Data Democratization With Generative AI: Context, Quality, Permissions
In a 2024 AI Connect webinar, executives from Informatica, Graft, Metaphor and Snowflake discussed how generative AI can widen access to data. They said semantic context, data quality, permissions and focused first use cases decide whether it succeeds.
By The AI Enterprise Magazine Team5 min read
Key takeaways
- Access is not enough. Oliner said data must be aggregated, summarized and corroborated before non-experts can use it.
- Context is the hard part. LLMs need a semantic layer describing how tables relate or they will make things up.
- Rethink permissions. AI inferences combine data with different access rules; Oliner suggested giving AI the access of a typical team member.
- Define the outcome first. Data readiness depends on the use case, not on abstract cleanup.
- Start small, and buy before building. Focused outcomes build trust; Oliner advised trying off-the-shelf tools before building a stack in-house.
In a 2024 AI Connect webinar sponsored by Informatica, Metaphor, Graft and Snowflake, four vendor executives discussed how generative AI could widen access to company data and what stops it from working. The panelists were Rudra Ray, Director of Technical Marketing at Informatica; Adam Oliner, CEO and founder of Graft; Kirit Basu, Head of Product at Metaphor; and Julian Forero, AI Product Marketing Lead at Snowflake. Brian Mink, co-founder and president of AI Connect, moderated.
Their shared message was that natural-language access opens data to non-technical staff, but quality, context, permissions and focus decide whether it works.
What does data democratization mean in the age of generative AI?
Forero said the long-running goal is a single source of truth that everyone can reach. The hard part, he said, is not putting data in one place but making it useful, which means someone must own its quality and freshness.
Basu said past democratization mainly served the data-savvy part of an organization. Natural language changes that, he said, because people can now ask simple questions.
Oliner argued that access alone falls short, using a library analogy: having a library card does not teach anyone surgery. "Access to data is not enough," he said. Information must be aggregated, summarized, enriched and corroborated, and AI is now a second consumer of company data alongside employees.
What challenges do companies face when democratizing data?
Oliner split the problems into political and technical. On the technical side, he said real company data is rarely complete, correct, clean or current, so search on top of it means "you get the wrong answer faster." He cited a statistic he had seen that only 9% of executives who put an AI solution into production in manufacturing were happy with it. He did not name the source.
Forero said the hardest technical problem for text-to-SQL is context. Sales and marketing each have their own tables, and if an analyst struggles to relate them, an LLM will make something up. The challenge, he said, is "how do we put human knowledge into technical systems," through a semantic layer that describes how data relates.
Ray added that poor data leads models to share wrong answers confidently, and that a resume-screening model trained on past hires could become biased.
Where should organizations start?
Basu said the worst mistakes he sees come from using a tool because it came free, such as a copilot, without a clear goal. He advised teams to start small with "extremely focused outcomes" and expand once they work.
Oliner said readiness depends on the use case, so organizations should decide the outcome first.
Forero suggested starting with tasks that don't require changing the model, such as text analytics on documents, before moving toward fine-tuning and agents. Many teams want to chat with their data first, he said, when other use cases can prove value sooner.
How do permissions change when AI joins the team?
Oliner called permissions an overlooked problem. AI can combine data from systems such as a CRM, a ticketing tool and Slack to infer something new, he said. If the inputs have different access rules, either everyone can see the inference and information leaks, or nobody can and the AI is unusable.
His suggestion was to think of the AI as a teammate and give it the access a typical member of a team would have. Anyone on that team can then see its results. He said tools that simply pass through underlying permissions would disappoint.
Basu added that frameworks for judging the quality of unstructured data are immature. He and Ray both favored narrow assistants for high-stakes decisions. Ray said critical business decisions need what he called zero hallucination, while exploratory analysis can tolerate some.
How should companies measure the return?
Forero said return has two sides, what a project brings in and what it costs, so teams should keep their stack simple enough to track costs. Early gains are mostly productivity, he said, which later may show up as revenue lift or lower churn. Oliner said defined outcomes set the metrics, and the right comparison is the status quo rather than perfection. Ray recommended small early wins with retrieval-augmented generation, which avoids the cost of fine-tuning.
Should enterprises build or buy?
Oliner said engineers like to build, but embeddings, vector databases and foundation models are new, so developers who say they can build the stack themselves speak "from a place of ignorance." His advice: "almost always it makes sense to buy something that solves the problem today rather than try to build something yourself."
Basu said companies outside the data and AI business should choose best-of-breed tools. Forero noted that everyday applications will ship their own copilots, which may suffice for teams working in one application.
What comes next?
Forero expected easier ways to talk to data, through semantic layers, graphs and vector storage, and LLMs tuned to specific jobs. Ray hoped for more citizen data scientists using low-code tools. Oliner described systems of record, which now include unstructured data and knowledge in people's heads, systems of intelligence that synthesize it, and systems of engagement such as chatbots and APIs. Basu said frameworks for measuring AI outcomes have not yet been invented.
In the closing round, Oliner said "the worst thing that generative AI did was trick people into thinking easy solutions to hard problems." Forero said data strategy comes first, Basu urged clarity of outcome, and Ray said an AI is only as good as its data.
Hear more from data and AI leaders in 2027
AI Connect hosts 24 free live webinars on its 2027 schedule, in four series: the AI × Enterprise Panel Series, AI Demo Days, Agentic AI Showcases and the Plug & Play Series. Details are at /webinars/. The AI Enterprise Conference 2027 takes place Tuesday, August 31, 2027, at Pier Sixty in New York City (/events/ai-enterprise-conference/).
Frequently asked questions
What is data democratization?
In the session, it meant giving everyone, not just data specialists, usable access to trustworthy data, with ownership of quality and governance in place.
Why can't generative AI just search company data?
Adam Oliner said company data is rarely complete, correct, clean or current, so simple search returns wrong answers faster.
How should AI permissions work?
Oliner suggested treating AI as a teammate with the access a typical team member would have, so results are visible to that team.
Where should companies start with generative AI on data?
Panelists advised choosing a focused outcome, defining readiness for that use case and starting with tasks that don't require changing the model.
This article is based on remarks by Rudra Ray, Director, Technical Marketing, Informatica; Adam Oliner, CEO and Founder, Graft; Kirit Basu, Head of Product, Metaphor; Julian Forero, AI Product Marketing Lead, Snowflake; moderator Brian Mink, Co-Founder and President, AI Connect at AI Connect webinar, 2024 on April 25, 2024. More coverage: Webinars.
Filed under: Data & AI Infrastructure, AI Governance & Trust, Context, Semantics & RAG
