deepak sharma

AI UX guidelines

Introduction: A shared internal framework for designing consistent, trustworthy, and reusable AI experiences across the organization.

The system covers full-page, side-panel, non-conversational, and agentic interactions. It translates real product use cases into reusable principles, patterns, components, and experience blueprints.

Goal: Help teams implement AI through a shared experience language instead of solving similar interaction problems independently.

Scope note: The guidelines establish shared UX standards while allowing product teams to adapt them to their domain, users, and technical requirements.

CASE STUDY · 2026

Creating a unified interaction system for AI experiences

Role
AI UX Researcher and Design Lead
Time line
Ongoing · 2026
Team
AI UX team; co-led with manager

As a member of the core AI UX team, I co-led the creation of an internal guideline system from the ground up. My contribution included research, concept design, component design, and defining a unified experience across AI-enabled products.

We translated departmental use cases into repeatable patterns for full-page, side-panel, non-conversational, and agentic interactions. This gave teams a shared foundation while leaving room for product-specific user and business requirements.

CASE STUDY · 2026

3

AI experience modes

5

Foundational principles

38+

Reusable components

THE CHALLENGE

AI adoption was accelerating, but the experience was becoming fragmented

Product teams across the organization were exploring AI simultaneously. Each department had different use cases, but teams were designing and implementing similar AI interactions in their own way. Without a shared experience framework, the same AI capability could look, behave, and communicate differently across products. This created a need for consistent guidance covering how AI is introduced, how users interact with it, how the system communicates its status, and how users remain in control. Our challenge was to create a unified UX foundation that established consistency without forcing every product into the same solution.

Fragmented experiences

  • Similar AI capabilities were being designed differently across departments and products.

Multiple integration models

  • Teams needed guidance for full-page, side-panel, and non-conversational AI experiences.

Consistency with flexibility

  • The system needed to provide shared standards while supporting different users, workflows, and product requirements.
Flow from department use cases, through the core AI UX team and shared guidelines and patterns, to unified product experiences

EXPERIENCE FRAMEWORK

Choosing the right AI experience for the task

Our research showed that AI should not default to a chat interface. The appropriate experience depends on the user’s task, the amount of context required, and the role AI plays within the workflow.

We organized the guidelines around three integration models, giving teams a shared starting point for designing AI experiences.

Standalone/Workspace copilot

AI as the primary workspace

Best suited to complex or sustained tasks where users need space to explore, generate, compare, refine, and manage AI-produced content.

  • Supports longer and more involved workflows
  • Provides room for input, output, history, and supporting context
  • Keeps controls and system activity visible
  • Allows users to review and refine results
Full-page copilot with a greeting, suggested prompts, and chat history
Full-page copilot: a welcome state with categorized prompt suggestions, a command input, and chat history.

Contextual copilot

AI as contextual assistance

Best suited to situations where users need AI support while remaining inside their primary product workflow.

  • Maintains the user’s current context
  • Provides assistance without taking over the workspace
  • Supports suggestions, summaries, and contextual actions
  • Allows the panel to be opened or dismissed as needed
Product dashboard with a copilot side panel open on the right
Contextual copilot: a side panel that assists alongside an existing dashboard without replacing it.

Non-conversational AI

AI as a focused product capability

Best suited to predictable actions that do not require a conversational exchange.

  • Uses direct commands, suggestions, or generated alternatives
  • Reduces unnecessary conversational steps
  • Integrates AI into familiar product controls
  • Keeps the interaction focused on the user’s immediate task
Grid of standard components, such as checkbox, table, form, and modal, each with an AI action
Non-conversational AI: AI actions built into familiar components such as tables, forms, message bars, and modals.
CriteriaFull-Page AISide-Panel AINon-Conversational AI
AI’s rolePrimary workspaceContextual assistantEmbedded capability
Use whenThe task is complex, sustained, or requires explorationUsers need assistance without leaving their current workflowAI can complete a focused, predictable action
EngagementImmersive and continuousOptional and contextualLightweight and moment-based
ContextBroad task and workspace contextCurrent page, object, or workflowSelected content, field, or immediate action
User controlReview, refine, compare, and manage outputsOpen, dismiss, accept, or apply suggestionsTrigger, preview, accept, edit, or undo
ExampleOne full-page product exampleOne side-panel exampleOne embedded-action example

PATTERN SYSTEM

From recurring use cases to reusable interaction guidance

As similar interaction needs appeared across departmental use cases, we converted the repeated solutions into reusable AI UX patterns.

Each pattern explained when it should be used, how it should behave, which states and controls were required, and how teams could adapt it without breaking the unified experience.

  1. 01

    Foundations

    Establish the shared language and visual foundation.

    • Terminology
    • Legal considerations
    • Voice and tone
    • Grid and layout

  2. 02

    Interaction behavior

    Define how users initiate, guide, and control AI.

    • Basic interactions
    • Session control
    • Bidirectional control
    • Prompt guidance

  3. 03

    System communication

    Help users understand AI responses, status, and limitations.

    • Model selection
    • Error handling
    • Feedback
    • Loading and system states

  4. 04

    Reusable implementation

    Connect guidance with repeatable product solutions.

    • Components
    • Interaction patterns
    • Full-page blueprint
    • Contextual side-panel blueprint
    • Non-conversational AI blueprint
Pattern walkthrough: each copilot pattern, such as the session header, is documented with its purpose, placement, and behavior.

AGENTIC UI

Designing understandable and controllable multi-agent experiences

Beyond individual AI interactions, I researched how users might work with systems in which multiple agents coordinate to complete a larger goal.

This introduced new UX questions around agent responsibilities, delegation, configuration, visibility, permissions, human approval, and recovery when an agent cannot complete its task.

  • How should users understand the role of each agent?
  • How are tasks delegated between an orchestrator and specialized agents?
  • What information should remain shared across agents?
  • How can users monitor progress and dependencies?
  • When should an agent request approval?
  • How can users pause, redirect, or stop agent activity?
  • How should failures and incomplete handoffs be communicated?

Module One: Multi-Agent Orchestration

Making coordination between agents visible

I explored an orchestration model in which a primary agent interprets the user’s goal, delegates work to specialized agents, and brings their outputs together. The experience needed to communicate responsibilities, handoffs, progress, and dependencies without overwhelming the user.

Orchestration concept walkthrough: how a primary agent delegates work, tracks handoffs, and brings results together.

Module Two: Agent Configuration Center

Giving teams a consistent way to configure agents

The Agent Configuration Center explored how agent setup could be organized into one unified experience. The goal was to make an agent’s purpose, behavior, capabilities, and boundaries understandable before it became active.

  • Purpose and responsibilities
  • Instructions
  • Knowledge or context
  • Models and tools
  • Permissions
  • Triggers
  • Human-approval requirements
  • Testing and publishing
Agent Configuration Center prototype: browsing agents by status and role, with a copilot that helps set one up.

Module Three: Activity Trace

Helping users understand what agents are doing

Activity Trace focused on making agent behavior inspectable. Users needed a way to follow progress, understand delegation, identify failures, and review important decisions or outputs.

  • Active agent
  • Current task
  • Task status
  • Agent-to-agent handoff
  • Inputs and outputs
  • Approval requests
  • Errors and recovery
  • Final result
Activity Trace prototype: the copilot shows each step it takes, such as searching documentation, while it builds the result.

ADOPTION

Making the guidelines accessible, reusable, and maintainable

Creating the patterns was only part of the challenge. Product teams also needed a clear way to discover the guidance, understand when to use it, and access the corresponding components and blueprints.

We published the guidelines internally through a shared Figma library. The library brings together foundational guidance, interaction patterns, reusable components, and experience blueprints in one structured resource.

  • Getting started
  • Terminology
  • Structured pattern navigation
  • Experience-model guidance
  • Pattern documentation
  • Reusable components
  • Full-page AI blueprint
  • Contextual side-panel blueprint
  • Non-conversational AI blueprint
  • Change log
  • New departmental use cases
  • Pattern refinements
  • Agentic UI extensions

IF YOU GOT THIS FAR…

Thank you