The chat-with-your-data craze died for a reason. Users opened the chat box once, asked a confused question, got a confused answer, and never came back. Every product team that shipped 'an AI chatbot' is now removing it.
The pattern that won: the copilot. Embedded directly in the workflow, suggesting actions inline, never asking the user to think in prompts.
- Chatbot: Empty input field. User must invent the query. Most users don't know what to ask.
- Copilot: Sees user context. Suggests next action. User accepts or rejects with one click.
- Chatbot adoption: 5-15% of users try it once, 1-3% return weekly.
- Copilot adoption: 50-80% of users use it daily because it's in the path of their existing workflow.
The 5 copilot patterns that work
Inline suggestions (the GitHub Copilot pattern)
AI offers a completion right where the user is typing. Tab to accept, esc to dismiss. Zero friction.
Smart defaults (the Linear pattern)
User clicks 'create issue' - the title, description, labels, and assignee are pre-filled by AI from the conversation that triggered it. User edits or accepts.
Action buttons (the Notion AI pattern)
Highlight text, get 4 contextual buttons: 'rewrite', 'summarize', 'translate', 'expand'. Each one runs a focused prompt the user never has to write.
Background drafts (the SmartLead pattern)
AI drafts the email/document in the background. User opens it and sees a half-written version, not a blank page.
Suggest-then-confirm (the Stripe Workbench pattern)
AI proposes a refund, change, or operation. User reviews the diff and clicks confirm. Never executes without human approval on irreversible actions.
Patterns that fail
Floating chat bubble in the corner
Users banner-blind it within 2 sessions. Death by ignore.
'Ask me anything' empty state
Users have no idea what to ask. Should be replaced with 3 specific suggested prompts based on context.
Chat history that doesn't persist or sync
If the conversation isn't synced across devices, users won't trust it for real work.
AI that asks clarifying questions before doing anything
One round of clarification is fine. Two is annoying. Three and the user gives up.
"Users don't want to invent prompts. They want suggestions they can accept with one click."
Chatbot vs copilot - by the numbers
| Metric | Chatbot UX | Copilot UX |
|---|---|---|
| Adoption (try once) | 5-15% | 60-80% |
| Adoption (daily active) | 1-3% | 40-60% |
| Time to first value | 30+ seconds | Under 5 seconds |
| Cognitive load | High (invent prompt) | Low (review + accept) |
| Discoverability | Hidden in a corner | In the user's path |
| Failure mode | Bad question = bad answer | Bad suggestion = ignore |
The 6 copilot patterns we ship
Inline completion
AI offers a completion at the cursor. Tab to accept. The GitHub Copilot model.
Smart defaults
Form fields pre-filled from context. User edits or accepts. The Linear pattern.
Action buttons
Highlighted text gets contextual buttons (Rewrite, Summarise, Translate). The Notion AI pattern.
Background drafts
Documents pre-drafted in background. User opens to a half-written version, not blank.
Suggest-then-confirm
AI proposes an action. User reviews diff. Confirms or rejects. Critical for irreversible actions.
Embedded quick search
Cmd+K opens semantic search across the user's data. Returns actions, not just results.
We rebuild chatbots into copilots.
Same models, same RAG, same data. Different UX. Adoption typically rises 5-10x.
Stop the chatbot, start the copilot
- Empty chat boxes are the new banner ads - users learn to ignore them.
- Embed AI in the workflow, not next to it.
- Suggest, don't ask. One-click acceptance beats prompt invention.
- Always show the diff before irreversible actions.
- If your AI feature has a 'How can I help?' empty state, redesign it.
Build the copilot, kill the chatbot
The chatbot is a 2023 idea. The copilot is what users actually want in 2026. The technical work is similar - same models, same RAG patterns, same eval harness. The product work is completely different.
If you're starting an AI feature today, the question isn't 'where does the chat box go?' It's 'what specific action can we suggest with one click?'
The IRPR engineering team ships production software for 50+ countries. Idea → Roadmap → Product → Release. 200+ products live.
About IRPR