Not another chatbot

You ask ChatGPT to draft a sensitive email to a customer. It gives you something fluent. You read it. You close the tab and write the email yourself.

You ask Claude to update a strategy document in place. It gives you a new version. You read it. You open the document and apply the two edits that actually mattered, by hand.

The pattern is consistent. The higher the stakes, the faster you give up on the assistant and do the work by hand.

It's a constraint problem.

Two kinds of knowledge work

There is work where iteration is cheap. Brainstorming. First drafts of low-stakes copy. Code you will review line by line before running. Research notes. The assistant is useful because the cost of a wrong output is a few seconds of your attention.

There is also the other kind. Work where the output has to end up inside a document someone else will read, commit to, or act on. The email the customer will forward. The spec the engineering team will build against. The roadmap the board will plan against. The PRD that determines what ships.

The second category is most of a senior operator's week. The assistant fails it, every time, for the same reason: it produces a reply, in a chat window, detached from the doc or thread the decision is about.

Why the category is the problem

Work in the second category has three demands a chat interface cannot meet.

The first is context that holds. The documents, the history, the prior decisions, the people on the thread: the assistant needs all of it, and it needs it across conversations, not just in the prompt you remember to paste.

The second is action bound to the document itself. The output is an edit to the spec, a revision to the roadmap, a draft in the customer thread. A reply in a chat window forces the human to do the last and most fragile step: translating a suggestion back into the actual system of record.

The third is discipline around commit. The edits must be proposed, reviewable, and unambiguously gated on a human yes. An assistant that sends emails or overwrites documents on its own authority is worse than useless at this altitude. One wrong move and the trust is gone for a year.

A chat interface is either useless or dangerous. There is no middle ground, not when it's just chat.

Why narrowing wins

Senza is narrow by intent. The wedge: durable work state underneath the documents and plans you already use, context that carries across conversations, action proposed against those documents directly, waiting on a human yes.

That is what the second category of work actually needs. The chatbots are playing a different game: producing fluent replies in a window next to the work. Senza is playing the game of getting the right edit onto the right doc, plan, or thread, reviewable, every time.

Those are different games.