Databricks embeds evidence into AI
- Databricks used a new banking compliance blog and fresh product launches to argue that AI evidence should be captured during model building, not reconstructed later, as regulated firms face revised U.S. model-risk rules. - The company put Lakeflow Designer into public preview on April 22 and said Unity Catalog Business Semantics is now generally available, tying visual pipelines, lineage, governed metrics and audit trails together. - The push follows April 17 guidance from the Federal Reserve, Federal Deposit Insurance Corporation and Office of the Comptroller of the Currency that replaced older model-risk rules. (federalreserve.gov)
Banks are rewriting how they document artificial intelligence after U.S. regulators replaced older model-risk guidance on April 17, and Databricks is pitching its platform as the place to do it. (federalreserve.gov) (databricks.com) The Federal Reserve, Federal Deposit Insurance Corporation and Office of the Comptroller of the Currency rescinded SR 11-7, OCC 2011-12, FIL-22-2017 and related issuances, and replaced them with a more risk-based framework for model risk management. (federalreserve.gov) (occ.gov) (fdic.gov) Databricks said that change means firms need evidence generated inside the model-development process itself, including lineage, approvals, testing records and controls, instead of assembling documentation after deployment. (databricks.com) Model risk management is the paperwork and testing that shows a bank can explain how a model was built, what data it used, where it can fail and who approved it. Databricks is arguing that this record should be produced automatically by the platform that trains and serves the model. (occ.gov) (databricks.com) That argument arrived alongside a product release. On April 22, Databricks put Lakeflow Designer into public preview as a visual, no-code, AI-native tool for preparing and analyzing data inside Databricks. (databricks.com) (learn.microsoft.com) Databricks said Lakeflow Designer keeps data in place, runs under Unity Catalog governance and produces production-ready code, while showing step-by-step previews of each transformation. The company said those previews are meant to make AI-generated data changes easier to review and trust. (databricks.com) (docs.databricks.com) A second piece is business semantics, the shared definitions behind terms like revenue, active customer or churn. Databricks said Unity Catalog Business Semantics is now generally available so dashboards, SQL queries, notebooks and AI agents can use the same governed metrics and dimensions. (databricks.com) Databricks also said it is open-sourcing the core implementation in Apache Spark, extending those definitions beyond one product and into the broader data ecosystem. (databricks.com) Put together, the company is selling a stack where the meaning of the data, the pipeline that shaped it and the evidence around the model all live in one governed system. That is the pitch to banks and other regulated firms now adapting to the April 2026 guidance. (databricks.com 1) (databricks.com 2) (databricks.com 3) The closing message from Databricks is less about one feature than about where compliance work happens. In its telling, evidence is no longer a report written after the model ships, but a byproduct of building the model on the platform. (databricks.com)
