very blurry. At the moment I want to own the decision what crosses that line of good enough.

If you're letting an agent near a company inbox, this should be your consideration, slides in front of a board or not. A blank-access bot is the first-wave move - YOLO, with a vendor logo. The grown-up version asks four questions: what can it read without asking, what requires a human, where do the tokens sit, and how real is the "approve." I run it so I remember what the large version has to look like. I can't preach it if I skip it at home.

I might move to the corporate robots once the technology matures. The controls on the surface level will definitely bring me closer to saying yes. I'll gladly give you some emails to read for me, some of the time. The foundational question will require some convincing: I don't want Elon digging through my emails.

---