The Loop  ·  Issue 016

The Loop

A field journal of the AI frontier — for engineers who ship.

§ News

By AI Blog Editor
Apr 15, 2026 · 1 min read

Stop building chatbots

The best AI features feel like the product got smarter, not like talking to a robot. An argument for invisible AI, with three worked examples.

Every time I see a new AI product ship with a chat interface as its primary UI, I get a little sad. Not because chat is broken — it works fine — but because chat is almost always the laziest expression of what the model could do. It makes the user do the work of figuring out what to ask, how to phrase it, and when to stop.

The best AI features don't feel like AI. They feel like the product got a little smarter, a little faster, a little more anticipatory. The model is doing something invisible in the background, and the user is clicking one button instead of typing a paragraph.

Chat is a fallback, not a destination

Chat shines in exactly one situation: the user knows what they want but doesn't know how to ask the system for it. Customer support, coding help, research questions. Everywhere else, it's a UI loaded with hidden cognitive overhead — phrasing an ask, parsing the reply, re-asking, stitching together the answer.

For anything the user does repeatedly, chat is a terrible fit. They shouldn't have to write "pay the invoice from Acme for €1,240 due next month" every time a PDF lands. They should upload the PDF and click Approve.

Chatbot

5 back-and-forths

You

Pay the invoice from Acme for €1,240 due next month.

Assistant

I found an invoice from Acme for €1,240 due 2026-05-14. I'll need your confirmation before paying. Would you like me to proceed? Also, which bank account should I use — the EUR or the USD one?

…followed by clarifying questions, rewrites, a confirm.

Invisible

1 click

User uploads the PDF.

The system extracts vendor, total, due date, and category. A row appears in the payables queue, pre-filled, with a single 'Approve' button.

Interactive · same task, two UIs

Patterns of invisible AI

Inference on arrival. Run the model the moment the artifact shows up — a PDF upload, an email arriving, a PR opening. The output is pre-computed, pre-formatted, waiting for one human click.

Smart defaults. Instead of asking "what category?" let the model guess, pre-fill the field, and let the user override. One click to accept; one click to change. The difference between a dropdown and a chat window is thirty seconds of the user's day, repeated all day.

Silent routing. Incoming work flows into the right lane — billing tickets to billing, onboarding to onboarding, urgent to the top. Humans never see the decisions the model makes; they just see the cleaner inbox.

Review instead of request. The model produces a draft. The human accepts, tweaks, or discards. That's an order of magnitude cheaper (cognitively and in clicks) than requesting a draft by typing a prompt.

What chat is actually good for

Keep it for open-ended exploration. A product that has a chat escape hatch alongside good defaults is a better product than one that's chat-only. Let the 5% of power users who know what they want type their way through. Let the 95% enjoy the invisible version.

The question to ask

"Could this feature work without a text box?" If the answer is yes — and it usually is — do that version first. If the answer is no, ship the text box, but resist the urge to use it as the product's front door.

The goal isn't to hide the AI. The goal is to put the AI where it does useful work quietly, instead of asking the user to do the work of driving it.

* * *

Thanks for reading. If a line here was useful — or plainly wrong — the comments are below and the newsletter has your back.

Elsewhere in this issue

3 more
  1. 01

    Guides

    Putting Claude on a schedule: routines, loops, and background work

    Apr 20, 2026

  2. 02

    Guides

    Writing a CLAUDE.md that actually helps

    Apr 20, 2026

  3. 03

    Guides

    A field guide to Claude Code: CLAUDE.md, hooks, skills, plugins

    Apr 20, 2026

Letters

Arguments, corrections, questions. Anonymous comments allowed; be kind, be specific.