In a Data Science Connect webinar sponsored by Fivetran, data leaders from Fivetran, General Mills and Webster Bank explained why lakehouses suit enterprise AI. They covered open table formats, fine-grained governance, cost control and how to migrate from a warehouse without losing momentum.
Key takeaways
- Lakehouses combine both worlds. Governed data on the lake side supports business-specific consumption on the warehouse side, Srivastava said.
- File formats matter. Lynn said open table formats such as Apache Iceberg and Delta Lake beat defaulting to CSV because they add schema and smarter query planning.
- Fine-grained governance needs a catalog. Object-storage permissions stop at the file level; row and column controls require a catalog layer.
- Fewer copies, lower cost. A single governed source of truth and right-sized compute engines cut storage and processing spend.
- Start with one use case. The panel advised proving value early instead of launching a long project that delivers late.
Data lakehouses combine the governance of a data warehouse with the flexibility of a data lake, and that combination is what makes them a strong foundation for enterprise AI, according to a Data Science Connect expert panel webinar sponsored by Fivetran. The panel featured Dan Lynn, Vice President of Product Management, Databases, at Fivetran; Sanchit Srivastava, Senior Manager, Data Analytics, at General Mills; and Ashish Dibouliya, Managing Director, Data Architecture, at Webster Bank. Brian Mink, co-founder and president of Data Science Connect, moderated.
The panelists, who together brought decades of experience from relational databases through Hadoop to cloud object storage, focused on practical questions: file formats, governance, cost control and how to migrate without stalling.
What is a data lakehouse and why does it matter for AI?
Srivastava said warehouses struggled with speed, since requirements often changed by the time a warehouse was built, while lakes were built without clear plans for how data would be consumed. A lakehouse, he said, takes the best of both: governed, well-modeled data on the lake side and business-specific consumption on the "house" side.
Dibouliya said a lakehouse supports structured and unstructured data on one platform, reduces data movement and streamlines the AI lifecycle from ingestion and preparation to model training.
Lynn warned against simply pointing AI agents at operational systems, which he said introduces scalability problems. Lakehouses build on a decade of lessons about decoupling compute and storage. Data sits in the organization's own object storage, such as S3, Google Cloud Storage or ADLS, and teams can bring different query engines to it. Tools such as the Model Context Protocol, retrieval-augmented generation and vector databases then act as "a last layer of transformation," he said, putting trusted data in front of agentic workflows.
How should teams handle structured, semi-structured and unstructured data?
Dibouliya listed the recurring challenges: ingesting and normalizing diverse formats, keeping metadata and governance consistent, handling schema evolution and securing access.
Lynn said file format choice deserves careful thought. He said he has seen many customers "get burned" by defaulting to CSV, a habit he called common in financial services. Open table formats such as Apache Iceberg and Delta Lake add schema and let query engines prune files intelligently, avoiding the old Hadoop pattern of scanning petabytes to query megabytes.
Srivastava said colocating all data types is a major accelerator because AI can find patterns developers were not looking for. "AI is a scavenger," he said.
How do lakehouses support AI data governance?
Lynn said object-storage permissions only control access at the file level. For row-level filtering and column-level permissions, teams need a catalog in front of the data, such as the Apache Iceberg REST catalog or its Polaris implementation. Because storage lives in the organization's own cloud accounts, operations teams can apply the same access policies they already use elsewhere.
Open table formats also help with lineage, Lynn said. A single transaction log records changes to a table even when different engines, such as Apache Spark, Trino or a cloud data warehouse, write to it.
Dibouliya named four governance pillars that lakehouses address: data quality, data management, fine-grained access control, and built-in lineage and auditability. He said those features support compliance with frameworks such as GDPR and HIPAA.
How can enterprises control lakehouse storage and compute costs?
"Copies of data are expensive," Lynn said. A single governed source of truth that many departments can query eliminates duplication. Organizations also control their own storage policies, can use negotiated cloud pricing and can tier old records to cold storage.
On compute, Lynn said Fivetran's analysis found a lot of cloud warehouse spend goes to merging incoming records into existing tables. Using an engine sized for that job frees capacity for other work. Running expensive warehouse engines on routine processing, he said, is like buying a sports car to mow the lawn.
Dibouliya recommended columnar formats, autoscaling, lifecycle management, reduced duplication, intelligent caching and cost-monitoring tools. Srivastava's advice was simpler: since most cost is compute, "be smart for your use cases" and match low-priority work to low-cost environments.
What are the pitfalls of migrating from a data warehouse to a lakehouse?
Srivastava said the biggest challenge is the mindset shift and change management. Technically, teams should watch large tables, since lakehouses depend on file sizing and chunking that a warehouse handles behind the scenes.
Lynn said a lakehouse lifts "the hood a little bit." Iceberg and Delta Lake keep snapshots for time travel, and without table maintenance they pile up and erase cost savings. Teams also need to choose between copy-on-write and merge-on-read for each table. He said Fivetran's managed data lake service writes Iceberg and Delta tables and handles compaction and snapshot expiration. He added that many cloud warehouses now let users query Iceberg tables with familiar SQL, easing the transition.
Dibouliya said existing regulatory reporting must keep working after migration, and that training and adoption are the hardest parts.
In audience Q&A, the panel said federated queries let data stay in its region while compute moves to it. For teams with siloed data, all three panelists recommended starting with one use case, proving value and turning it into a framework. Lynn cautioned against a 12- or 18-month plan that delivers nothing until the end, because "it's very easy to lose faith along the way."
More data architecture discussions in Data Science Connect webinars
Data Science Connect hosts 24 free live webinars on its 2027 schedule across four series: the AI × Enterprise Panel Series, AI Demo Days, Agentic AI Showcases and the Plug & Play Series. Sessions bring enterprise practitioners and vendors together on topics like lakehouse architecture and AI-ready data. See the webinar schedule.
Frequently asked questions
What is the advantage of a data lakehouse for AI?
Panelists said a lakehouse stores structured and unstructured data on one governed platform, reduces data movement and lets teams use different query engines on the same data.
Why use Apache Iceberg or Delta Lake instead of CSV?
Lynn said open table formats add schema and let query engines prune files, avoiding full scans and the quality problems that come with CSV.
How do you govern access to data in a lakehouse?
Object-storage permissions control files; for row-level filtering and column-level permissions, Lynn said teams need a catalog such as the Iceberg REST catalog in front of the data.
How should an organization with siloed data start a lakehouse?
The panel recommended picking one use case, bringing in only the data it needs, proving value and reusing that as a framework.
Watch: From Chaos to Clarity: How Data Lakehouses Empower AI at Scale on YouTube
Event: Webinars · More coverage: Webinars insights
