Data democratization efforts often stall for reasons that have little to do with tools. In a Data Science Connect panel, leaders from AtScale, T-Mobile and Mars said success depends on balanced governance, deep data literacy and leadership, and that AI is now exposing every gap.
Key takeaways
- Neither extreme works. Fully central control slows access; fully distributed stacks destroy trust in data, Mariani said.
- Use hub and spoke. A center of excellence supplies tools, governance and training while business teams build their own data products.
- Teach the full arc. Literacy programs should cover hypothesis to statistical interpretation to create real power users.
- Start small. Build semantic models for one painful use case instead of waiting for a full migration.
- AI audits your data. Agents expose weak data foundations, and the lasting competitive edge will be data, not AI itself.
Enterprises keep spending on self-service data platforms, yet many data democratization efforts stall. In "The Readiness Gap: Why Many Enterprises Fail at Data Democratization," an AI for Enterprise Webinar Series panel sponsored by AtScale and moderated by Data Science Connect co-founder and president Brian Mink, three practitioners argued the gap is mostly about governance design, data literacy and leadership, not tools. The panel featured David Mariani, CTO and founder of AtScale; Ugandhar Dasi, Enterprise AI Architect at T-Mobile; and Shashank Kadetotad, Senior Director, Enterprise Data Science & AI Delivery in the Mars Snacking Chief Data Office. Mink noted that panelists' comments don't necessarily reflect their organizations' views.
Why do data democratization initiatives fail?
Mariani drew on his time running analytics at Yahoo, where each business unit ran its own stack. Without a way to reconcile them, the company couldn't agree on basic metrics like clicks and impressions, which made comparing sites or search and display performance next to impossible. Swinging from fully central to fully distributed control removes governance and "destroys trust in data," he said.
Kadetotad said he has seen the pendulum swing the other way too, with overgoverned processes that make timely access almost impossible. Both he and Dasi said the people and process side is often an afterthought: training, usability and executive support matter as much as the platform.
What does organizational readiness for self-service analytics look like?
For Dasi, readiness starts with data people trust: reliable, consistent and governed, with catalogs and lineage that explain what the data means. Then comes user experience and data literacy programs that invite nontechnical users in.
Mariani recommended a hub-and-spoke model. A center of excellence, typically the BI or IT team, provides shared tools, governance and training without acting as gatekeeper, while business teams build their own data products. "All IT controlled or all business controlled, neither one of those to me are a good idea," he said.
Kadetotad said literacy programs should go beyond access to teach the full problem-solving arc, from forming a hypothesis to interpreting statistical models. That, he said, is how organizations actually develop power users.
Asked how to start semantic modeling when infrastructure is still immature, Mariani said "you don't need to boil the ocean." Pick a use case with real business pain and build semantic models for it. That focuses the work, can guide the order of a cloud migration, and shields users from the switch when data moves from one platform to another.
How do you balance data access with security and compliance?
Kadetotad pointed to role-based and more granular attribute-based access, plus automated classification and tagging. "You can't govern what you don't know," he said, noting that even mature companies don't always know where all their data lives.
Dasi described security and privacy by design built on least privilege. To keep that from blocking experimentation, organizations can offer sandboxes with anonymized or synthetic data. He added that newer masking and access-control technology is narrowing the gap between security and usability.
Mariani took a more provocative line. He would invest in role-based access controls at the data platform layer so policies apply however the data is consumed, then aim for everything to be available by default, with access denied by exception. Agents make that shift urgent, he said, because no one can predict what data they will need. "Sipping data through a straw doesn't make sense in the era of AI and agents and LLMs that are starving for data," he said.
How is AI raising the stakes for data readiness?
AI is "a brutal auditor of your data infrastructure," Kadetotad said. It promises business users basic analysis with minimal technical skill, but exposes every weakness in the data underneath. Dasi said chatbots and SQL agents ask questions no one prescribed, so poor data leads straight to hallucinated answers. Mariani said customers also worry about proprietary data leaking to general-purpose LLMs, a problem he said he hasn't seen solved.
Looking ahead, Mariani described a move from people-driven analytics to autonomous analytics. Some AtScale customers run 120,000 queries a day from people, he said; agents could push that into the millions or more, raising the bar on governance and accuracy. Kadetotad expects AI itself to become commoditized. "The competitive advantage won't necessarily be AI per se," he said. It will be the data.
How do organizations sustain data democratization?
Dasi credited data councils and architecture review boards for helping his teams navigate new technology. Kadetotad favored a strong central chief data officer function that sets standards, with data stewards embedded in domains as bridges, plus guilds, open office hours and goals that reward data-driven decisions. Mariani measures success by usage: "people vote with their feet," and the number of users and queries should keep rising.
All three stressed leadership. In the lightning round, Kadetotad named leadership endorsement, Dasi named data and culture, and Mariani named an organization that supports the business with governance, adding a plug for semantic layers.
Join the next DSC panel
The AI × Enterprise Panel Series is one of four series on Data Science Connect's 2027 schedule of 24 free live webinars, alongside AI Demo Days, Agentic AI Showcases and the Plug & Play Series. See the lineup on the webinars page.
Frequently asked questions
Why do data democratization initiatives fail?
The panel cited fragmented stacks with inconsistent metrics, overgoverned access processes, weak training and usability, and a lack of leadership support.
What is the hub-and-spoke model for data governance?
A central center of excellence, usually BI or IT, provides shared tools, governance and training, while business units build their own data products on that foundation.
How do you balance data access and security?
Panelists recommended role- and attribute-based access, automated classification and tagging, least-privilege design with sandboxes of anonymized or synthetic data, and controls applied at the data platform layer.
How do you measure the success of data democratization?
Dave Mariani measures consumption: the number of users and queries against data products should keep rising over time.
Watch: The Readiness Gap: Why Many Enterprises Fail at Data Democratization on YouTube
Event: Webinars · More coverage: Webinars insights
