Join clinical volume, cost and payer data on one governed foundation — with an allocation model your clinical directors will actually accept.
The annual allocation argument is not an arithmetic problem, it is a participation problem. This guide covers the model build and, more importantly, the sequence that gets clinical leadership to sign up to the drivers before anything is engineered.
Lakehouse and medallion design in OneLake, ingestion from EHR and finance sources, allocation logic, semantic model design with row-level security, and Direct Lake configuration for reporting without a refresh window.
Shared costs — theatres, imaging, pathology, overhead — have to be attributed somehow, and every method disadvantages somebody. There is no neutral answer, which is precisely why the model has to be agreed rather than imposed.
A defensible allocation that clinical leadership signed up to is worth more than a technically superior one they dispute every year.
Check these before beginning. Most stalled implementations stall on one of them.
A Fabric capacity. F64 or above if you intend to use Copilot experiences against the semantic model.
Fabric administrator, a data engineer, and a finance owner who can approve the allocation basis.
A decision about who owns metric definitions. Without a named owner this becomes a permanent negotiation.
Allocation drivers agreed with clinical and finance leadership jointly, in writing, before engineering starts.
Configuration is straightforward once these are clear. Skipping them is why most first attempts produce something that works and cannot be maintained.
Clinical volume, cost and payer data land in one logical lake in open table format, and every workload reads the same copy. The value is not speed; it is that a published figure can be traced back to source rather than to an analyst's monthly process.
Bronze holds source data as received. Silver holds cleaned, conformed entities. Gold holds the allocated, business-ready model. Keeping the allocation logic in a defined layer means it can be inspected and changed without touching ingestion.
Power BI reads Delta tables in OneLake directly, without an import step. For reporting that clinical directors check weekly, removing the refresh lag matters more than query speed.
Settings shown are the ones that matter, not every field on the form. Values are starting points to validate against your own environment.
Run a workshop with clinical and finance leadership together. Work through each shared cost pool and agree the driver — theatre minutes, imaging studies, bed days, weighted activity.
Write down the reasoning, not just the driver. When somebody questions it in eighteen months, the reasoning is what settles it.
Land EHR extracts, general ledger detail, payer contract data and cost centre mappings exactly as received. Use shortcuts to reference data already in a lake rather than copying it.
Resist the temptation to clean during ingestion. Bronze exists so you can prove what the source said.
This is where most of the effort goes. Encounter identifiers, provider identifiers, cost centre codes and service-line mappings all need reconciling across sources that were never designed to agree.
Build a service-line dimension explicitly rather than deriving it in report logic. It is the axis everything reports on and it must be stable.
Direct costs attach to the encounter. Shared pools allocate by the agreed driver. Keep each allocation step as a separate, named transformation so the chain is inspectable.
Produce an allocation reconciliation output: total cost in, total allocated out, and the residual. If the residual is not zero, something is wrong and you want to know before anyone sees a margin figure.
Use Direct Lake on OneLake for compatibility with OneLake security and better modelling features. Name tables and columns in business terms — this matters more now that Copilot may reason over the model.
Apply row-level security so a service line sees its own performance, and design it before publishing rather than after the registrar or compliance officer raises it.
Every one of these is avoidable, and every one of them is common enough that we check for it by default.
A technically excellent model with a disputed allocation is a model nobody uses. The workshop is the project's critical path, not a preliminary.
It will diverge between reports within months. Build the dimension once, in the model, and let everything inherit it.
Consumption-based compute behaves differently from licensed software. An inefficient notebook or an over-refreshed model is expensive in a way a fixed licence never was.
Retrofitting an access model onto a published semantic layer is disruptive and it always happens at the worst possible moment.
We will build the model for one service line against your own data and walk your clinical and finance leads through the result together, before anything wider is committed.
Describe the situation in your own words.
