AI Agent Workflow
AI Agent Workflow
AI Agent Workflow
A three-stage "explainable recommendations + human-in-the-loop" process that helps local businesses confidently adopt AI-suggested actions.
A three-stage "explainable recommendations + human-in-the-loop" process that helps local businesses confidently adopt AI-suggested actions.
View full deck
Role
Role
Role
Product/UX Design
Product/UX Design
Product/UX Design
Year completed
Year completed
Year completed
2025
2025
2025
My contributions
My contributions
My contributions
Strategy to Execution
Quality and Constraints
Engineering Collaboration
Strategy to Execution
Quality and Constraints
Engineering Collaboration
Strategy to Execution
Quality and Constraints
Engineering Collaboration
Background
Background
Background
AtlasNova is a B2B AI operations partner for small and medium-sized businesses, currently focused on the food and beverage industry. It features built-in AI CMO/COO/CDO roles, integrating marketing, operations, analytics, and multi-platform capabilities.
Project Overview
Project Overview
Project Overview
Target Users
Target Users
Target Users
Local business owners and store managers with 1–20 locations ("wearing all hats—operations/marketing/finance")
Local business owners and store managers with 1–20 locations ("wearing all hats—operations/marketing/finance")
Local business owners and store managers with 1–20 locations ("wearing all hats—operations/marketing/finance")
Mike (3 stores Coffee Shop Owner)
Mike (3 stores Coffee Shop Owner)
Mike (3 stores Coffee Shop Owner)
User Persona
Pain
Pain
Pain
Data scattered across 8+ tools
Data scattered across 8+ tools
Data scattered across 8+ tools
Marketing/review management are manual
Marketing/review management are manual
Marketing/review management are manual
Lengthy chain from insights to execution
Lengthy chain from insights to execution
Lengthy chain from insights to execution
Impact
Impact
Impact
High labor costs
High labor costs
High labor costs
Missed growth opportunities
Missed growth opportunities
Missed growth opportunities
Decision fatigue
Decision fatigue
Decision fatigue
Core Problem
Core Problem
Core Problem
18%
18%
18%
V1 experience was fragmented with unclear entry points.
Resulting in only 18% of users taking action after seeing AI recommendations (baseline).
V1 experience was fragmented with unclear entry points.
Resulting in only 18% of users taking action after seeing AI recommendations (baseline).
V1 experience was fragmented with unclear entry points.
Resulting in only 18% of users taking action after seeing AI recommendations (baseline).
Primary Goal
Primary Goal
Primary Goal
Improve user understanding and trust in AI recommendations, enabling more users to confidently execute next steps (rather than just reading).
Improve user understanding and trust in AI recommendations, enabling more users to confidently execute next steps (rather than just reading).
Improve user understanding and trust in AI recommendations, enabling more users to confidently execute next steps (rather than just reading).
Design Strategies
Design Strategies
Design Strategies
Trust is the foundation of efficiency, "Less doubt, not just fewer clicks".
Trust is the foundation of efficiency, "Less doubt, not just fewer clicks".
Trust is the foundation of efficiency, "Less doubt, not just fewer clicks".
AI Agent Persona
AI Agent Persona
AI Agent Persona
Calm, Competent Chief of Staff.
Calm, Competent Chief of Staff.
Calm, Competent Chief of Staff.
Agent Persona
Calm
Communicates with calm and confidence, never rushing or overwhelming.
Communicates with calm and confidence, never rushing or overwhelming.
Communicates with calm and confidence, never rushing or overwhelming.
Competent
Offers clear, data-backed reasoning for each suggestion.
Offers clear, data-backed reasoning for each suggestion.
Offers clear, data-backed reasoning for each suggestion.
Respectful
Always leaves the final decision to the human.
Always leaves the final decision to the human.
Always leaves the final decision to the human.
Design Levers & Metrics
Design Levers & Metrics
Design Levers & Metrics
Outcome
Outcome
Outcome
Design Lever
Tracked Metrics
Increase recommendation adoption rate
Increase recommendation adoption rate
Increase recommendation adoption rate
Explainability
Suggestion expansion rate; percentage of recommendations with explanations; adoption rate after reading explanations
Increase execution completion rate
Increase execution completion rate
Increase execution completion rate
Human-in-the-loop
Prepare page completion rate; confirmation page abandonment rate; post-edit launch rate
Reduce interruption perception
Reduce interruption perception
Reduce interruption perception
Low noise/appropriate tone
Weekly recommendation volume; single-page display limit; rejection reason distribution; accept/reject ratio
Solutions Overview
Solutions Overview
Solutions Overview
Information Architecture Redesign
- User Research
With 5 users with semi-structured interviews, we find there are two styles of users. And we are thinking maybe we needed two rhythms:
With 5 users with semi-structured interviews, we find there are two styles of users. And we are thinking maybe we needed two rhythms:
With 5 users with semi-structured interviews, we find there are two styles of users. And we are thinking maybe we needed two rhythms:
Act Now（Push）
Firefighters
AI suggests next steps proactively.
“Just tell me what’s wrong.”
Expect quick answers.
Act fast, trust results.
Explore & Ask（Pull）
Thinkers
User asks questions or explores data.
“I want to check it myself”
Need reasons, not orders.
Act carefully, trust logic.
- Vibe Code Prototype Demo
Using Figma Make to quick explore and develop. Named two rhythms entrance: Command Center and Analytics Analytics Dashboard.
Using Figma Make to quick explore and develop. Named two rhythms entrance: Command Center and Analytics Analytics Dashboard.
Using Figma Make to quick explore and develop. Named two rhythms entrance: Command Center and Analytics Analytics Dashboard.
- Simplified the idea
Because the project requires rapid iteration, we discussed splitting the old dashboard into two pages first.
Because the project requires rapid iteration, we discussed splitting the old dashboard into two pages first.
Because the project requires rapid iteration, we discussed splitting the old dashboard into two pages first.
Three-stage Human-in-the-loop Process
Respectful
AI do 80–90% of the work. Humans review and make the final call.
AI Agent
Users
1
Suggest
AI proposes
User approves or skips
2
Prepare
AI drafts detail plan
User tweaks
3
Confirm
AI generates
User signs off
1
Suggest Stage
Displays actionable next steps with essential elements such as goals, metrics, and platforms, along with "why" (predicted impact & confidence level). CTA: Act now.
Displays actionable next steps with essential elements such as goals, metrics, and platforms, along with "why" (predicted impact & confidence level). CTA: Act now.
Displays actionable next steps with essential elements such as goals, metrics, and platforms, along with "why" (predicted impact & confidence level). CTA: Act now.
Act now
2
Prepare Stage
Form information is pre-filled by AI; users can edit and refine details (such as campaign goals/audience/assets, etc.).
Form information is pre-filled by AI; users can edit and refine details (such as campaign goals/audience/assets, etc.).
Form information is pre-filled by AI; users can edit and refine details (such as campaign goals/audience/assets, etc.).
Generate Content
3
Confirm Stage
AI generates drafts (copy/images/scheduling); users can rewrite/regenerate/replace images, then Launch; Pause/Reschedule options available both before and after launch.
AI generates drafts (copy/images/scheduling); users can rewrite/regenerate/replace images, then Launch; Pause/Reschedule options available both before and after launch.
AI generates drafts (copy/images/scheduling); users can rewrite/regenerate/replace images, then Launch; Pause/Reschedule options available both before and after launch.
Launch
- After launch
Pause/Reschedule options available both before and after launch.
Pause/Reschedule options available both before and after launch.
Pause/Reschedule options available both before and after launch.
Explainable Action Suggestion Card System
Competent
Each Action Suggestions cards included:
1
What you can do next
Provide basic information for the next action proposal.
Provide basic information for the next action proposal.
Provide basic information for the next action proposal.
2
Why it matters
Predicted Impact: Expected results vs baseline.
Confidence: How certain we are.
Predicted Impact: Expected results vs baseline.
Confidence: How certain we are.
Predicted Impact: Expected results vs baseline.
Confidence: How certain we are.
3
Why triggered it
Signals: Live signals that triggered this.
History: Evidence from similar past cases.
Assumptions: Conditions required to succeed.
Signals: Live signals that triggered this.
History: Evidence from similar past cases.
Assumptions: Conditions required to succeed.
Signals: Live signals that triggered this.
History: Evidence from similar past cases.
Assumptions: Conditions required to succeed.
Calm
Keep It Clear
Supports collapse/expand to reduce information overload; limit to 5–7 highest-priority recommendations per page, with others accessible via "View all."
Supports collapse/expand to reduce information overload; limit to 5–7 highest-priority recommendations per page, with others accessible via "View all."
Supports collapse/expand to reduce information overload; limit to 5–7 highest-priority recommendations per page, with others accessible via "View all."
Respectful
Enable Feedback Loop
Instant disappearance doesn't sound great, we add a new status of action suggestion card. Let usres feel their decisions are be seen and let them believe experts can do better. Also help product for backend database.
Instant disappearance doesn't sound great, we add a new status of action suggestion card. Let usres feel their decisions are be seen and let them believe experts can do better. Also help product for backend database.
Instant disappearance doesn't sound great, we add a new status of action suggestion card. Let usres feel their decisions are be seen and let them believe experts can do better. Also help product for backend database.
Dashboard Exploration Experience Enhancement
Respectful
Empower Users to Navigate Insights
Insights sidebar positioned on the right, allowing side-by-side comparison with charts. And clicking on insights highlights related charts (context motion)
Insights sidebar positioned on the right, allowing side-by-side comparison with charts. And clicking on insights highlights related charts (context motion)
Insights sidebar positioned on the right, allowing side-by-side comparison with charts. And clicking on insights highlights related charts (context motion)
Competent
Guide Users to What Matters
When users click the insights, the cards also will be open. View these dates and Factors affecting will alsoe be shown. Provide clear instructions for users when provide insights.
When users click the insights, the cards also will be open. View these dates and Factors affecting will alsoe be shown. Provide clear instructions for users when provide insights.
When users click the insights, the cards also will be open. View these dates and Factors affecting will alsoe be shown. Provide clear instructions for users when provide insights.
Respectful
Open / Collapsed Modes
Users can hide to explore data without brother. Give users room to explore. Be there when needed.
But the initial status keeps the insights Side panel open to encourage user engagement.
Users can hide to explore data without brother. Give users room to explore. Be there when needed.
But the initial status keeps the insights Side panel open to encourage user engagement.
Users can hide to explore data without brother. Give users room to explore. Be there when needed.
But the initial status keeps the insights Side panel open to encourage user engagement.
Full Deck
Full Deck
Full Deck
Case Study: Designing Trust in AI agent workflow
Case Study: Designing Trust in AI agent workflow
Case Study: Designing Trust in AI agent workflow
View more
Button Text
Button Text
