In a Data Science Connect webinar, leaders from Actian, Fiserv and Slalom said AI is only as reliable as the business context behind enterprise data. They covered scalable stewardship, risk-based governance, the metadata AI needs, auditability and where automation should stop.
Key takeaways
- Stewardship is an operating model, not a person, and scales when it is embedded in everyday work and augmented with automation.
- Match governance to risk. Low-risk processes can pass automated checks while critical ones get formal review.
- Build guardrails into tools. Classifications set in the catalog should carry through every downstream use.
- AI needs context to know what it doesn't know: definitions, lineage, quality and freshness signals in a shared semantic layer.
- Let AI do the first pass. Automate drafting and lineage, but keep approval and accountability with people.
Enterprise data is only as useful as the context behind it, and for AI that context has to stay current. That was the starting point for "Context Is the New Foundation for AI-Ready Governance," a Data Science Connect webinar sponsored by Actian.
Brian Mink, co-founder and president of Data Science Connect, moderated a panel with Betty Wang, Senior Engineer at Actian; Amit Palkar, Senior Director of Data Science at Fiserv; and Jared Maslin, Senior Principal at Slalom. The session opened with a short Actian product demo.
What did Actian's data steward agent demo show?
Wang said stale metadata, unclear ownership and drifting business definitions leave AI systems working from incomplete or conflicting information. She demonstrated Actian's data steward agent, which she said is embedded in Actian's data intelligence platform and works inside catalog workflows to enrich and maintain metadata.
In the demo, she asked the agent why a staging customer dataset was incomplete. It explained the gaps against a configurable metamodel, such as missing glossary terms and descriptions, then proposed a bulk fix. Wang said its suggestions draw on peer datasets, table contents and sensitivity patterns, and that a human must review and approve changes before anything is applied. The agent isn't meant to replace data stewards, she said, but to remove the manual bottleneck.
What does scalable data stewardship look like?
Stewardship has usually been an extra duty handed to already busy teams, which is rarely sustainable, Maslin said. "Stewardship isn't a person, it's an operating model," he said. The goal is to embed it in everyday processes and augment it with metadata platforms, catalogs, lineage, data quality monitoring and, increasingly, agentic workflows.
Maslin described governance as an enablement capability rather than a control function. His shorthand: governance should create "seat belts and airbags, not speed bumps."
Palkar compared the shift to the move from prompt engineering to context engineering, from handcrafting to building systems that maintain context.
How can governance build trust without slowing teams down?
Palkar said governance should match the risk of the process. At Fiserv and other organizations, he said, a decision framework sends low-risk processes through automated checks, scripts and agents, while more critical work goes through formal review.
He also argued for building guardrails into the tools themselves, not downstream processes. Tagging a sensitive column in a catalog such as Unity Catalog, for example, carries that classification through later analysis, data contracts and verification loops. That makes governance continuous instead of a series of gates.
Maslin said the least friction comes when governance is almost invisible to users. With rich business context, organizations can move from role-based to attribute-based access controls, which make access decisions at runtime based on the request and the person, instead of forcing users to work out which role to request.
What metadata and business context does AI actually need?
It depends on the organization, Wang said. A regulated bank may weight lineage and PII classification more heavily than a marketing team focused on self-service analytics. Common needs include technical metadata, lineage and quality signals paired with clear definitions and ownership, brought together in a semantic layer so a term like "customer" means the same thing everywhere.
"AI will act on the data with total confidence whether it deserves it or not," Wang said. Sometimes the missing piece is a definition, sometimes lineage for explainability and audit, sometimes live quality and freshness signals. Palkar added that metadata now has to be machine readable as well as human readable.
Maslin warned that bad data feeding AI context can push outputs in directions that don't serve the business, in ways users may not notice.
How do enterprises make AI decisions auditable?
Responding to an audience question about reconstructing a model decision months later, Palkar pointed to end-to-end traceability of models, features, transformations and source systems. He also suggested treating catalogs and lineage as durable memory that persists beyond a single model context, alongside versioned model registries.
Maslin recommended model cards that record not just what a model was trained on but what it was intended to do. Incidents can surface six weeks or six months later, he said, when the original developer may no longer be in the role.
Where does automation help, and what stays human?
Wang said AI can take the first pass at repetitive work such as drafting definitions and classifications, assigning ownership, detecting common data patterns and stitching together lineage. Humans should still approve definitions, settle ownership disputes and sign off before anything reaches production.
On adoption, Maslin cited three factors: visible executive sponsorship, clear accountability and showing business value quickly rather than promising ROI 18 months out.
In the closing round, Palkar said context engineering "is a governance problem." Wang advised investing in the context foundation before AI use cases. Maslin's advice: "Don't boil the ocean."
More on data governance and AI 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. See the webinar schedule for upcoming sessions.
Frequently asked questions
What is AI-ready data governance?
As the panel described it, governance that keeps business definitions, ownership, lineage and quality signals current so people and AI systems can trust and safely use enterprise data.
What metadata does AI need?
It depends on the organization, Actian's Betty Wang said, but common needs are technical metadata, lineage and quality signals paired with clear business definitions and ownership in a semantic layer.
How do you scale data stewardship?
Treat stewardship as an operating model, embed it in existing workflows, and use catalogs, lineage, quality monitoring and AI agents to automate routine tasks.
How can AI model decisions be audited months later?
Panelists recommended end-to-end traceability, catalogs and lineage that persist as durable memory, versioned model registries and model cards that record intended behavior.
Watch: Context Is the New Foundation for AI-Ready Governance with Actian | Webinar Replay on YouTube
Event: Webinars · More coverage: Webinars insights
