In a Data Science Connect webinar, experts from Snowflake and Precisely, moderated by Verizon's Donald Spaulding, argued that data built for human analysts isn't ready for autonomous agents. They called for rich metadata, shared data quality rules, lineage and auditability as the foundation for agentic AI.
Key takeaways
- BI-ready isn't agent-ready. Mittal said analysts catch numbers that look wrong, but agents take data at face value and act.
- Metadata is the biggest gap. Metz said metadata must now tell AI how to use a dataset, including its known errors.
- Meaning should ship with the data. Precisely is packaging metadata envelopes with semantic registries and dataset narratives.
- Shared quality rules resolve conflicts. Sharma said teams with competing definitions of good data need one central set of rules.
- Trust is an architecture decision. Mittal named lineage, enrichment confidence and auditability as requirements for autonomous action.
Data that works for dashboards isn't necessarily ready for AI agents that act on their own. That was the central argument of a Data Science Connect expert panel webinar, Building the Data Foundation for Agentic AI with Snowflake and Precisely, brought to you by Precisely. Donald Spaulding of Verizon moderated a discussion with Sam Mittal, Partner Solutions Engineering at Snowflake; Amy Metz, Product Management Director at Precisely; and Shantanu Sharma, Product Manager at Precisely.
What does agentic AI demand from enterprise data?
Each panelist defined agentic AI in terms of what it asks of the data layer. For Metz, the data foundation now has to serve an autonomous machine rather than a human. "There needs to be guard rails put around your data foundation to make sure that the AI isn't guessing," she said. Repeatability matters too: three users asking an agent the same question should get the same answer.
Mittal described agentic AI as "what happens when the gap between insight and the action closes." With no human reviewing output before something happens in the real world, he said, data integrity has moved from a data team concern to a company-wide risk conversation.
Metz drew a line between generative and autonomous use cases. Relatively clean, structured data can be good enough for generative AI when a human catches errors. Autonomous agents need data that is integrated, governed and enriched, she said, backed by robust metadata that feeds semantic layers and gives agents strict instructions.
What data gaps undermine AI agent performance?
Mittal said the most common gap he sees isn't dirty data in the traditional sense. Organizations spent years building data for dashboards and reports, and human analysts are good at spotting a number that looks wrong. "Agents don't do that. They take the data at the face value and act," he said. Data optimized for BI hasn't been recalibrated for autonomous action, and teams often don't notice until an agent does something consequential.
Sharma pointed to competing definitions of good data. Teams keep their own master or golden datasets, which creates discrepancies. He described a telecom customer whose billing and customer support teams defined good address data differently. Precisely's answer, he said, is a central repository of data quality rules applied to Snowflake datasets.
Metz, who said she has worked in geospatial data for 29 years, argued that no dataset of hundreds of billions of records will ever be perfect, no matter how many checks run in the pipeline. "I don't think AI is going to solve that problem," she said. The bigger gap is metadata.
Why does metadata matter more for agentic AI?
Metadata used to be a box to check, Metz said: the coordinate system, the source of the data. Now it needs to tell the AI how to use a dataset, what is known about its errors, which attributes fit a given insurance or telecom use case, and how to run data flows for a specific problem.
Organizations can get quality, governance and access right and still have AI misread their data, she said, because nothing tells the model what fields mean, how they relate, or whether a dataset is fit for purpose. Precisely's "agentic ready" initiative is built on the principle that "meaning really needs to ship with the data." Each dataset gets what she called a metadata envelope, with cataloging specifications, semantic registries that map attribute values to their meanings, consistent spatial indexing, cloud-native formats such as GeoParquet, and narratives describing how datasets relate through a global ID system.
How do Snowflake and Precisely work together on AI-ready data?
Sharma walked through a modernization flow. Customers replicate business-critical data from legacy systems such as IBM mainframe DB2 into Snowflake. Precisely's Data Integrity Suite then catalogs data in both places, links it to business definitions to build a semantic layer, recommends quality rules and runs remediation and enrichment pipelines.
Connecting the suite's MCP server to Snowflake Cortex was easier than he expected, Sharma said. Using sample address data, he built an enrichment pipeline in less than three minutes. The success has created a new problem: IT teams now face tens or hundreds of MCP servers. Precisely is extending its governance tools to scan MCP servers, AI models and AI use cases to address that sprawl, he said.
Mittal outlined three requirements for trusted autonomous action: lineage showing where every piece of data came from, which he said Snowflake's Horizon governance layer provides; confidence that an address, entity or identity is correct, the layer partners such as Precisely add natively inside Snowflake; and auditability of who authorized an action, what data it used and what it decided. That record "lives in the platform architecture, not the agent prompt," he said.
What should data teams do first?
Mittal said the fastest-moving organizations share three traits. They "treat the data integrity as a continuous operational layer not a project," they commit to one data platform instead of relitigating it every quarter, and they bring in specialized partners where they need the deepest data trust.
His practical advice: turn on Snowflake Horizon's native data quality checks, lineage tracking and data contracts first. Then, within 90 days, audit one high-value AI use case by asking what data quality standard an agent needs to act on it safely without human review. Most organizations discover they need richer data, better entity resolution or deeper location context, he said.
Looking ahead, Sharma said Precisely plans to package skills and agents alongside its MCP server to make enrichment easier to try.
Catch more data and AI panels in 2027
Data Science Connect's 2027 schedule includes 24 free live webinars across four series: the AI × Enterprise Panel Series, AI Demo Days, Agentic AI Showcases and the Plug & Play Series. Find upcoming sessions on the webinars page.
Frequently asked questions
What is agentic-ready data?
As the panel described it, data that is integrated, governed and enriched, with metadata and semantic layers that tell an autonomous agent what fields mean and how to use them without human review.
Why isn't data built for BI good enough for AI agents?
Mittal said human analysts flag numbers that look wrong, while agents take data at face value and act, so data calibrated for dashboards can lead agents to consequential mistakes.
What is MCP server sprawl?
Sharma described IT teams ending up with tens or hundreds of MCP servers across pilots and use cases, which then need to be inventoried and governed.
What should data teams do first to prepare for agentic AI?
Mittal advised turning on native data quality checks, lineage and data contracts, then auditing one high-value AI use case for the data standard an agent needs to act safely.
Watch: Building the Data Foundation for Agentic AI with Snowflake & Precisely | Webinar Replay on YouTube
Event: Webinars · More coverage: Webinars insights
