Trusted Data for AI: Robust, Relevant, Responsible (Informatica)

COLLIDE Data + AI Conference 2024Data Management for AI: What Does It Take to Build AI Success on Trusted Data?Watch the full session · 19 min

At COLLIDE 2024 in Atlanta, Peter Manta, AI and Data Strategy Leader at Informatica, told attendees that most AI conversations skip the part that decides whether projects succeed: the data. In a session titled "Data Management for AI: What Does It Take to Build AI Success on Trusted Data?", sponsored by Informatica, he used a mock vacation, a Zillow cautionary tale and a three-part checklist to make the case. COLLIDE ran October 10-11, 2024, at the Sandy Springs Performing Arts Center.

Why should data teams be involved in AI projects?

Manta opened with a show of hands. Nearly everyone in the room had an AI initiative under way, fewer worked with data, and fewer still were data people involved in those initiatives. He argued they should be.

He cited an Informatica survey of chief data officers who were trying to implement AI use cases. The roadblocks they named, he said, all involve data: quality, privacy, ethics, quantity and governance. His point was that the data office can help with exactly the issues that stop AI goals.

What are robust, relevant and responsible data?

Manta illustrated the idea with an exercise in which an AI model picks a group vacation. In the first scenario, nobody fed it commercial shipping logs, so the group lands on a beach covered in oil from a grounded tanker. In the second, a data error makes a lodge look like $50 a night when it is far more, and the bill balloons. In the third, the model sends everyone to Rome without accounting for a gluten allergy affecting half the group.

Each failure mapped to a pillar:

• Robust: Look at all the data needed, including mainframes, streaming sources, applications, JSON documents and unstructured data.
• Relevant: The data has to be correct, and "two decimal points off" can change the outcome.
• Responsible: Treat data with an eye to ethics, bias, transparency, accountability and lineage.

What happened when Zillow got its models wrong?

Manta said real consequences follow when this goes wrong. His favorite example was Zillow. Its Zestimate home valuations, he said, had become quite accurate over the years. Zillow then launched Zillow Offers, which bought homes it thought were undervalued and resold them.

According to Manta, the company used a model that was too brittle to adapt to changing market conditions and trained it on different data than the valuation product. When the market changed, he said, the business lost $300 million. He also noted lighter examples, such as chatbots that swear and a truck sold for a dollar.

Where does the effort in AI projects really go?

Manta walked through the data science lifecycle: data work on the left, modeling in the middle, and operationalization on the right. Teams are tempted to dive into modeling, he said, because it is the exciting part. But the data exploration, governance, cleaning and feature engineering come first. As an example, he said a column of phone numbers is not useful until you strip out the area code and can use it geographically.

He cited an 80/20 rule: 80% of the work is not modeling. "The modeling is becoming a commodity," he said, pointing to cloud services and Python libraries. "The focus is on the middle but it should really be on the book ends."

On the right side, he expected AI regulation to make drift monitoring, governance, transparency and accountability much more important, so that when AI "goes off the rails," companies can explain what happened.

How do context, governance and regulation apply to generative AI?

For generative AI, Manta stressed context. The same words can mean different things in a bar and at the bottom of a staircase, and a question about the color blue should get different answers from a paint seller's chatbot and a general one. Models have to be grounded in the right data.

Governance and security matter too, he said. An HR employee asking an internal chatbot for North American salaries should see them; someone in manufacturing should not.

On regulation, he pointed to the EU AI Act's focus on high-risk systems, and said states such as Colorado and California are moving ahead in the U.S. while Brazil, India, Canada and Japan are looking at rules. He said the safe prediction is "more lineage more transparency more accountability," supported by data quality, lineage and cataloging.

What next steps did Manta recommend?

Manta recommended starting with a maturity assessment of data management practices, including governance, cataloging and standardized data quality. Teams should then bring the chief data officer and data office into AI projects and pilot with a small, high-impact use case. Gather feedback on the data, iterate and scale, he said. Some organizations start internally and later open chatbots to external customers.

Where can data and AI leaders continue the conversation?

COLLIDE 2027, the largest AI conference in the Southern U.S., will bring 1,000+ enterprise data and AI leaders to the Sandy Springs Performing Arts Center in Atlanta on Wednesday, October 13, 2027. Details are on the COLLIDE event page, and registration is coming soon.

Frequently asked questions

What is trusted data for AI?

Manta described it as robust, relevant and responsible: complete enough, correct, and handled with attention to ethics, bias and lineage.

Why do AI projects fail because of data?

He said CDOs cite data quality, privacy, ethics, quantity and governance as the main roadblocks.

What went wrong with Zillow Offers?

Manta said Zillow used a model too brittle for changing market conditions, trained on different data than its valuation tool, and lost $300 million.

How should companies start an AI initiative?

He recommended a data-focused maturity assessment, involving the data office, and piloting a small high-impact project.

This article is based on remarks by Peter Manta, AI and Data Strategy Leader, Informatica at COLLIDE Data + AI Conference 2024 on October 10, 2024. More coverage: COLLIDE AI Conference.

Filed under: Data & AI Infrastructure, Observability & Data Quality, AI Governance & Trust