Never face a blank chat box: let the page suggest the question

The blank box is a knowledge problem, not a typing problem

An empty prompt with a blinking cursor looks like the simplest interface anyone ever shipped. It is one of the hardest, because before you type a single character it quietly asks you to have already done three separate things: know what the system is capable of answering, know which of those answers you actually want right now, and phrase the request precisely enough to be understood. Three independent places to fail, and the box gives you feedback on none of them until after you've committed to a sentence and watched it come back empty.

Compare it to the thing everyone complains about — a menu of canned reports. A menu is clumsy, rigid, and usually missing the exact cut you want. But it does one thing a chat box cannot: it shows you the boundary. You learn what is answerable by reading, not by guessing and getting burned. Nielsen Norman Group has a name for the gap the prompt leaves behind — the articulation barrier: with a prompt-only interface you have to be articulate enough to describe your intent in prose before you get anything at all, and their proposed fix is a hybrid interface that pairs free text with GUI affordances rather than replacing them.

That's the whole design problem in one line. The fix for a blank box isn't a smarter parser on the other side of it. It's putting the question where the user already is.

Recognition beats recall — so the page does the asking

Recognizing a good question off a short list is a fundamentally cheaper mental operation than composing one from nothing. Recognition rather than recall is one of the oldest usability heuristics on the books, and a chat box is pure recall — the most expensive mode, offered at the moment you know least.

But a list only helps if it's a real affordance rather than a poster. A generic "try asking about your sales" teaches nothing; it restates the category you were already stuck inside. Suggestions have to be two things at once: page-aware, so they're about what's in front of you, and data-aware, so they're about your actual store rather than a demo one.

That's what the Ask Rail does. It suggests questions worth asking on every page, generated from your own data — questions like "Which items run out before my next order arrives?", "What did I sell most last weekend?", and "Which supplier has the longest lead time?" Read those three and you've learned the capability boundary without a single failed attempt. You now know Atlas can answer about stock, sales history, and supplier lead times. That's a tutorial disguised as a shortcut.

The number on screen is the prompt

There's a step past picking from a list, and it's the one that removes the last bit of work: you don't pick a question at all, you click the thing you were already staring at.

Sparkle marks sit next to the numbers on screen. Click one and you ask about that exact value — no typing. What that eliminates is the second failure mode of every analytics conversation: restating context the screen is already showing you. Without the mark, asking about a number means naming the product, the location, the date range, and the metric back to a system that just rendered all four. It's a transcription tax on a question you didn't even know you had until the mark appeared beside the figure.

The mark is doing something subtler than saving keystrokes, though. It is proposing the question. A number you would have scrolled straight past becomes one you're curious about, because something beside it implied there was more to know — and the answer that comes back in plain English from your live store data hands you the action that follows as a one-tap chip.

The second visit shouldn't start from zero either

Cold start isn't a one-time event. It happens every time you open the page, because whatever you asked last Tuesday is gone and the box is empty again. So the same principle has to apply on visit fifty, not just visit one.

In the AI Report Builder it does. Suggested questions greet you on every visit — "What are my best sellers this month?", "What should I reorder now?", "Where am I losing revenue to stockouts?" — and your recently asked reports are one click from running again, which turns your own history into the list you recognize from. Twelve ready-made templates cover sales, inventory, forecasting, and financials, so the capability boundary is legible before you type rather than after. There are no measure pickers and no filter trees; you ask, and Atlas answers with a live table.

The exit ramp matters as much as the entrance. An answer can carry a follow-up chip built from its own result — asking why one specific product in the list is slowing down, say — and you can refine what you got by talking to it. Which means the second question doesn't come from a blank box either. It comes from the answer to the first.

Where the question comes from when you don't have one.