Building Trust in AI: Data Governance and Data Management for Generative AI
How enterprises build data trust before they ship generative AI
About this talk
This Data Science Connect panel works through the practical question most teams hit after their first generative AI demo: what has to be true about your data before an LLM is allowed anywhere near it.
The discussion covers the governance layer that sits between raw enterprise data and a model prompt — classification, access control, lineage, PII handling in retrieval, and the evaluation loop that tells you whether output is trustworthy over time.
Henry Clavo draws on production data platform work in healthcare and government, where governance is not optional and the cost of an ungoverned answer is measured in regulatory exposure, not just user experience.
Key takeaways
- Data trust is a prerequisite, not a phase two — ungoverned source data produces confidently wrong model output at scale.
- Access control has to follow the data into the vector store: retrieval must respect the same permissions as the system of record.
- PII handling in RAG needs explicit treatment — classify, mask, or exclude at ingestion rather than filtering at the prompt.
- Lineage is what makes an AI answer auditable; without it you cannot explain why a model said what it said.
- Evaluation is continuous. Ship with a measurable baseline for accuracy, grounding, and refusal behavior, then monitor drift.
- Governance succeeds when it is owned jointly by data engineering, security, and the business — not delegated to a committee.
Discussion highlights
- OpeningWhy trust comes first
The panel frames trust as a data problem before it is a model problem — most failed GenAI pilots stall on data quality and permissions, not model capability.
- Mid-panelGoverning the retrieval layer
Discussion of what changes when enterprise content is embedded and retrieved: classification at ingestion, permission-aware retrieval, and keeping regulated data out of shared indexes.
- Mid-panelPII, lineage, and audit
How to keep an audit trail from prompt to source document, and why lineage metadata becomes the evidence base for compliance conversations.
- ClosingOperating model
Where governance ownership actually sits in an organization, and how to keep guardrails from becoming a bottleneck on delivery.
Who this is for
- Data and analytics leaders standing up their first governed GenAI use case
- Data engineers designing RAG pipelines over regulated or sensitive content
- Security, privacy, and compliance partners defining AI guardrails
Tags
Further reading
- GuideData Governance for Generative AI
How to build data governance for generative AI: trust, access control, lineage, PII, RAG guardrails, and evaluation — with examples and an FAQ.
- GuideDatabricks vs. Snowflake
Databricks vs. Snowflake compared: lakehouse vs. cloud warehouse, pricing, governance, AI workloads, and how to plan a migration.