Home / Insights / Operating model
Facts, meaning, decisions: a masterclass in the Domain Operating Model.
Analytics and AI rarely stall for lack of technology. They stall because nobody decided who owns facts, who owns meaning and who owns decisions. This is the operating model that decides it, and the red lines that keep it honest.
The meeting that never ends
It is the quarterly review of a large health system. The CFO opens with total cost of care: $612 per member per month. The population health lead has $587 from the claims warehouse. The new AI assistant answers $641 and cites a table nobody recognizes.
Everyone at that table is competent. Every number is defensible. And the next forty minutes will be spent on reconciliation instead of on what the meeting was for: deciding what to do about the cost of care.
An illustrative scene. The figures are invented for the example.
I have watched versions of this meeting for most of my career. The instinct is to blame the platform, the data quality or the team. It is almost never any of those. Nobody decided where a fact ends and a meaning begins, and who owns each.
A scale problem, not a tooling problem
A central data team works at twenty dashboards. It breaks at two thousand questions. The model asks one team to do three different jobs for the whole enterprise, and demand grows faster than any single team can.
The symptoms are familiar:
- Demand outruns delivery. Every new question joins a queue.
- Definitions fragment. Each team encodes its own version of revenue or readmission.
- Governance becomes a bottleneck. Manual reviews are slow, so they are skipped.
- AI creates parallel data universes. Each initiative pulls its own extract and defines its own labels.
Enterprises usually respond by buying a new platform. The same operating model moves onto it, and the same arguments happen faster. Replatforming a confused operating model does not remove the confusion. It modernizes it.
The primary constraint on analytics and AI at scale is not technology. It is operating-model design.
Any stalled analytics transformation has three pairs of responsibilities fused that should have been kept apart.
| Fused together | What goes wrong | What should happen |
|---|---|---|
| Creating data and interpreting it | The team that loads the data also decides what it means, one report at a time. | Facts are produced by one owner; meaning is assigned by another. |
| Storing data and giving it meaning | KPIs are baked into tables, so every table becomes a competing definition. | Storage holds facts. Meaning lives in one governed layer. |
| Governing and delivering | The governance team becomes the delivery team, and a queue. | Standards are set centrally and enforced by the platform; domains deliver. |
The one principle: facts and meaning are different jobs
The Domain Operating Model rests on one non-negotiable distinction: two kinds of domain, with different responsibilities, owners and homes in the architecture.
| Domain type | Produces | Owns | Lives in | Measured by |
|---|---|---|---|---|
| Data domains | Trusted, reusable facts | Quality, availability, evolution and service levels of data products | The lakehouse: Bronze → Silver → Gold | Reuse, freshness, trust |
| Business domains | Trusted meaning and decisions | Metric definitions, KPI logic and decision logic | The semantic and metrics layer, and the decisions built on it | Decision speed, adoption, outcomes |
Data domains produce facts. Business domains produce meaning.
This is where many data mesh programs quietly went wrong. They gave every domain end-to-end ownership, from source to dashboard. Each produced its own facts and its own meaning, and the enterprise ended up with a well-governed version of the original problem: six certified definitions of revenue, each correct inside its own domain.
Separating the two jobs gives the enterprise three single points of truth, each with one owner: certified facts, published meaning and decision intelligence. Governance runs across all three, on one shared platform.
From pipelines to products
In the old model, data was a by-product of reporting: a request arrived, a pipeline was built, a report shipped. Success was measured in pipelines delivered.
In a domain operating model, each data domain delivers durable data products. A data product is not a table with a nice name. It has:
- A named owner accountable for it, not a team alias
- A published contract: what it means, and what will never change without notice
- Service levels and quality tests, visible to consumers
- Lineage and known consumers, so a change can be assessed before it is made
The measures change with it: a data domain is judged on reuse, trust and time-to-value. Pipeline count stops being a measure of anything.
We stop delivering pipelines and start delivering products.
Draw domains around capabilities, not systems
The most consequential design decision is where you draw domain boundaries, and the most common mistake is to draw them around systems. "The Epic domain." "The SAP domain." It feels natural because that is how the data arrives. It is also why so many domain designs fall apart at the first migration.
Anchor domains to enduring business capabilities, the things the organization will always need to do, whatever it runs them on.
| Capability-aligned data domain | Example source systems today |
|---|---|
| Clinical Care | EHR / EMR, departmental clinical systems |
| Financial Transactions | ERP, patient accounting, claims |
| Workforce | HRMS, scheduling, time and attendance |
Consider a health system that acquires a hospital on a different EHR. In a system-aligned design, that is a new domain, new pipelines and new definitions for every downstream report. In a capability-aligned design, the new hospital's data is onboarded into the Clinical Care domain and published through the same data products. The dashboards, models and executives downstream should barely notice.
Systems change. Capabilities persist.
The lakehouse certifies what happened
The lakehouse is the enterprise system of record for analytical data. Inside each data domain, data is refined in three layers, and each layer has one job.
| Layer | Purpose | Rule |
|---|---|---|
| Bronze | Immutable ingestion: the record of source | No business logic. Data lands as it arrived, so it can always be replayed. |
| Silver | Standardization and conformance: the record of trusted data | Enterprise identifiers resolved here: patient, provider, location, payer. Sources integrated once. |
| Gold | Certified data products: the record of certified data | Facts only. Owned, contracted, tested and versioned. |
The modeling pattern should fit the estate. What does not change is the boundary at the top of the stack.
Gold describes what happened. It does not say what it means.
Gold tables must never contain:
- Business KPIs, such as margin or readmission rate
- Executive logic, such as targets or thresholds
- Cross-domain conclusions, or aggregations built for one dashboard
This is the red line that prevents metric drift. The moment a readmission flag is computed inside an encounters table, the enterprise has two definitions of readmission. They will agree on the day they are built. They will not agree a year later, and nobody will know which one the AI model was trained on.
The semantic layer is the enterprise's translation engine
The semantic and metrics layer is where the enterprise decides what certified facts mean. It is the most strategically important layer in the model, and the one most often treated as a feature of a BI tool.
Business logic changes far faster than data. The definition of an avoidable readmission or the fiscal calendar after a merger change while the underlying encounters do not. Put that logic in the data and every change is a data migration. Put it in the semantic layer and every change is a governed, versioned release.
Stability in data. Flexibility in meaning.
It is the only place business meaning is assigned, and to serve that role it has to be headless:
- Independent of any BI tool, so changing tools does not change the numbers
- Versioned and tested, like any production software
- Served to everything: BI, APIs, applications and AI
One metric, one registry
At the heart of the layer is a metric registry. Every enterprise metric exists in it exactly once, owned by a named business domain and certified before anyone may use it in an executive setting. Each entry reads less like a formula than a contract: who owns the metric, what it means, the certified facts it is built from, its version, the tests it must pass and who consumes it.
Business domains own decisions
Business domains, such as Finance, Revenue Cycle or Clinical Quality, sit on top of the semantic layer. Each consumes the data products of several data domains, owns the KPIs and decision logic for its part of the business and is measured on outcomes, not pipelines.
The critical word is owns. A business leader is accountable for what the metrics mean and the decisions they drive. This is a business operating model, not an IT program with a sponsor's name on the slide.
One direction of flow
The model holds only if every layer consumes from the layer below it.
- Source systemsEHR, ERP, claims, HRMS, CRM, devices
- Capability-aligned data domainsBronze → Silver → Gold: certified facts
- Semantic & metrics layerPublished meaning: every metric once
- Business domainsDecisions, owned by the business
- BI · AI · APIs · applicationsConsumers of truth, never creators of it
No business domain consumes physical data directly.
The rule sounds bureaucratic until you see what it prevents. The day a dashboard calculates its own margin from a Gold table, the enterprise has a second definition of margin, created one reasonable shortcut at a time.
A worked example: total cost of care
Total cost of care draws on utilization facts from Clinical Care, financial facts from Financial Transactions and pharmacy cost facts from Supply Chain. The composite KPI is defined once, in the semantic layer, and Finance owns the decisions built on it. The CFO's dashboard, the population health report and the AI assistant from the opening scene all get the same answer, with its owner, version and lineage.
Cross-domain metrics are where executives live
Most of the decisions that matter to an executive team span domains. Margin draws on Finance, Workforce and Supply Chain. Quality draws on Clinical, Operations and Finance. This is where operating models break: nobody is sure who owns a number that belongs to everyone.
Correct pattern
- Composite logic defined once, in the semantic layer
- Owned by a named executive, such as the CFO for margin
- Versioned, certified and traceable to every contributing fact
Forbidden pattern
- An "enterprise KPI" table in the lakehouse
- Cross-domain joins rebuilt inside each dashboard
- A composite metric with no single owner
Governance without bottlenecks
Centralized governance and centralized delivery are different things; conflating them is how governance becomes a queue.
| Role | Responsibility |
|---|---|
| Central platform team | Guardrails, tooling, templates, CI/CD, and the shared platform as a service to domains |
| Domain teams | Data products and semantic models, and their quality, documentation and evolution |
| Central governance | Standards, policy and certification of data products and enterprise metrics |
The shift that makes it work is moving governance from meetings into the architecture. Certification is a stage in the release, not a committee date. Policies are enforced by the platform. The central team sets guardrails; it does not stand at the gate reviewing every change.
Healthcare-grade by default
- Access enforced centrally, not re-implemented in each tool
- PHI and PII protected by role and purpose, with minimum-necessary access
- Every access path auditable, from the dashboard back to the source
AI without a second data universe
Enterprises that struggle to scale AI usually have two data estates: a governed one for reporting and an ungoverned one assembled, project by project, for models. The model removes the second.
| AI stage | Consumes |
|---|---|
| Feature engineering | Silver and Gold data products |
| Training labels | Certified semantic metrics |
| Inference and decisions | Certified KPIs and thresholds |
| Monitoring | The same metric registry the business uses |
Training labels matter most. A readmission model trained on its own definition of readmission predicts something the quality team does not measure. Train it on the certified metric and model performance and business performance become the same conversation.
The rule applies with more force to generative AI and agents. An assistant asked for last quarter's readmission rate should call the metric registry, not write its own SQL against raw tables. Grounding AI in published meaning turns a confident answer into a correct one.
AI consumes enterprise meaning. It does not reinvent it.
The domain lifecycle
- Register. Name the capability and owner before anything is built.
- Build. Bronze to Silver to Gold, on platform templates.
- Certify. Data products and semantic models pass the same automated gate.
- Consume. Through BI, AI, APIs and applications, never around them.
- Monitor. Quality, service levels and usage, in production.
- Evolve or retire. Version openly; retire what nobody uses.
How this goes wrong
- Mesh by memo. Domains declared on an org chart, with no platform or funding behind them.
- KPIs creep into Gold. A margin column added "just this once" for a board deadline, never removed.
- The semantic layer lives inside one BI tool. Meaning locked in a tool that AI and the next BI tool cannot reuse.
- Domains are drawn around systems. Every migration becomes a reorganization.
- Governance acts as a gate. Delivery slows, and shadow data quietly returns.
- AI builds its own features from extracts. A second, ungoverned data universe grows.
- Metrics without owners. Definitions nobody can certify, change or retire.
The red lines
The rules I ask executive sponsors to hold. Each prevents a failure described above.
Lakehouse
- No KPIs
- No business meaning
- No cross-domain logic
Semantic layer
- One metric, one definition
- No logic embedded in BI
Governance
- No data outside governance
- No asset without an owner
BI and AI
- BI consumes truth; it does not create it
- AI uses the same metrics as analytics
What success looks like
The model shows up in behavior, not technology. Watch for:
| Signal | How to measure it |
|---|---|
| Fewer reconciliation meetings | Hours a month spent making numbers agree |
| Faster time to insight | Days from a new question to a certified answer |
| Higher reuse | Consumers per data product; share of new use cases built on existing products |
| Growing trust | Share of executive metrics certified, owned and traceable |
| AI on the same foundation | Share of model features and labels drawn from governed products and metrics |
When leaders stop asking “whose number is right?”, the model is working.
Where to start: one decision, ninety days
No enterprise adopts this model in one program. Start with one decision that matters, and prove the model on it.
- Days 1–30. Choose one decision, such as total cost of care. Name its business owner. Inventory the metrics behind it and the data domains they draw on.
- Days 31–60. Certify the facts those metrics need. Define each metric once, with its owner. Retire the duplicate extracts.
- Days 61–90. Move the executive dashboard and one AI use case onto the certified metrics. Measure reconciliation time before and after.
Then do it again with the next decision, reusing the domains and metrics of the last.
The mandate: clarity of responsibility
Strip away the architecture and four sentences remain:
- Data domains own facts.
- Business domains own meaning, published once in the semantic layer.
- Business leaders own the decisions made on it.
- Central teams enforce guardrails, not gates.
Get those four right and the technology choices become far less dramatic. Get them wrong and no platform will save you. Done well, the model delivers autonomy without anarchy, governance without bottlenecks, and scale without chaos.