AI Design in High Risk Settings - Aligning business impact, risks, and innovation
As AI becomes integral to business strategy, many organizations have struggled to navigate the delicate balancing act between technical complexity and business value, especially with the advent of generative AI. This is especially true in high-risk settings like healthcare, defense, and financial services.
In this session, actionable strategies for enhancing AI and data literacy among stakeholders, aligning AI projects with business objectives, and fostering trust between AI builders and users will be explored. While many strategies are applicable to both traditional machine learning and generative AI projects, special considerations for projects incorporating generative AI will also be explored.
Attendees will walk away with practical insights to propel AI projects from the design board to real-world impact and pave the way for a culture of informed and responsible AI innovation within their organizations.
Session Outline:
Participants will learn how to:
· Develop greater AI and data literacy among stakeholders and users
· Align between business outcomes and AI project goals
· Build trust and effective communication channels between AI builders and users
· Identify and communicate the potential risks and unintended consequences that arise from both traditional and generative ML models
Background Knowledge:
Suitable for all backgrounds - especially those looking to communicate more effectively with stakeholders, or stakeholders learning to get more comfortable with technical considerations.
4. Left unchecked,
models can learn
the wrong things
from data
"Why Should I Trust You?": Explaining the
Predictions of Any Classifier
Marco Tulio Ribeiro, Sameer Singh, Carlos
Guestrin
5. Left unchecked, deep
learning models can
learn the wrong
things from data
Data Science Institute, American
College of Radiology (2022)
8. Fewer than 25% of
the workforce
perceive self as
data literate (HBR 2021)
Digital literacy
• Word processing
• Spreadsheets
2000s
Data literacy
• What data?
• Can we build?
2010s
AI literacy
• Should we build?
• When do we trust?
2020s
9. Enabling users saved the day
• 50 million+ patient lives
• The team reaches out to qualified (accepted) leads
• Missing potential accepts reduces potential revenue
• Accepting unqualified leads reduces profit
Ingest Outreach Qualify
Approve
Accept
Increased monthly
revenues by 18% (to date)
Outcome
11. Annual Solid-Tumor Cancer Readmissions
Hospitalizations
Readmissions (22k per
readmission)
Identified by model
(potential savings)
Prevented by intervention
(~$2.3M saved)
2.8 million cancer in-patient
visits in US
15 – 25%
(national avg)
~70%
~20%
• Resource
efficiency
• Patient
outcomes
• Care quality
• Care
coordination
and continuity
12. The value of data
science is limited by
the value of decisions
impacted