A user highlights a paragraph, taps your copilot’s suggestion and gets something confidently wrong. There is no undo and no way to tell where it came from. They close the panel, and most never open it again. The model did not fail them. The interface around it did.
If you are adding an assistant to an existing SaaS product, the interface decides whether people use it or switch it off. This post covers seven copilot patterns, a way to choose how much control each one gives the user and a table of when to use which. It is written for product and design leads.
Key takeaways
A copilot is a set of interface patterns, not a chat box. Pick the pattern per task.
Match the pattern to the cost of a wrong answer. Cheap and reversible tasks can run ahead of the user. Costly ones should wait for approval.
Every pattern needs three exits: dismiss, correct and undo.
Show what the copilot knows and how sure it is, in the place the user is already looking.
Measure acceptance, edits and undo, not only usage.
What is an AI copilot in a SaaS product?
An AI copilot is an assistant inside a product that suggests, drafts or performs part of a task while the user stays in charge. The user keeps the goal and the final call. The copilot handles the repetitive middle.
That definition matters because it separates a copilot from an autonomous agent and from a chatbot. A copilot works beside the user in the same screen. An agent works alone and reports back. A chatbot answers questions but usually cannot change anything.
Microsoft researchers set out 18 guidelines for human AI interaction in a CHI 2019 paper. Several apply directly to copilots: make clear what the system can do, make clear how well it can do it, support efficient dismissal, support efficient correction and make clear why it did what it did. Google’s People + AI Research group makes similar points in its guidebook chapters on feedback and control and on errors and graceful failure.
What are the seven AI copilot UX patterns?
There are seven patterns that cover most copilot features. Each gives the user a different amount of control and asks for a different amount of attention.
Pattern | What it looks like | Best for | Main risk |
|---|---|---|---|
1. Inline suggestion | Ghost text or a chip at the cursor | Writing, code, form fields | Distraction when wrong |
2. Side panel chat | A conversation next to the work | Open questions, exploration | Context gets lost |
3. Command bar | A prompt box that runs one action | Power users, known tasks | Users do not know what to type |
4. Contextual action | A button on a selected object (“Summarize”) | Single clear jobs | Buttons multiply |
5. Draft and review | The copilot produces a draft the user edits | Longer outputs | Rubber stamping |
6. Proactive nudge | A prompt that appears unasked | Spotting issues | Annoyance |
7. Background action | The copilot acts and reports after | Low risk, reversible work | Silent mistakes |
Most products need three or four of these, not all seven. A reporting tool might use a command bar, contextual actions and draft and review. A writing tool might lean on inline suggestions and a side panel.
Also read: AI Medical Scribe UX Teardown: What Abridge, Suki, Nuance DAX and Freed Get Right
How do you choose the right pattern? The Copilot Control Ladder
Choose by asking how costly a wrong output is and how easily the user can undo it. We call this the Copilot Control Ladder. It has three rungs, and each pattern sits on one.
Rung | The copilot may | Patterns | Use when |
|---|---|---|---|
Suggest | Offer, never apply | Inline suggestion, nudge, contextual action preview | A wrong answer is costly or hard to reverse |
Approve | Prepare a change the user confirms | Draft and review, command bar with preview | The output is long or touches shared data |
Act | Apply the change, then report | Background action with undo | A wrong answer is cheap and reversible |
Start every new capability on the Suggest rung. Move it up only when users accept it most of the time and undo it rarely. Never move up because the model got better in a benchmark. Move up because users showed you they trust it.
The ladder also gives your team a shared word for arguments. “Should this auto apply?” becomes “Is it cheap and reversible enough for the Act rung?”
Also read: Onboarding and Empty States for AI Tools, How to Design Trust Into AI Interfaces, AI Medical Scribe UX Teardown.
To see how we think about these decisions in practice, our AI SaaS dashboard design post shows controls in real interfaces.
What does a good copilot interface get right?
A good copilot makes three things visible at the point of use: what it can do, how sure it is and how to undo. Put them in the same screen as the output, not in a help page.
Scope. A one line prompt such as “Ask about your pipeline, drafts or reports” beats a blank input. Blank inputs cause the cold start problem, where users do not know what is possible.
Confidence. Mark generated content so people know what to check. Label by source and by state (“Draft”, “From your notes”), and show how the answer was reached when a decision depends on it.
Exits. Every suggestion needs a dismiss action that costs one click and a correction path that does not force a restart.
Pro tip: Put the dismiss action closer to the thumb or cursor than the accept action on patterns that appear unasked. People accept less when they feel they cannot refuse, and they stop trusting the panel altogether.
How do you design for errors and corrections?
Plan for the wrong answer first, because it will happen. PAIR’s guidance on errors and graceful failure separates errors the system knows about from errors it does not, and asks designers to give users a way forward in both cases.
Use this short checklist for each copilot feature:
What does the user see when the model has low confidence or no answer?
What can the user change without starting over?
How does the user get the previous state back?
Where does a correction go, and does the copilot learn from it or only for the session? Say which.
Who is told when something with side effects was applied?
Corrections are also your best product data. Log accept, edit and reject separately. A high accept rate with a high edit rate means the draft is a starting point. A high reject rate with low usage means the pattern is wrong, not the model.
What are the most common AI copilot UX mistakes?
Defaulting to chat. A chat panel is flexible, but it makes the user write the prompt. Use contextual actions where the task is known.
No undo on background actions. Acting without a way back converts one wrong output into lost trust.
Hiding the scope. If users cannot tell what it can do, they test it with the wrong tasks and conclude it is weak.
Interrupting at the wrong time. A nudge during typing costs more than a nudge after a pause.
Treating usage as success. Opens and messages count curiosity. Accepted outputs count value.
One pattern for every task. The same assistant can suggest on a draft and act on a tag. Give each task its rung.
How do you test your copilot patterns this week?
Run a five person task test on your riskiest pattern before you ship it.
Pick one task where a wrong output matters.
Give five users the task with the copilot, seeded with one wrong output.
Watch whether they notice, how they fix it and whether they can get back.
Pass if four of five spot the error and recover without help.
If fewer pass, move the pattern down one rung on the ladder and try again.
Pro tip: Seed the wrong output on purpose. Users who never see an error tell you nothing about recovery.
Also read: How to Show AI Accuracy and Trust on a Landing Page
To see how we work on AI product interfaces, read our ORYN AI agent monitoring landing page and Nura healthcare platform dashboard case studies, or browse all of our case studies.
Keep reading: How to Design Trust Into AI Interfaces: 8 Moments Where Users Decide, Onboarding and Empty States for AI Tools: How to Get Users to a First Win and AI Medical Scribe UX Teardown: What Abridge, Suki, Nuance DAX and Freed Get Right
Which of your copilot patterns is on the wrong rung?
Send us a screen of your assistant and we will mark which pattern fits each task and where the exits are missing. Book a design review · See our work
AI copilot UX patterns FAQs
Seven patterns cover most features: inline suggestion, side panel chat, command bar, contextual action, draft and review, proactive nudge and background action. They differ in how much control the user keeps. Choose by the cost of a wrong output and how easily the user can undo it.
A copilot works beside the user in the same screen and leaves the final decision to them. An agent takes a goal and works on its own, then reports back. Many products need both, with copilot patterns for risky work and agent patterns for safe, reversible tasks.
Not by default. Chat suits open questions and exploration. For known tasks, a contextual action or inline suggestion is faster because the user does not write a prompt. Many strong products use chat for questions and buttons or suggestions for repeat work.
Show what it can do, how sure it is and how to undo. Label generated content, show the source where a decision depends on it, and make dismissal and correction one click. Trust grows when users can recover cheaply from mistakes.
Yes. Kraftbase is a product messaging and UI/UX design studio for AI and B2B SaaS companies. We design copilot panels, dashboards and the marketing sites that explain them, so the words and the interface come from the same decisions.
Track accepted outputs, edits per accepted output, rejects and undo, alongside retention of users who tried it. Opens and message counts show curiosity, not value. A rising accept rate with a stable undo rate is the signal to move a feature up the control ladder.
Only when a wrong result is cheap and easy to reverse, such as tagging or sorting. Anything touching shared data, money, customers or health records should wait for approval. Start every feature on the suggest rung and move it up when user behavior supports it.
About the author
Aneri is a designer at Kraftbase, a product messaging and UI/UX design studio in Vadodara, India, that helps AI and B2B SaaS teams in the US, UAE and India explain and design what they’ve built.
Sources
Copilot Control Ladder: our own framework, not a published standard
Design review
Which of your copilot patterns is on the wrong rung?
Send us a screen of your assistant and we will mark which pattern fits each task and where the exits are missing.
A user highlights a paragraph, taps your copilot’s suggestion and gets something confidently wrong. There is no undo and no way to tell where it came from. They close the panel, and most never open it again. The model did not fail them. The interface around it did.
If you are adding an assistant to an existing SaaS product, the interface decides whether people use it or switch it off. This post covers seven copilot patterns, a way to choose how much control each one gives the user and a table of when to use which. It is written for product and design leads.
Key takeaways
A copilot is a set of interface patterns, not a chat box. Pick the pattern per task.
Match the pattern to the cost of a wrong answer. Cheap and reversible tasks can run ahead of the user. Costly ones should wait for approval.
Every pattern needs three exits: dismiss, correct and undo.
Show what the copilot knows and how sure it is, in the place the user is already looking.
Measure acceptance, edits and undo, not only usage.
What is an AI copilot in a SaaS product?
An AI copilot is an assistant inside a product that suggests, drafts or performs part of a task while the user stays in charge. The user keeps the goal and the final call. The copilot handles the repetitive middle.
That definition matters because it separates a copilot from an autonomous agent and from a chatbot. A copilot works beside the user in the same screen. An agent works alone and reports back. A chatbot answers questions but usually cannot change anything.
Microsoft researchers set out 18 guidelines for human AI interaction in a CHI 2019 paper. Several apply directly to copilots: make clear what the system can do, make clear how well it can do it, support efficient dismissal, support efficient correction and make clear why it did what it did. Google’s People + AI Research group makes similar points in its guidebook chapters on feedback and control and on errors and graceful failure.
What are the seven AI copilot UX patterns?
There are seven patterns that cover most copilot features. Each gives the user a different amount of control and asks for a different amount of attention.
Pattern | What it looks like | Best for | Main risk |
|---|---|---|---|
1. Inline suggestion | Ghost text or a chip at the cursor | Writing, code, form fields | Distraction when wrong |
2. Side panel chat | A conversation next to the work | Open questions, exploration | Context gets lost |
3. Command bar | A prompt box that runs one action | Power users, known tasks | Users do not know what to type |
4. Contextual action | A button on a selected object (“Summarize”) | Single clear jobs | Buttons multiply |
5. Draft and review | The copilot produces a draft the user edits | Longer outputs | Rubber stamping |
6. Proactive nudge | A prompt that appears unasked | Spotting issues | Annoyance |
7. Background action | The copilot acts and reports after | Low risk, reversible work | Silent mistakes |
Most products need three or four of these, not all seven. A reporting tool might use a command bar, contextual actions and draft and review. A writing tool might lean on inline suggestions and a side panel.
Also read: AI Medical Scribe UX Teardown: What Abridge, Suki, Nuance DAX and Freed Get Right
How do you choose the right pattern? The Copilot Control Ladder
Choose by asking how costly a wrong output is and how easily the user can undo it. We call this the Copilot Control Ladder. It has three rungs, and each pattern sits on one.
Rung | The copilot may | Patterns | Use when |
|---|---|---|---|
Suggest | Offer, never apply | Inline suggestion, nudge, contextual action preview | A wrong answer is costly or hard to reverse |
Approve | Prepare a change the user confirms | Draft and review, command bar with preview | The output is long or touches shared data |
Act | Apply the change, then report | Background action with undo | A wrong answer is cheap and reversible |
Start every new capability on the Suggest rung. Move it up only when users accept it most of the time and undo it rarely. Never move up because the model got better in a benchmark. Move up because users showed you they trust it.
The ladder also gives your team a shared word for arguments. “Should this auto apply?” becomes “Is it cheap and reversible enough for the Act rung?”
Also read: Onboarding and Empty States for AI Tools, How to Design Trust Into AI Interfaces, AI Medical Scribe UX Teardown.
To see how we think about these decisions in practice, our AI SaaS dashboard design post shows controls in real interfaces.
What does a good copilot interface get right?
A good copilot makes three things visible at the point of use: what it can do, how sure it is and how to undo. Put them in the same screen as the output, not in a help page.
Scope. A one line prompt such as “Ask about your pipeline, drafts or reports” beats a blank input. Blank inputs cause the cold start problem, where users do not know what is possible.
Confidence. Mark generated content so people know what to check. Label by source and by state (“Draft”, “From your notes”), and show how the answer was reached when a decision depends on it.
Exits. Every suggestion needs a dismiss action that costs one click and a correction path that does not force a restart.
Pro tip: Put the dismiss action closer to the thumb or cursor than the accept action on patterns that appear unasked. People accept less when they feel they cannot refuse, and they stop trusting the panel altogether.
How do you design for errors and corrections?
Plan for the wrong answer first, because it will happen. PAIR’s guidance on errors and graceful failure separates errors the system knows about from errors it does not, and asks designers to give users a way forward in both cases.
Use this short checklist for each copilot feature:
What does the user see when the model has low confidence or no answer?
What can the user change without starting over?
How does the user get the previous state back?
Where does a correction go, and does the copilot learn from it or only for the session? Say which.
Who is told when something with side effects was applied?
Corrections are also your best product data. Log accept, edit and reject separately. A high accept rate with a high edit rate means the draft is a starting point. A high reject rate with low usage means the pattern is wrong, not the model.
What are the most common AI copilot UX mistakes?
Defaulting to chat. A chat panel is flexible, but it makes the user write the prompt. Use contextual actions where the task is known.
No undo on background actions. Acting without a way back converts one wrong output into lost trust.
Hiding the scope. If users cannot tell what it can do, they test it with the wrong tasks and conclude it is weak.
Interrupting at the wrong time. A nudge during typing costs more than a nudge after a pause.
Treating usage as success. Opens and messages count curiosity. Accepted outputs count value.
One pattern for every task. The same assistant can suggest on a draft and act on a tag. Give each task its rung.
How do you test your copilot patterns this week?
Run a five person task test on your riskiest pattern before you ship it.
Pick one task where a wrong output matters.
Give five users the task with the copilot, seeded with one wrong output.
Watch whether they notice, how they fix it and whether they can get back.
Pass if four of five spot the error and recover without help.
If fewer pass, move the pattern down one rung on the ladder and try again.
Pro tip: Seed the wrong output on purpose. Users who never see an error tell you nothing about recovery.
Also read: How to Show AI Accuracy and Trust on a Landing Page
To see how we work on AI product interfaces, read our ORYN AI agent monitoring landing page and Nura healthcare platform dashboard case studies, or browse all of our case studies.
Keep reading: How to Design Trust Into AI Interfaces: 8 Moments Where Users Decide, Onboarding and Empty States for AI Tools: How to Get Users to a First Win and AI Medical Scribe UX Teardown: What Abridge, Suki, Nuance DAX and Freed Get Right
Which of your copilot patterns is on the wrong rung?
Send us a screen of your assistant and we will mark which pattern fits each task and where the exits are missing. Book a design review · See our work
AI copilot UX patterns FAQs
Seven patterns cover most features: inline suggestion, side panel chat, command bar, contextual action, draft and review, proactive nudge and background action. They differ in how much control the user keeps. Choose by the cost of a wrong output and how easily the user can undo it.
A copilot works beside the user in the same screen and leaves the final decision to them. An agent takes a goal and works on its own, then reports back. Many products need both, with copilot patterns for risky work and agent patterns for safe, reversible tasks.
Not by default. Chat suits open questions and exploration. For known tasks, a contextual action or inline suggestion is faster because the user does not write a prompt. Many strong products use chat for questions and buttons or suggestions for repeat work.
Show what it can do, how sure it is and how to undo. Label generated content, show the source where a decision depends on it, and make dismissal and correction one click. Trust grows when users can recover cheaply from mistakes.
Yes. Kraftbase is a product messaging and UI/UX design studio for AI and B2B SaaS companies. We design copilot panels, dashboards and the marketing sites that explain them, so the words and the interface come from the same decisions.
Track accepted outputs, edits per accepted output, rejects and undo, alongside retention of users who tried it. Opens and message counts show curiosity, not value. A rising accept rate with a stable undo rate is the signal to move a feature up the control ladder.
Only when a wrong result is cheap and easy to reverse, such as tagging or sorting. Anything touching shared data, money, customers or health records should wait for approval. Start every feature on the suggest rung and move it up when user behavior supports it.
About the author
Aneri is a designer at Kraftbase, a product messaging and UI/UX design studio in Vadodara, India, that helps AI and B2B SaaS teams in the US, UAE and India explain and design what they’ve built.
Sources
Copilot Control Ladder: our own framework, not a published standard
Design review
Which of your copilot patterns is on the wrong rung?
Send us a screen of your assistant and we will mark which pattern fits each task and where the exits are missing.