Explainable AI for B2B SaaS
Vunet Systems
In this fast-paced mid-sized AI startup, I was a solo designer. I designed AI interfaces focused on transparency and explainablity. I worked closely with the CEO, the ML Team to understand algorithm metrics and the Development team to understand what is possible to build.
Solution
The product is used by major Indian banks, for end to end monitoring of business journeys. We wanted to introduce the insights generated by our AI algorithms into the user workflow. We wanted to explore different ways that this ‘AI - human’ interface could be designed.
Insight Cards on Dashboard
Proactive and predictive insights that explain what could go wrong, and what went wrong. Confidence score, correlated events and historical data charts make it explainable.
ML Management Dashboard
Managing ML models for users wishing to learn more about or manage the algorithms. Having 3 levels of abstraction make it explainable to people with different technical skills.
Discovery
User Research
Our initial goal was to find a suitable form for the insights that are delivered by the AI algorithm. However, our research revealed an important opportunity that made us revisit the problem - trust. Our users had a mixed perception of AI. Too much trust and too little trust, both are a result of not understanding how algorithms make decisions for a job that our end user is accountable for.
Monitoring happens manually, a process called ‘eyeball monitoring’
When failures happen, the recovery is also a very human and manual process
When we introduce AI into a workflow that is mostly human, trust becomes paramount.
Contextual Inquiries
Interviews with various stakeholders - CXOs, Development team, Sales team
5 semi-structured interviews & cognitive walkthroughs with our (proxy) users
Building user flows, touch points, time and clicks per task
Opportunity
Trust, ethics and explainability are important factors to consider when implementing AI in the real world. Our research showed us an opportunity to make AI explainable to all user types - with low, medium and high AI expertise.
Significance
AI can attract a lot of new customers, but having a trustworthy product makes them stay. A human centred approach to AI design helps Vunet stand out amongst competitors. We soften the stance of perfection placed on AI algorithms by showing users how decisions and predictions are made.
Ideation, Design & Prototyping
Insight Cards
Iteration 1
Insights in natural language
Machine generated insights is quantitative data, whereas our minds make sense of this data qualitatively. We want an insight card to bridge this gap, and reduce the interpretation time required.
Pros
Easier to read
Very effective if accurate
Cons
Subjective and open to misinterpretation
Lack of contextual data
Iteration 2
Insights under a panel
We were inspired by Google Analytics’ AI insights panel. Before it could even be tested, it was shot down by development due to framework scalability and feasibility issues.
Pros
Maintains context
Can be shown or collapsed
Technically feasible
Cons
Not scalable
Non-conventional
Iteration 3
Insights as a card component
Algorithm outputs -> visual information
A single glimpse of this card can help the users prepare for adverse events, or analyse events that have occurred. It helps answer the following questions:
What is the insight and when did it happen/will occur?
Why did the algorithm predict or analyze this insight?
How confident is the algorithm?
What should the user do next?
Perfecting the Insight Card
A single glimpse of this card can help the users prepare for adverse events, or analyse events that have occurred. It helps answer the following questions:
What is the insight and when did it happen/will occur?
Why did the algorithm predict or analyze this insight?
How confident is the algorithm?
What should the user do next?
Ideation, Design & Prototyping
ML Management Dashboard
Why do we need this?
An insight card is great for users who frequently monitor the dashboard and have a low ML expertise. Our client companies often have Data Scientists and ML Engineers, who would likely want a more granular view of the algorithms. Based on these varying needs, we designed an ML management dashboard which has 3 layers.
Layer 1 - Overall Health
These components can be used by low ML users to escalate issues if the system health is down
Layer 2 - Model Performance
Used by mid level ML users to understand performance, training and input data quality. They can further drill down to know more by clicking the ‘More’ button.
Layer 3 - Individual Model Details
Highest level of granularity. Can be used by high level ML users to redesign, analyse or fix algorithms.
Takeaways
This project made me a research first designer
A strong foundation of user research helps make ambiguous problems less challenging. AI is changing the world and how we work at light speed, and design plays an important role in bringing that power to people. Trust, transparency and explainability are crucial concepts in designing for AI
Designing at startups is like shapeshifting in and out of different teams
I learned early on that each team speaks a different language. Especially in enterprise teams, the voice of the user can be easily forgotten.
Presenting my work regularly helped me soothe my defensiveness
I'll admit that often, I unintentionally get attached to an idea or concept. Presenting and sharing my work opened a safe space for ideas to flow through, failing early on bad ideas and building up quickly on the good ones.