Microsoft Copilot Skill

Microsoft Copilot Skill

Microsoft wanted to integrate help and learning content into Copilot to provide accurate, timely answers to questions about how to use products like Word and Excel. The goal was to build a solution to achieve speed to market.

Microsoft wanted to integrate help and learning content into Copilot to provide accurate, timely answers to questions about how to use products like Word and Excel. The goal was to build a solution to achieve speed to market.

OVERVIEW

The challenge

When a user asks for instructions on how to do something in Microsoft Copilot like how to add a watermark, an answer is generated using the large language model (LLM). This is costly and could require a wait. The business wanted to integrate help and learning content as a skill. The goal was to provide fast, accurate and version relevant instructional guidance to help enterprise and consumer users easily complete tasks (and achieve goals) using Microsoft products - improving both user experience and satisfaction.


The business objective was reduced wait time < 1 second.


Note: A skill is a tool that augments AI. It interacts with APIs from software or services to retrieve real-time info, include business data or perform computations.

My role

Led design of the MVP / short-term solution to achieve speed to market. The extended and core team included UX, product management, dev, research and content design across the organization including Office AI and Fluent teams.

OVERVIEW

The challenge

When a user asks for instructions on how to do something in Microsoft Copilot like how to add a watermark, an answer is generated using the large language model (LLM). This is costly and could require a wait. The business wanted to integrate help and learning content as a skill. The goal was to provide fast, accurate and version relevant instructional guidance to help enterprise and consumer users easily complete tasks (and achieve goals) using Microsoft products - improving both user experience and satisfaction.


The business objective was reduced wait time < 1 second.


Note: A skill is a tool that augments AI. It interacts with APIs from software or services to retrieve real-time info, include business data or perform computations.

My role

Led design of the MVP / short-term solution to achieve speed to market. The extended and core team included UX, product management, dev, research and content design across the organization including Office AI and Fluent teams.

OVERVIEW

The challenge

When a user asks for instructions on how to do something in Microsoft Copilot like how to add a watermark, an answer is generated using the large language model (LLM). This is costly and could require a wait. The business wanted to integrate help and learning content as a skill. The goal was to provide fast, accurate and version relevant instructional guidance to help enterprise and consumer users easily complete tasks (and achieve goals) using Microsoft products - improving both user experience and satisfaction.


The business objective was reduced wait time < 1 second.


Note: A skill is a tool that augments AI. It interacts with APIs from software or services to retrieve real-time info, include business data or perform computations.

My role

Led design of the MVP / short-term solution to achieve speed to market. The extended and core team included UX, product management, dev, research and content design across the organization including Office AI and Fluent teams.

Minimum viable product scope

For the initial release, the team was focused on two scenarios:


  • As a user, I can view Quick Answers (including actions) in Copilot when I need help with a task. 


  • As a user, if a Quick Answer doesn’t help, I can quickly access the help pane from Copilot for additional Help resources

Copilot Help Skill Prototype

Copilot Help Skill Prototype

In-product help experience

In-product help experience

In-product Help generates an efficient and accurate response to a query in the form of a “quick answer” or a Help article without a call to the LLM.

In-product Help generates an efficient and accurate response to a query in the form of a “quick answer” or a Help article without a call to the LLM.

What is a quick answer?

What is a quick answer?

  • Concise step-by-step answers to common queries (eg. how do I add a shape).


  • Action buttons help users complete the task (eg. Show me).

  • Concise step-by-step answers to common queries (eg. how do I add a shape).


  • Action buttons help users complete the task (eg. Show me).

Comparative audit

Comparative audit

We did an audit to learn from internal and external apps. Key themes included: Use suggested prompts and actions to help users, build credibility with reference links, use an informal tone (eg. Emojis, interjections), balance concise vs. lengthy responses.

We did an audit to learn from internal and external apps. Key themes included: Use suggested prompts and actions to help users, build credibility with reference links, use an informal tone (eg. Emojis, interjections), balance concise vs. lengthy responses.

Bing Copilot

Canva

Einstein

Bing Copilot

Canva

Salesforce Einstein

Aligned on user and system flow

Aligned on user and system flow

We worked with product managers and developers to review, define and align on the optimal flow for the skill.

We worked with product managers and developers to review, define and align on the optimal flow for the skill.

What will success look like?

What will success look like?

Based on the architecture approach and user flows - we explored and defined engagement measures:


  • Seen: number of times users were shown suggestion for Help skill

  • Try: number of times users request the Help skill in Copilot

  • Kept: number of times a user engages with a Help skill response 

  • Number of times Quick Answer was shown as a response 

  • Number of times action button was clicked 


Pre-launch the team aligned on org wide and skill specific engagement measures:


  • Seen: number of times users were shown suggestion for Help skill

  • Try: number of times users request the Help skill in Copilot

  • Kept: number of times a user engages with a recommended skill response 

  • Number of times Quick Answer was shown as a response 

  • Number of times QA action button was clicked 

Establishing coherence

Establishing coherence

We worked with visual and content design teams to review and align on emerging patterns for card layout and conversation design of skills within Microsoft Copilot.

We worked with visual and content design teams to review and align on emerging patterns for card layout and conversation design of skills within Microsoft Copilot.

Ideation

Ideation

Based on the elements of this emerging design system, we explored treatments to the help and learning skill.

Based on the elements of this emerging design system, we explored treatments to the help and learning skill.

Finalizing designs for dev handoff

Finalizing designs for dev handoff

After extensive collaboration, we aligned on the final treatment and documented the final specs with light and dark mode treatments, anatomy and accessibility guidance for every potential card type.

After extensive collaboration, we aligned on the final treatment and documented the final specs with light and dark mode treatments, anatomy and accessibility guidance for every potential card type.

Lessons learned

Lessons learned

Our team had to make fast short-term decisions while exploring future possibilities. Here is a bit of what I learned:


  • Building the plugin required almost daily collaboration to align with Fluent AI’s evolving design system and to justify intentional product-level differences.


  • We evaluated two technical paths 1) grounding Copilot in Help data or 2) creating a separate Help Skill. This required weighing tradeoffs in performance, user experience, and maintenance.


  • Although the initial MVP focused on simple, single-turn commands, we used design thinking to explore future scenarios for onboarding, personalization and multi-turn dialogues for complex workflows.


  • We encountered a need to educate users in research data early on. This meant that the transition to a truly AI-first experience would mean intentionally shaping new user behaviors beyond GUI interactions (e.g., ribbon commands).

Pre-launch the team aligned on org wide and skill specific engagement measures:


  • Seen: number of times users were shown suggestion for Help skill

  • Try: number of times users request the Help skill in Copilot

  • Kept: number of times a user engages with a recommended skill response 

  • Number of times Quick Answer was shown as a response 

  • Number of times QA action button was clicked 

Minimum viable product scope

For the initial release, the team was focused on two scenarios:


  • As a user, I can view Quick Answers (including actions) in Copilot when I need help with a task. 


  • As a user, if a Quick Answer doesn’t help, I can quickly access the help pane from Copilot for additional Help resources