Siemens Healthineers’ path to Genie starts with rebuilding the data estate

Siemens Healthineers is preparing to put Databricks Genie One in front of more than 1,000 analytics users—but the useful lesson is what had to happen before anyone could ask a question in plain language. The company says it first rejected an expensive cloud lift-and-shift, rebuilt a two-decade-old SQL estate as governed data products and began migrating those products into Unity Catalog. Databricks’ customer account describes Genie as the self-service layer at the end of that sequence, not the starting point.

The NL2SQL launch is still a target

The source is careful about deployment status. Siemens Healthineers says its migration is “nearly complete” and that the team is finishing both the Unity Catalog move and dataset descriptions so more than 1,000 users can eventually query with Genie One. Students are prototyping cloud-based AI, but the account does not report a production Genie accuracy rate, adoption number or completed rollout.

That distinction matters. The company already has more than 1,000 people using Qlik Sense dashboards, mostly through Excel rather than SQL. Questions outside those dashboards currently flow to a central team. Genie is intended to reduce that bottleneck, but Databricks’ own case study presents governed cataloging and description work as prerequisites.

Why the rebuild came first

The underlying XMART platform ingests roughly 100 TB of MRI scanner data each month for service, field-stability and post-market-surveillance work. According to the case study, the estate includes billions of event-log lines accumulated since 2007. An early proof of concept that moved the existing architecture to the cloud unchanged was estimated to cost two to four times as much as staying on premises.

Siemens Healthineers instead rebuilt XMART on Delta Lake and split the monolith into discrete data products. It is moving those products into a single Unity Catalog metastore one at a time. The operational change is straightforward: other teams receive permissions to shared products instead of requesting encrypted exports and physical transfers. Historical access that could take weeks or months is expected to become near-instant.

The company reports that the redesigned platform should cost about 50% less than the on-premises setup, while targeted work cut the costs of two individual data products by 80% and 90%. Those are vendor-published customer claims, not an independently audited comparison, but they explain why cost attribution and data-product boundaries precede the conversational interface.

The practical takeaway

For teams evaluating enterprise NL2SQL, this case argues against attaching a chat layer to a legacy warehouse and calling the migration complete. Siemens Healthineers’ order of operations is: reject the uneconomic lift-and-shift, separate the monolith into measurable products, replace copies with governed access, describe the datasets, and only then broaden natural-language querying.

Genie may eventually absorb routine questions now queued behind a central team. The case study does not yet prove that outcome. What it does document is the infrastructure contract required to attempt it at 100-TB-per-month scale.