How to show the AI is working without making people wait

You press "Summarise" on a long report. A grey spinner appears in the middle of the page and stays there. After fifteen seconds you click the button again, because you honestly cannot tell whether anything is happening. Now it is writing two summaries.

That small moment is the whole problem with AI features. A language model takes real time to think and write, often several seconds and sometimes much longer. You cannot always make it faster. You can decide what the person sees while it works, and that decision changes whether the wait feels like progress or like a broken button.

Why AI waits feel worse than normal waits

Most of the web trained us to expect answers in under a second. A search result, a page, a "liked" heart: all close to instant. AI breaks that habit in two ways.

First, the wait is longer. Writing a paragraph, reading a file or generating an image can take anywhere from a couple of seconds to a minute.

Second, the wait is unpredictable. The same button can take three seconds on one request and thirty on the next, depending on how much the model has to read and write. People handle a long wait they can predict much better than a short wait they cannot.

So the job is to answer the question every waiting person is silently asking: "Is it working, and how long will this take?"

The three time limits that still hold

The most useful rule here is more than thirty years old. In his 1993 book Usability Engineering, Jakob Nielsen set out three response-time limits, building on earlier work by Miller (1968) and Card and colleagues (1991). Nielsen Norman Group still publishes them as the three important limits:

Almost every AI feature lives past the 1-second line, and many go past 10. That tells you the design work is unavoidable. It is also the first of Nielsen's ten usability heuristics, "visibility of system status": the design should always keep people informed about what is going on, through feedback within a reasonable amount of time.

Pick the indicator by the length of the wait

Nielsen Norman Group's guidance on progress indicators (Katie Sherwin, 2014) gives a simple ladder:

• Under 1 second: show nothing extra. A spinner that flashes for a quarter of a second is more distracting than helpful.
• About 2 to 10 seconds: a looping indicator, such as a spinner, is fine, ideally with a few words that say what is happening ("Reading your document…").
• 10 seconds or more: use a percent-done indicator or another sign of real progress, and give people a way to stop.

The same article describes a University of Nebraska-Lincoln study where people who saw a moving progress bar were willing to wait, on average, three times longer than people who saw nothing. The wait was the same length. Only what they could see had changed.

Two things the article says to avoid are worth repeating, because AI products still do both: a static "Loading…" label that never changes (people cannot tell if it has frozen), and a "Please don't click again" warning. Show that the first click was accepted instead. That alone would prevent the double summary from the opening.

For a whole page that is still assembling, a skeleton screen helps. A skeleton is a grey outline of the layout that fills in as content arrives. Nielsen Norman Group's Skeleton Screens 101 (2023) recommends them for full-page loads of up to about 10 seconds, spinners for a single card or module, and a progress bar with a time estimate for anything longer.

Five patterns made for AI features

The ladder above was written for ordinary software. AI adds a few patterns of its own.

1. Stream the answer as it is written

Most large language model APIs can send the response piece by piece instead of all at once. Anthropic's documentation, for example, describes how you can stream a response incrementally using server-sent events, a standard way for a server to keep sending updates to the browser. In plain terms: the first words can appear on screen while the rest is still being written.

Streaming turns a long blank wait into reading time. The person starts taking in the answer instead of staring at a spinner. If your feature produces text, this is usually the single biggest improvement you can make.

2. Name the steps, not just the spinner

When the AI does several things in a row (searching, reading, drafting, checking), say which one it is on. A short line that changes, like "Searching 3 sources" then "Reading the contract" then "Writing the summary", tells people that time is being spent on their task. It also gives them a rough sense of how far along it is, even without a percentage.

Keep these lines honest. Only show a step the system is actually doing. Invented "thinking" messages that cycle on a timer are a small lie, and people notice when the same three phrases repeat.

3. Set expectations before the click

If a task usually takes a while, say so on or near the button: "Takes about a minute." Google's People + AI Guidebook advises teams to communicate what the product can and cannot do, clearly and early. Time is part of that. A person who knows a report takes a minute will go and make tea. A person who expected two seconds will start clicking things.

4. Let people keep working, and let them stop

For long jobs, do not lock the whole screen. Run the job in the background, show its status in one place, and let people carry on with something else. Always offer a clear Stop or Cancel. Sherwin's article recommends a way to stop for long processes, and for AI it matters even more: often someone realises halfway through that they asked the wrong question.

5. Show the partial result, then the final one

When you can, show something useful early: a first draft, the first image, the list of sources found. Then refine it. People can judge whether the AI is on the right track and stop it if not, which saves them time and saves you compute.

A worked example: the "Summarise report" button

Here is the same feature, designed twice.

Before

1. Person clicks "Summarise".
2. Button stays the same. A spinner appears in the middle of the page.
3. Nothing changes for 15 to 40 seconds.
4. The full summary appears all at once.

After

The model did not get faster. The person now knows, at every moment, that it is working, roughly what it is doing and what they can do about it.

Microcopy you can copy

Small words carry a lot of this. Here are some starting points. Change them to match what your system really does.