AI-first, with no AI in the room

12Gradus is built AI-first. It is also, deliberately, a product with no AI anywhere near the relationship between a sponsor and a sponsee.

Those two sentences look contradictory. They aren't, and the distinction is worth spelling out, because "AI-powered recovery app" is a phrase that should make anyone in the rooms nervous.

Where the AI actually is

Behind the product, not inside it.

Engineering. Most of the code in 12Gradus is written with AI agents and reviewed by a human. That includes the unglamorous work small teams usually skip: the adversarial test suite that tries to break our own access rules, the migration files, the responsive fixes at 320 pixels wide, the error paths nobody sees. A one-person company that ships a payments integration, a carrier-registered SMS bridge, per-operation database security and a web app is doing it with leverage. This is the leverage.

Operations. Catalogue updates, release checks, content builds, the boring recurring chores that otherwise get done late or not at all.

Words. Blog posts and help text start as drafts from a model and are edited by a person before they go out. This post did. We mention it because a company that writes about honesty should say so.

This website. The assistant in the bottom corner of every page reads the text of these public pages and answers from them. It has no access to the app, to any account, or to any recovery data — there's nothing for it to look anyone up in.

Where it will never be

Four commitments. These are the interesting part, because they're the ones that cost us features we could otherwise ship.

1. No model reads private recovery content. Journals, private reflections, step-work answers, chat between a sponsor and sponsee. Not for summarising, not for search, not for "insights", not for a helpful weekly digest. The obvious product idea here — let AI summarise your sponsee's week for you — is one we've decided against permanently. Somebody wrote those words to one person. Sending them to a third party, even a well-behaved one, is a betrayal of that even if it never leaks.

2. Nothing you write trains anything. Not now, not later, not anonymised, not aggregated. Recovery content is the single worst category of data to have floating in a training corpus, and the fact that it would probably be fine is not a good enough reason.

3. No AI stands in for a sponsor. No auto-replies to a sponsee. No generated encouragement sent under your name. No chatbot that does step work with someone at 3am. This is the one people ask for most, and it's the clearest no: the entire mechanism of sponsorship is that a person who has been there is choosing to pick up the phone. Simulating that hollows out the only part that works.

4. No scoring people. No relapse-risk percentages, no engagement grades, no algorithmic ranking of which sponsee needs you most this week. We could build it. The data is right there. We won't, for two reasons: it would be wrong often, and — worse — it would be trusted. A sponsor who reads "low risk" next to a name is a sponsor who makes one fewer phone call. Nobody should ever be talked out of a call by a number.

Why publish this at all

Because the alternative is a privacy policy nobody reads and a slide deck nobody sees.

Recovery communities have earned their suspicion of software. People have been surveilled, scored, marketed to, and had their worst weeks sold on as data. Any company arriving in this space with the word "AI" in its pitch owes a specific, checkable account of what it's doing — not a reassurance.

So the list above is a public one, on its own page, in plain language. If we ever cross it, you'll be able to point at the exact line. And if you catch us first, write to <a href="mailto:support@12gradus.com">support@12gradus.com</a> and we'll either fix it or take the page down. What we won't do is quietly edit the promise.

The uncomfortable bit

There's a version of this argument that says: if AI could genuinely spot the sponsee who's about to relapse, refusing to build it is a choice with a body count.

We've sat with that. Two answers.

The first is empirical: it can't, not reliably, not from the thin behavioural signal an app can see. What it can do is produce a confident-sounding number, and confident-sounding numbers change human behaviour far more than their accuracy deserves.

The second is about what the tool is for. 12Gradus doesn't need to predict anything. It needs to show a sponsor that it has been eleven days since they heard from someone, and then get out of the way. The sponsor already knows things about that person that no model will ever have — how they sounded on the phone last month, what their mother's illness did to them, what step they're stuck on and why.

Our job is to make sure that human notices. Not to replace what they'd notice.