Unpopular opinion: "no AI" is becoming a feature, not a gap

Every second launch here is AI native right now. I went the other way and shipped a tool with no AI in it at all. Because it is basically a text writing/automation tool, I expected that to be a problem. It's turned out to be the easiest thing to explain and actually push for when it comes to a tool claiming privacy as one of its features.

Here's what keeps happening. We get into a deal, things are going fine, and then someone from their security team asks whether our staff can read what they store in the product. For most tools in this category the answer is yes. Nobody's being shady about it, that's just what it means to keep data on a server you control.

The thing people underestimate is what piles up in a snippet library once a team shares one:

• support macros with account states quoted in them

• sales templates carrying pricing that hasn't shipped yet

• legal wording somebody spent two weeks getting approved

• runbooks with internal hostnames in them

• the security questionnaire answers, saved so nobody rewrites them from scratch

Read that back and it's basically a summary of how the company runs. It got there by accident, because it started as one person's typing shortcuts and nobody ever went back and classified any of it.

Teams make this worse rather than better, which took me a while to see. You centralize wording because you don't want eleven people sending eleven versions of the refund policy with three of them wrong about the window. Fair enough. But now all of it sits in one shared object with everyone's access on it, and somebody has to sign off on where that lives.

^ This is where I stopped being able to justify AI. A model that helps you write has to see what you're writing, that's the trade. And for a paragraph that already went through legal, I don't want a good paraphrase. I want the same string that got approved, every time. Boring and identical is the requirement.

So we went the other direction. Encryption happens on the device, the server only ever holds ciphertext (basically gibberish), no keystrokes get read. It costs us something real: lose your passphrase and it's gone, and we can't recover it for you either.

Any product offering both zero-knowledge and passphrase recovery is quietly doing one of them badly!

Obviously I'm not neutral here, I built TypeHalo. But I don't have a settled view on where the line sits.

Where do you put it? Is "the vendor can read our templates" fine with you, or is that something you'd reject a tool over?