Author: Norma: Practice Director
There is a genuinely fast way to move a dashboard from Tableau or Power BI into Databricks: You import the workbook, and it appears inside Databricks Genie looking much as it did before. For one workbook, or a handful, that native import is exactly the right tool, and it is a useful on ramp. But I spend my time with enterprises that do not have a handful of workbooks. They have hundreds or thousands. And when you import an estate that size exactly as it is, you have not modernized anything. You have copied a decade of accumulated redundancy and conflicting definitions straight into your AI layer, where it causes far more trouble than it ever did in a dashboard.
Here is the pattern I see. Legacy BI estates carry redundancy that teams have accommodated over years. One workbook we came across had ninety tabs where four would have done the job, each tab slicing the same measures a slightly different way. In another, a single metric, net revenue, had twelve different definitions across teams, each with its own small variations. In a dashboard, that is a maintenance burden. In AI, it becomes something worse.
The native import creates one metric view and one agent per workbook, one at a time, with no cleanup. Point that at an estate of several thousand workbooks and you get several thousand metric views, an unmanageable sprawl of agents, and routing that no longer knows which definition to trust. The agent will still answer, confidently, using whichever version of net revenue it happened to infer, and it may be wrong. This is why so much enterprise AI disappoints today. It is enabled on top of models that were never cleaned up or governed, so it inherits the ambiguity and returns confident, incorrect answers. McKinsey found that nearly two-thirds of enterprises have experimented with agents, but fewer than 10 percent have scaled them to real value, with eight in ten citing data limitations as the roadblock. The agents are rarely the problem. The foundation underneath them is.
Our approach starts before anything moves. We read the legacy BI, whether it is Tableau, Power BI, or another analytical tool, and assemble one unified semantic layer: The tables, joins, and measures, enriched with the descriptions and business meaning each element needs, and cleaned up so that twelve definitions of net revenue become one. Every field is traceable backward to where it came from, which workbook, who authored it, and who approved it, so the single definition is not just cleaner, it is accountable. This is the same discipline the best legacy BI always had, a governed, central semantic layer, brought into the AI world.
The second principle is that a migration should be driven by how dashboards are actually used, not by moving everything. In practice, each dashboard tends to go one of four ways.
The estate stops growing and starts shrinking toward what the business truly relies on.
The native Databricks import is an as-is migration, one workbook at a time, and for the right situation it is perfect. Our value is in everything an enterprise-scale move needs that a straight import does not do: Unifying conflicting definitions into one governed model, rationalizing the estate by usage, standing up the agentic framework with its governance and validation, and proving parity with the reports people trust today.
It is an accelerator-led approach delivered with services around it, because at this scale the judgment calls, which dashboards stay, which become a conversation, which merge, and which retire, matter as much as the automation. For a Head of Analytics or Business Intelligence, and for the Chief Information Officer (CIO) sponsoring the move, the result is analytics people can trust and self-serve through a simple chat, on a governed foundation that gets smaller and cleaner rather than larger and noisier.
To make this journey fast and consistent, we bring three accelerators, produced largely automatically.
This is how we prove the numbers still tie out, rather than asking anyone to take it on faith.
Because the target is a set of domain agents, governance cannot be an afterthought. A supervisor agent, orchestrated through Agent Bricks, routes each question to the right domain agent, and each agent draws on its own metric views in Unity Catalog. Around that sits the framework we apply to every agentic solution: Role-based security, observability that logs the sessions, questions, and responses, end-to-end lineage, and evaluation scripts that keep checking whether each agent is still accurate as it changes and as feedback comes in. The Unity AI Gateway handles budgeting and model routing. And because everything traces back to one unified model, a single change to a definition is updated everywhere at once, instead of being re-edited in workbook after workbook.
The temptation with any migration is to measure success by how fast you move dashboards. With BI to AI, that is the wrong measure. Move an estate as-is and you will have recreated your old chaos in a place where it does more harm. Unify the semantic layer first, migrate by how things are actually used, and prove every answer against the BI people already believe, and you get something better than a faster dashboard. You get a business that can talk to its data and trust what it hears.
Related reading: KPI Partners Enterprise Analytics Accelerator on Databricks
Ready to Transform Your Data Strategy? Talk to our experts and discover how KPI Partners can accelerate your data and analytics initiatives.