At COLLIDE 2025, leaders from Quest Software, ADP, WellStar Health System and Carter's discussed governance that keeps up with AI. They covered why AI breaks traditional governance, trust scoring for data, governance by design, fast risk assessments and the team structures that let governance scale without slowing delivery.
Key takeaways
- AI adds inference. Governance must handle bad data arriving at run time, not just clean training data.
- Governance by design. ADP requires governance approval at proof of concept, pilot and general availability.
- Governance isn't only IT's job. Yateski recommends executive policy groups where IT is no more than about 15% to 20%.
- Assess risk fast. Agree on a risk profile with the business upfront, and pause pilots that exceed it.
- Don't wait for perfect data. Start AI projects with the data you have while governance keeps improving it.
Data governance has to change shape to keep up with enterprise AI, according to a panel at COLLIDE Data + AI Conference 2025 in Atlanta, in a session sponsored by Quest Software. Susan Laine, who described herself as CTO of Quest's data product line, joined Nick Yateski, former chief AI officer of WellStar Health System, and Julia, who leads the effort to scale generative AI at ADP. Selma, who leads global data and analytics strategy at Carter's, moderated.
The panel focused on how to build governance that earns trust without slowing down innovation.
Why doesn't traditional data governance work for AI?
Traditional governance is good at cleaning data for reports and dashboards, Yateski said, but AI adds inference. Even after training data is cleaned, the data that arrives when a model runs will include bad data, and systems must account for it.
Clean data also isn't enough for agentic AI, he said. Columns and fields need context that explains what they mean so generative systems can connect the dots beyond table references.
Laine said many organizations are rushing. "Our clients are skipping the metadata. They're skipping the governance," she said. Data modelers, business analysts and stewards still work in separate silos. Julia said ADP's shift from static data to data in motion, including agents that act autonomously, has pushed it toward policy-driven governance with role-based and access-based controls.
How do you measure trusted data?
Data quality profiling used to be the main signal, Laine said, but it didn't capture whether data is actually used or liked. Quest's approach scores data on nine facets, which clients can weight. As data is cataloged, it is classified into gold, silver and bronze data sets and data products. The facets include profiling results, end-user ratings, how widely data is used, whether it is curated and whether it has business policies and definitions.
The moderator noted that adoption itself can signal whether data and models are trustworthy.
How do you embed governance without slowing AI delivery?
At ADP, "number one thing is governance by design," Julia said. That means clear direction for developers, automated testing and proactive monitoring in lower environments. Some people see governance as a blocker, she said, but ADP treats it as the safe way to grow. Every generative AI use case submission asks the business to propose its governance, and governance approval is required at the proof-of-concept, pilot and general-availability stages.
"I think of AI governance as a non-IT responsibility," Yateski said. He recommended two tiers: an operational center of excellence and an executive group that sets policy and direction rather than ruling case by case. IT should make up no more than about 15% to 20% of that group, he said, with privacy, compliance, legal and every business vertical represented.
Yateski also urged fast, upfront risk assessments agreed with the business. If a pilot exceeds the agreed risk profile, pause it and move on.
How should governance handle drift, bias and hallucinations?
Laine described observability that alerts people in real time when data quality rules are violated, data drifts or personal information seeps into a model, and routes the alert into familiar channels such as Microsoft Teams or Slack. She also pointed to guardian agents mapped to an organization's AI strategy and its accepted bias thresholds, an example of "using AI for AI."
Julia said ADP pairs observability tools with human-in-the-loop review by subject matter experts for critical, sensitive client use cases.
Yateski said modern AI systems are never 100% accurate, and people often accept human advice that is far less reliable than an AI system that is right 99.9% of the time. Some cases still call for fully deterministic systems, he added. He also offered an irony: "The better your data governance becomes over time, the more you will cause model drift," because models trained on uncleaned data may not handle cleaned sources. Data governance and AI governance have to work together.
What team structures help scale governance?
ADP uses a hub-and-spoke model, Julia said. The hub sets rules, selects tools and approves use cases, while business units execute within that framework. She advised companies to start with basics, noting that ADP centralized data from acquired payroll companies on one platform before its AI journey.
Yateski expects data science, data engineering, software engineering and ML engineering skills to merge into single roles, with product managers becoming more technical. At WellStar, with 32,000 employees, he relied on change champions in each business unit to drive upskilling and gather feedback.
In a closing speed round, Laine offered two words, "converge and orchestrate." Julia stressed adaptability. Yateski's advice: "Don't wait for your data to be perfect to start an AI project because it never will be."
See sessions like this at COLLIDE 2027
COLLIDE, the largest AI conference in the Southern U.S., brings more than 1,000 enterprise data and AI leaders to Atlanta for keynotes and panels on governance, data and AI. The 2027 conference takes place Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center. Registration is coming soon at COLLIDE.
Frequently asked questions
Why does AI require a different approach to data governance?
Traditional governance cleans data for reports, but AI must also handle bad data at inference time and needs business context for columns and fields, according to Nick Yateski.
What is governance by design for AI?
At ADP it means clear developer guidance, automated testing, proactive monitoring and governance approval built into each stage from proof of concept to general availability.
How can organizations measure trusted data?
Quest's Susan Laine described scoring data on nine weighted facets, including quality profiling, usage, user ratings and business definitions, then classifying it as gold, silver or bronze.
Can better data governance cause model drift?
Yes, Yateski said: as source data gets cleaner, models trained on older, uncleaned data may not handle it, so data and AI governance must work together.
Watch: Scaling Trust: Building Data Governance That Keeps Up With AI, brought to you by Quest Software on YouTube
Event: COLLIDE AI Conference · More coverage: COLLIDE AI Conference insights
