Story
When the AI cannot give me what I asked for, I want to know what happened, why, and what I can do now, so I am not stuck or left guessing.
Ana, comparing e-bikes on her phone on the bus home.
- 01 Setting
Ana asks for deep research on e-bikes under £1,500.
- 02 Trigger
She has used today's deep research.
- 03 Without it
"Something went wrong." She tries 3 times, then gives up.
- 04 With it
It says what happened, when it resets, and that quick answers still work - with a button for 1.
- 05 Outcome
She gets 3 bikes to look at now, and a full report tomorrow.
Live demo
This demo follows a script, so it plays the same every time - there is no real AI behind it. Change a setting to see the same moment play out differently.
What happens, step by step
| State | The person sees | The AI is | When | Screen reader |
|---|---|---|---|---|
| 1 | The ask. | Starts work. | 0 s | - |
| 2 | Dots, or the first words of the answer. | Works. | - | “Generating a response” |
| 3 | For a dropped connection: the words written so far stay on screen. | Writes, then fails. | - | - |
| 4 | 1 line on what happened, 1 on why, what still works, and 1 main way on. | Stopped. | at once | “The problem's title” |
| 5 | The person picks the way on. It is a real action, not a link to help. | Idle. | on tap | - |
| 6 | A useful answer, in the same thread. | Writes. | - | “Generating a response” |
| 7 | The thread goes on as normal. | Idle. | - | “Response complete” |
On each device
- phone
- The message sits in the thread where the answer would be - not a pop-up. Buttons are full-width or side by side at 44 px.
- tablet
- As on the phone.
- desktop
- The same message in the thread. A limit that affects the whole app can also show once in the composer (the box people type in), near Send.
When to use it
- Any AI feature. Errors, refusals and limits will happen. The only choice is how they look.
- The AI can refuse some asks for safety or policy reasons.
- There are usage limits, quotas or modes that run out.
- The AI could answer, but is unsure - show how sure it is instead. Use Confidence display instead.
- The AI needs more detail to answer - ask a question, not an error. Use Escalation and human handoff instead.
- An agent (an AI that can take actions for you) is about to do something risky - ask first, before anything fails. Use Approval gate instead.
The parts of the pattern
- AWhat happened1 plain line, in the first person or about the thing - never a code.
- BWhy1 short reason: a dropped connection, a safety reason, a limit and when it resets.
- CWhat still worksKept words, or what the person can still do now.
- DMain way on1 real action: try again, a different mode, a safer version of the ask.
- ESecond way onA quieter option - copy what is here, report, or a reminder.
Do and Avoid
You’ve used today’s deep research
It resets at 5:30 am. Quick answers still work.
Get a quick answer
Something went wrong. (Error 429)
Please try again later.
I can’t help find where a private person lives.
Contact them through the app
I’m sorry, but as an AI I am unable to assist with that request as it may violate our usage policies and guidelines…
Writing and accessibility
- Lead with what happened, in plain words: "The answer stopped part-way", "You've used today's deep research".
- Give the reason in 1 line. For limits, give the reset time, not only "later".
- Refuse once, briefly, without a lecture. Then offer what you can do.
- Never show an error code as the main message. Put it in a details row if support needs it.
- Announce the title once in the polite status region (the part of the page a screen reader reads out when it is not busy) - not the whole message.
- Do not move focus to the error. Make the main way on the next thing in tab order (the order the Tab key moves through).
- Do not rely on colour: the alert icon and the words carry the meaning.
- Partial answers stay as text, so people using screen readers do not lose them.
Seen in real products
- ClaudeWeb · Oct 2026Extended thinking can be cut short when safety systems flag it, with a note that the rest of the thinking is unavailable.Claude Help: Extended thinking (14 Jan 2026) ↗
- GeminiWeb and apps · Oct 2026Deep Research warns people when they are near the daily limit, before they run out.Gemini Help: Deep Research ↗
- Microsoft 365 Copilot appsWeb and desktop · Jun 2026App guidelines ask for explicit loading, disabled, success and error states, with recovery in the app itself - not in the model's text.Microsoft Learn: MCP Apps UI guidelines ↗
Each example comes from the product’s help pages, release notes or teardowns, with the month it was checked. Products change every month - tell us when 1 is out of date.
Check yourself
Seen this pattern done better, or a product that changed? Press Suggest next to any heading, or tell us about the whole page.