AI strategies fail when the data underneath them isn't right. At COLLIDE 2024, leaders from Southern Company, Clorox, Morgan Stanley and Tecton explained how to tie data strategy to business value, break down silos with standardized features, and scale pilots that break at full volume.
Key takeaways
- Tie data and AI to business value. Whatley said neither strategy is a one-time exercise, and both need feedback loops.
- Centralize and standardize. Ferragut said a single features-as-code platform reveals repeated features and cuts pipeline costs.
- Three pillars hold it up. Gulati named data quality, governance and scalability as the critical elements.
- Pilots break at scale. A call center tool that worked for 10,000 customers stalled at full volume, Whatley said.
- Real-time data enables in-flight changes. Clorox shifts ad spending quickly when click-through or cost-per-click metrics disappoint.
A good AI strategy starts with a good data strategy, and both have to be tied to business value. That was the consensus of a panel at COLLIDE Data & AI Conference 2024 in Atlanta, in a session sponsored by Tecton. Prital Ullal, Executive Director of Wealth Management Technology at Morgan Stanley, moderated a discussion with Sergio Ferragut, Principal Developer Advocate at Tecton; Nick Whatley, Director, Enterprise Data & Customer Analytics at Southern Company; and Arpit Gulati, Associate Director, Data & AI at Clorox.
How should organizations align data strategy with AI goals?
At Southern Company, data and AI strategies run in parallel, and neither is a one-time exercise, Whatley said. The key to winning buy-in is tying both to outcomes for employees, shareholders and customers. "It's not just data for data's sake or AI for AI's sake," he said.
Data is the fuel for analytics, machine learning and AI, and if it isn't right, the strategies built on it will fail and need revisiting quickly, according to Whatley. He also stressed feedback loops: data quality and availability on one side, and model accuracy and usability on the other. Southern Company is now connecting strategies that were once built in separate pockets, including data, AI, compliance and governance. He said that's "easier said than done."
How do you inventory data products in a data-rich company?
Consumer packaged goods companies have data from retail point-of-sale systems, household panels, consumption, media and social listening, Gulati said. Clorox prioritizes data by the KPIs it wants to move, such as market share or media optimization. It then audits the data and uses data health trackers to catch incomplete or inaccurate feeds before they produce biased results.
The harder part is connecting sources, he said. Rather than looking at social media trends in isolation, the team tries to tie a campaign's search and social performance to its return on investment.
How do you break down data silos for AI?
Ferragut said Tecton often sees customers whose data science or data engineering teams have each built separate solutions. The result is wasted people, wasted compute and higher costs. His prescription: "centralize and standardize."
Tecton promotes "features as code," a standardized way to build data pipelines for structured and unstructured data. When customers migrate disparate streaming, real-time and batch pipelines to a single platform, Ferragut said, they find many repeated features, which reduces the number of pipelines and the cost of maintaining and monitoring them. Other teams can then reuse that work.
What are the critical elements of a data strategy for AI?
Gulati named three pillars. Data quality comes first, especially because generative AI tends to arrive "top down" from executives, unlike earlier machine learning work that data scientists brought from the bottom up. Governance is second, because Clorox handles first-party consumer data such as names and email addresses. Scalability is third. Proofs of concept built quickly for a CMO often break when scaled, and adoption falls, he said.
Whatley said quality is "in the eye of the beholder." Some regulated uses demand near-perfect data, while a model may only need data that's good enough, so teams should define quality with stakeholders. Ferragut added training-serving skew, when data is processed differently in production than it was for training. He said standardized features as code can eliminate it.
For future-proofing, Ferragut said the stack must evolve as the field shifts from structured data to embeddings and RAG. Tecton recently released chatbot personalization that combines RAG on unstructured data with structured data, he said.
Why do AI pilots fail to scale?
Whatley recalled sitting with call center representatives using a new tool. It worked well for 5,000 to 10,000 customers. At full scale, it sat spinning while representatives were on the line with customers, which explained low adoption. Seeing the tool from the user's seat is how teams find the technical details that drive performance and belief, he said.
He also encouraged teams to involve stakeholders throughout projects and to welcome unexpected model outputs as a chance to question assumptions. "It's data science for a reason," he said. Ullal noted that mergers and acquisitions and model drift both require retraining.
How does real-time data help marketing analytics?
Consumer trends in Clorox's beauty and skincare lines change quickly, often driven by influencers, Gulati said. Real-time data supports in-flight optimization. If click-through rates are low or cost per click is high, the team can quickly shift spending, such as from online video to connected TV ads. He said consumers will pay a premium for personalization, and that market share growth increasingly comes from value rather than volume.
In closing, Gulati summed up the panel: "Your data quality is really the fuel, and scalability is the engine."
Explore data and AI strategy 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. The 2027 conference takes place Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center, and registration is coming soon. Data Science Connect's free live webinars cover data infrastructure for AI throughout the year.
Frequently asked questions
How do you align data strategy with AI strategy?
Tie both to business outcomes, connect efforts that were built separately, and build feedback loops on data quality and model usability, Whatley said.
What is features as code?
A standardized way of defining data pipelines and machine learning features in code on one platform, which Ferragut said reduces duplicate pipelines and eliminates training-serving skew.
What is training-serving skew?
A mismatch between how data is processed for model training and how it is processed when the model serves predictions in production.
Why do AI pilots fail to scale?
Panelists said proofs of concept are often built quickly for a small group and break or slow down at full volume, which lowers adoption.
Watch: The Best AI Strategy Begins with a Good Data Strategy, brought to you by Tecton on YouTube
Event: COLLIDE AI Conference · More coverage: COLLIDE AI Conference insights
