MULTI-AGENT AI FOR ETHICAL DECISION-MAKING
Aura
MULTI-AGENT AI FOR ETHICAL DECISION-MAKING
Aura
MULTI-AGENT AI FOR ETHICAL DECISION-MAKING
Aura
I worked on a multidisciplinary team (research, design, and dev) to design for behavior change and develop a multi-agent AI decision support system that simplifies ethical decision-making for leaders in manufacturing and robotics.
DURATION
6 months
ROLE
Design lead
CLIENT
Honda Research Institute
SKILLS
UI/UX, user research, usability testing
TOOLS
Figma, Dovetail
I worked on a multidisciplinary team (research, design, and dev) to design for behavior change and develop a multi-agent AI decision support system that simplifies ethical decision-making for leaders in manufacturing and robotics.
DURATION
6 months
ROLE
Design lead
CLIENT
Honda Research Institute
SKILLS
UI/UX, user research, usability testing
TOOLS
Figma, Dovetail
I worked on a multidisciplinary team (research, design, and dev) to design for behavior change and develop a multi-agent AI decision support system that simplifies ethical decision-making for leaders in manufacturing and robotics.
DURATION
6 months
ROLE
Design lead
CLIENT
Honda Research Institute
SKILLS
UI/UX, user research, usability testing
TOOLS
Figma, Dovetail
INITIAL BRIEF
Use AI to demonstrate how non-technical end users may include their personal ethics in systems that transcend transportation, 10 to 15 years in the future.
Use AI to demonstrate how non-technical end users may include their personal ethics in systems that transcend transportation, 10 to 15 years in the future.
With a prompt this ambiguous, it was important to narrow our focus strategically.
With a prompt this ambiguous, it was important to narrow our focus strategically.
NARROWING SCOPE
Finding potential to make the greatest impact
40+ interviews with executives across manufacturing and robotics, healthcare, and emergency response; an extensive lit review; and decision-making surveys with 200 respondents, revealed that our greatest impact lies in organizational high-stakes decision-making.
NARROWING SCOPE
Finding potential to make the greatest impact
40+ interviews with executives across manufacturing and robotics, healthcare, and emergency response; an extensive lit review; and decision-making surveys with 200 respondents, revealed that our greatest impact lies in organizational high-stakes decision-making.
NARROWING SCOPE
Finding potential to make the greatest impact
40+ interviews with executives across manufacturing and robotics, healthcare, and emergency response; an extensive lit review; and decision-making surveys with 200 respondents, revealed that our greatest impact lies in organizational high-stakes decision-making.
SETTING
Why decision making?
Why decision making?
The decisions we make influence every action we take and are heavily rooted in ethics; this has the greatest ripple effect.
The decisions we make influence every action we take and are heavily rooted in ethics; this has the greatest ripple effect.
DOMAIN
Why manufacturing & robotics?
Why manufacturing & robotics?
The transition to Industry 5.0 makes ethical AI collaboration urgent and relevant. And, their familiarity with AI means lower adoption barriers.
The transition to Industry 5.0 makes ethical AI collaboration urgent and relevant. And, their familiarity with AI means lower adoption barriers.
CONTEXT
Why organizational high-stakes?
Why organizational high-stakes?
Organizational decision-makers are responsible for more than just themselves, and high-stakes decisions are more emotionally difficult. So, these decisions are riskier, more ethically complex, and impact more people.
Organizational decision-makers are responsible for more than just themselves, and high-stakes decisions are more emotionally difficult. So, these decisions are riskier, more ethically complex, and impact more people.
TARGET USER
Why executives?
Why executives?
They are responsible for strategic, often values-based decisions but lack access to stakeholder perspectives. Illuminating potential consequences and alternatives can enhance ethical reflection and lead to better decision outcomes.
They are responsible for strategic, often values-based decisions but lack access to stakeholder perspectives. Illuminating potential consequences and alternatives can enhance ethical reflection and lead to better decision outcomes.
CONTEXT
Current decision-making processes present many challenges
Gathering findings from interviews, I found that manufacturing leaders don't need more data. They just need access to it. As leaders have to make complex decisions with limited information, critical perspectives often get overlooked.
CONTEXT
Current decision-making processes present many challenges
Gathering findings from interviews, I found that manufacturing leaders don't need more data. They just need access to it. As leaders have to make complex decisions with limited information, critical perspectives often get overlooked.
CONTEXT
Current decision-making processes present many challenges
Gathering findings from interviews, I found that manufacturing leaders don't need more data. They just need access to it. As leaders have to make complex decisions with limited information, critical perspectives often get overlooked.
Fragmented information systems
Fragmented information systems
Information is scattered across departments, people, and systems; there does not exist a place to aggregate it.
Information is scattered across departments, people, and systems; there does not exist a place to aggregate it.
Lost institutional knowledge
Lost institutional knowledge
Decisions rely heavily on historical knowledge, but that knowledge gets lost due to the generational gap of experienced workers retiring.
Decisions rely heavily on historical knowledge, but that knowledge gets lost due to the generational gap of experienced workers retiring.
Hard to predict outcomes
Hard to predict outcomes
Implementation feasibility is hard to forecast, but failures could be prevented with input from floor workers.
Implementation feasibility is hard to forecast, but failures could be prevented with input from floor workers.
Current AI systems fail to consider contextual factors
AI systems like COMPAS—a risk assessment tool used by judges to decide prison sentences—treat data as objective, jumping to conclusions without explaining their internal logic.
Current AI systems fail to consider contextual factors
AI systems like COMPAS—a risk assessment tool used by judges to decide prison sentences—treat data as objective, jumping to conclusions without explaining their internal logic.
Current AI systems fail to consider contextual factors
AI systems like COMPAS—a risk assessment tool used by judges to decide prison sentences—treat data as objective, jumping to conclusions without explaining their internal logic.
REFRAMING
How might we support manufacturing and robotics leaders in making informed decisions, where critical perspectives like floor workers are often unconsidered?
How might we support manufacturing and robotics leaders in making informed decisions, where critical perspectives like floor workers are often unconsidered?
SOLUTION
A multi-agent AI decision support system to expand perspective
A multi-agent AI decision support system to expand perspective
By surfacing unintended consequences, stakeholder perspectives, and potential alternatives, Aura helps leaders see the larger context of their decision. It provides clear and actionable information that aligns with user needs.
By surfacing unintended consequences, stakeholder perspectives, and potential alternatives, Aura helps leaders see the larger context of their decision. It provides clear and actionable information that aligns with user needs.
ADDRESSING PAIN POINTS
Fragmented information systems?
Aura pulls it all together.
Fragmented information systems?
Aura pulls it all together.
Aura and its sub-agents bring together information scattered across people and departments and puts it in context, highlighting how various stakeholders will be impacted by the decision.
Aura and its sub-agents bring together information scattered across people and departments and puts it in context, highlighting how various stakeholders will be impacted by the decision.
Lost institutional knowledge?
Aura captures it.
Lost institutional knowledge?
Aura captures it.
Using a RAG pipeline, Aura preserves historical context, retrieving the most relevant information at the right time.
Using a RAG pipeline, Aura preserves historical context, retrieving the most relevant information at the right time.
Hard to predict outcomes?
Aura helps plan mitigation.
Hard to predict outcomes?
Aura helps plan mitigation.
By surfacing unintended consequences, executives gain insight into the ripple effects of their decision before making it. Aura also provides alternatives and trade-offs to help them plan ahead and prevent harm.
By surfacing unintended consequences, executives gain insight into the ripple effects of their decision before making it. Aura also provides alternatives and trade-offs to help them plan ahead and prevent harm.
DESIGN CONSIDERATIONS
Creating a new form of human-AI collaboration that supports reflection
I guided the team through 5 rounds of concept and usability testing, where we received feedback from over 170 users. My goal was to understand how best to present information to users to 1) increase decision confidence and 2) encourage ethical reflection. To do so, I considered the following:
DESIGN CONSIDERATIONS
Creating a new form of human-AI collaboration that supports reflection
I guided the team through 5 rounds of concept and usability testing, where we received feedback from over 170 users. My goal was to understand how best to present information to users to 1) increase decision confidence and 2) encourage ethical reflection. To do so, I considered the following:
DESIGN CONSIDERATIONS
Creating a new form of human-AI collaboration that supports reflection
I guided the team through 5 rounds of concept and usability testing, where we received feedback from over 170 users. My goal was to understand how best to present information to users to 1) increase decision confidence and 2) encourage ethical reflection. To do so, I considered the following:
How to visualize the decision environment
How to visualize the decision environment
How helpful is it for users to see the complexity of their decision? How does this affect cognitive load?
How helpful is it for users to see the complexity of their decision? How does this affect cognitive load?
What information to present, and when
What information to present, and when
What kind of information best helps users navigate their decision? When should additional context be surfaced?
What kind of information best helps users navigate their decision? When should additional context be surfaced?
How to preserve human agency
How to preserve human agency
How can we ensure users maintain autonomy? In what ways should the system communicate with users?
How can we ensure users maintain autonomy? In what ways should the system communicate with users?
ITERATIONS
Mapping the decision environment visually and digestibly
Mapping the decision environment visually and digestibly
I designed and tested 3 variations of a visualization. While users valued seeing the seeing the decision environment, they found the web of nodes confusing. So, I simplified the design to improve clarity while still conveying the interconnectedness of factors.
I designed and tested 3 variations of a visualization. While users valued seeing the seeing the decision environment, they found the web of nodes confusing. So, I simplified the design to improve clarity while still conveying the interconnectedness of factors.
Overwhelming and confusing
Overwhelming and confusing
Overwhelming and confusing
Lacks context
Lacks context
Lacks context
Improved clarity
Improved clarity
Improved clarity
Less distraction
Less distraction
Less distraction
Navigating the decision process with progressive onboarding
Navigating the decision process with progressive onboarding
I learned that users need more support throughout the process than they realize. I introduced progressive onboarding to give users context as they go, instead of bombarding them with this upfront.
I learned that users need more support throughout the process than they realize. I introduced progressive onboarding to give users context as they go, instead of bombarding them with this upfront.
Information overload upfront
Information overload upfront
Information overload upfront
Skipped or forgotten by users
Skipped or forgotten by users
Skipped or forgotten by users
Provides context as users go
Provides context as users go
Provides context as users go
Ever-present
Ever-present
Ever-present
Reducing interaction points with a focused info panel
Reducing interaction points with a focused info panel
This change surprised me. I thought a conversational interface would feel intuitive, but too many interaction points meant users weren't sure how to engage. I transformed it into a one-directional panel where sub-agents provide timely, relevant information.
This change surprised me. I thought a conversational interface would feel intuitive, but too many interaction points meant users weren't sure how to engage. I transformed it into a one-directional panel where sub-agents provide timely, relevant information.
LOOKING AHEAD
What might the future of human-AI collaboration look like?
What might the future of human-AI collaboration look like?
Ethical decisions are deeply complex, and the binary input of buttons limits nuance. To account for this, the future state of Aura is multi-modal: both voice and touch. Two-way dialogue allows users to more intuitively communicate their values and explore the intricacies of their decision.
Ethical decisions are deeply complex, and the binary input of buttons limits nuance. To account for this, the future state of Aura is multi-modal: both voice and touch. Two-way dialogue allows users to more intuitively communicate their values and explore the intricacies of their decision.
OUTCOMES
Increasing decision confidence
Increasing decision confidence
During usability testing, we measured target users' confidence levels before and after using Aura.
During usability testing, we measured target users' confidence levels before and after using Aura.
increase in decision confidence after using Aura in the decision process
increase in decision confidence after using Aura in the decision process
increase in decision confidence after using Aura in the decision process
7.5%
7.5%
Executives in manufacturing and robotics found Aura valuable:
Executives in manufacturing and robotics found Aura valuable:
"It really helped me understand the downstream effects on various areas of the business. I hadn't thought how departments like marketing or engineering might be affected."
"It really helped me understand the downstream effects on various areas of the business. I hadn't thought how departments like marketing or engineering might be affected."
"Aura would help me understand blind spots in my thought process and better understand the overall problem statement before I made a decision."
"Aura would help me understand blind spots in my thought process and better understand the overall problem statement before I made a decision."
Takeaways
Ambiguity requires action; put stakes in the ground to move forward.
This was my first time leading design on an interdisciplinary team, and it challenged my instinct to research until certain and test only after defining the problem. In a space this ambiguous, I had to make and test assumptions early—creating clarity through action rather than waiting for it.
Takeaways
Ambiguity requires action; put stakes in the ground to move forward.
This was my first time leading design on an interdisciplinary team, and it challenged my instinct to research until certain and test only after defining the problem. In a space this ambiguous, I had to make and test assumptions early—creating clarity through action rather than waiting for it.
Takeaways
Ambiguity requires action; put stakes in the ground to move forward.
This was my first time leading design on an interdisciplinary team, and it challenged my instinct to research until certain and test only after defining the problem. In a space this ambiguous, I had to make and test assumptions early—creating clarity through action rather than waiting for it.
