Insights at Scale With Agentic AI: How AI Ready Data is Transforming Commercial Life Sciences

Anuya Aher, Nishesh Aggarwal18 Jun 2026

Problem

A brand analytics team at Merck asked its AI agent how many new patients a brand got over the last rolling three months. The agent wrote plausible SQL and returned a confident wrong answer. The warehouse was mature (about 500 data products) but built for humans, who fill in what a dashboard leaves out. An agent does not do that and never says "I don't know".

Example from the article: for "How many New Patients have been prescribed Winrevair across R3M", the agent summed TRX in a generic sales table. It missed that the brand sells via specialty pharmacies, that a "new patient" means a first enrollment form, and that R3M has its own definition.

Solution in brief

Keep the data foundation and add six layers of business context on top, from column descriptions to KPI rules, access rules and feedback memory. Domain experts write the upper layers once, and every agent and persona reuses them. Merck builds them on native Databricks features (Unity Catalog, Genie Spaces, Agent Bricks).

Questions this case answers
  • How to make a text-to-SQL analytics agent return correct KPI answers over an existing warehouse by adding layers of business context
  • How to define rolling-window and cohort business metrics once so that an agent computes them the same way every time
  • How to make a text-to-SQL analytics agent return correct KPI answers on top of an existing data warehouse
  • How to prepare analytics-ready data products so that LLM agents can use them

Tools and methods

  • Six context layers — data, domain/system, use-case, enterprise and memory context above analytics-ready data products.
  • Genie Spaces — one per use case, brand or domain, with instructions, metrics and trusted SQL.
  • Multi-Agent Supervisor (Agent Bricks) — routes a question to the right Genie Spaces.
  • Unity Catalog — metadata, lineage, governance.
  • Feedback loop — user ratings and expert corrections feed back into context.
Continue reading — it's free

Google or a one-time email code. No password.