Inside The Garage SF with Cici Zhao

Two weeks ago, I had the privilege of moderating a panel discussion in San Francisco as part of the Kellogg School of Management alumni network. The event brought together AI professionals and alumni for an engaging conversation on one of the biggest challenges in AI today: Data Readiness. It was an evening of knowledge sharing, networking, and thought-provoking discussion.

As a Solutions Manager at Quantexa, I lead Go-To-Market efforts for AI-powered solutions. My background includes driving AI strategy and analytics at Square and consulting for Fortune 500 companies at Visa and EY. I’m deeply passionate about AI’s evolving role in business and love fostering discussions that bridge technical advancements with practical applications.

Our Panelists

We were fortunate to have three incredible speakers who each brought unique expertise to the conversation:

• Michelle Bonat – A transformative leader in AI and technology, Michelle has integrated AI into 90+ products for 1.5 billion logins and held leadership roles at JPMC, Chase, and Oracle. She holds multiple AI patents and serves on the Kellogg/McCormick MBAi advisory board.
• Tam Le – With over 15 years of experience leading data and analytics teams at Google, Adobe, and Asana, Tam now helps companies fine-tune AI models and improve data quality through her company, Tbrain.AI.‍
• Alexander Tsado – A passionate advocate for AI accessibility, Alex helped deploy the first Nvidia AI GPUs to major cloud providers and now focuses on global AI education and strategy through Ahura AI and Alliance4ai.org.

Our Discussion & Takeaways

We kicked off the discussion with a striking statistic: 85% of AI projects fail to deliver, with poor data quality being a key factor. The panelists shared real-world examples of how missing, biased, or inconsistent data created challenges in scaling AI solutions. Common pitfalls included inconsistent data collection, labeling errors, and lack of governance. Infrastructure decisions also play a crucial role in data readiness, influencing scalability and AI performance.

To navigate these challenges, we explored best practices for data readiness. High-quality training data, post-training validation, and thoughtful data curation such as synthetic data were highlighted as essential for building effective AI models. Aligning data collection with business objectives ensures relevance and accuracy, while governance frameworks help maintain consistency and reduce bias.

We also discussed how data challenges differ across industries. In financial services, compliance and security are paramount. Startups must prioritize scalability and flexibility in their AI data strategies, while larger enterprises focus on governance and operational efficiency. AI education and responsible data use were also emphasized, particularly in preparing younger generations for an AI-driven future.

The panel concluded by discussing what makes a dataset truly valuable—diversity, accuracy, and ethical sourcing. For professionals looking to break into AI, understanding data governance, privacy regulations, and responsible AI practices is crucial.

The event wrapped up with a fantastic Q&A session, where attendees brought insightful questions that sparked further discussion. Some other takeaways:

• Reframe data challenges. Instead of labeling data as "bad," consider it "not yet ready for AI." This mindset shift fosters problem-solving rather than frustration.
• Data preparation is 80% of every AI/ML project. The success of AI initiatives depends on high-quality data collection, curation, and validation.
• Align data, use cases, and AI technology. Mismatched data can lead to unreliable outputs. For example, using a public generative AI chatbot without fact-checking for legal research may result in hallucinated information, compromising reliability.
• Tackle industry-specific data issues. In fields like finance, handling Personally Identifiable Information (PII) securely is critical. Techniques like data masking, tokenization, and homomorphic encryption can help.
• Mitigate bias for accurate AI outcomes. Incomplete or skewed data results in non-representative AI models. Diverse and well-curated datasets are essential.

A huge thank you to Michelle Bonat, Tam Le, and Alexander Tsado for sharing their expertise and to everyone who attended and contributed to the conversation. Your enthusiasm made this event a success, and I look forward to more meaningful discussions on AI, data, and innovation!