Accelerating the Speed to Value of Analytics with Databricks AI/BI & Agents

After consolidating analytical datasets on Databricks, the team moved dashboards, semantics, and self-service there as well, then used agents to translate legacy models into metric views. For analytics leaders evaluating modern BI foundations, the complete guide to business intelligence and analytics in the AI era offers useful context for the approach Adapar took.

The talk is a practical blueprint for analytics leaders modernizing BI while preparing for agents and conversational analytics. The core lesson is simple: agents pay off only when metadata, semantics, and skills are treated as first-class assets.

Building analytics where the data lives

Keeping analytics on the same platform as ingestion reduced cross-warehouse friction and simplified governance. AI/BI Dashboards inherit Unity Catalog controls, which eliminated duplicative permissioning across a warehouse and a separate BI tier. The presence of Genie created a credible path to self-service for non-technical users, a group that had been hard to engage in prior tools.

Metric views were decisive because they provide a codified semantic layer that reduces "dueling dashboards." Since metric views are YAML-based, the team preserved flexibility to use partner BI tools for niche needs while keeping a single source of truth for metrics.

From plan to migration: translating semantics and dashboards

The migration started with an inventory of dashboards and models, guided by usage telemetry and stakeholder interviews. Low-usage assets were not assumed to be low value, so the team validated business criticality with domain owners. Work then proceeded in tranches prioritized by impact and adoption.

Semantic translation was the hardest lift. Early trials with general-purpose assistants involved heavy back-and-forth and limited platform knowledge. Progress accelerated with the Agent Bricks, which bundled current agent skills and Databricks context. The team initially moved some dashboards with SQL behind widgets to validate front-end behavior, then iterated toward full metric views. Parity checks against the legacy platform acted as a quality gate for every tranche.

Metadata as the multiplier for AI and BI

Agent performance correlated strongly with the richness of table and column metadata in Databricks. Where metadata was complete, first-pass semantic models were usable. Where it was sparse, translation required extended conversations before results matched benchmarks. AI can suggest column descriptions and candidate metrics, but human review remained non-negotiable given company-specific field usage.

Three metadata practices emerged as high leverage:

• Markdown-style backlinks that connect tables, schemas, and fields, plus column contracts that document keys and join cardinality.
• Narrative context that explains who uses a dataset, which processes it supports, and what questions it answers.
• An eye toward ontologies as they mature, without waiting on them to begin enriching metadata today.
   

A quick audience poll suggested most organizations have less than 25 percent metadata coverage, which the speakers framed as a sizable opportunity to boost agent accuracy.

Codifying agent skills to sustain quality

Skills turned ad hoc prompt knowledge into repeatable procedure. Drawing on published benchmarks, the team noted that agents without skills hovered around 21 percent accuracy, while combining metadata, semantics, and skills pushed accuracy into the mid to high 90s. Accuracy decayed without upkeep, reinforcing the need for active maintenance and potential ownership by a dedicated role.

Adapar captured translation nuances, diagnostic routines, dashboard conventions, and dataset discovery briefs as reusable skills. Genie Code improved over time, and both Genie and Claude were used to abstract successful prompts into team-wide instructions, promoting standardization across analysts and use cases.

Invest in the foundations agents depend on

AI did not replace core data work. Teams that enrich metadata, standardize logic in metric views, and maintain focused agent skills will compound gains as agent capabilities improve. Teams that skip these steps will spend cycles debugging conversations instead of delivering insight.