Leaser architecture / Field note 004
A shared language from lead to NOI.
Enterprise AI becomes reliable when every system agrees on what the business means.
AI does not reason from data alone. It reasons from the meanings an organization has attached to that data.
If “lease,” “exposure,” “qualified lead,” or “NOI impact” means something different in the dashboard, the recommendation engine, the owner report, and the agent’s context, there is no closed intelligence loop. There are competing versions of the business.
A shared semantic layer resolves that conflict. It gives every human and machine participant the same entities, definitions, grains, time windows, evidence labels, and lineage—from the first demand signal to the economic outcome.
The model can change.
The interface can change.
The meaning cannot drift.
Most enterprise AI errors begin before the prompt.
A language model can generate flawless SQL against a flawed business definition. It can retrieve five documents that use the same word differently. It can explain a metric with confidence while querying the wrong grain.
More context does not reconcile contradictory meanings. Retrieval does not decide which source is authoritative. Text-to-SQL does not know whether “lease” refers to an ad-platform conversion, an application event, a guest-card stage, or a signed record in the property-management system.
This is why the semantic layer is not merely analytics infrastructure. It is the trust boundary between probabilistic reasoning and enterprise reality.
A real architecture lesson
Two correct queries. Two different businesses.
Before Leaser centralized key metrics, services independently computed cost per lease. One used the floorplan target when no leases existed. Another returned zero. Both were internally coherent. Together, they were operationally dangerous.
Target CPL used as the neutral baseline
No signed leases interpreted as zero cost
Which number should the agent act on?
Leaser moved the definition into canonical database views. Every service now reads the same exposure, occupancy, and CPL logic instead of recomputing it. The fix was not a smarter model. It was one meaning.
A metric is a compressed decision policy.
A metric tells a system what counts, what does not, where to aggregate, how far back to look, and what to do when reality is incomplete. A name and formula are only the beginning.
What real-world thing does this describe?
Which canonical key makes records the same thing?
Property, floorplan, campaign, prospect, or portfolio?
Which inputs, filters, and fallback behavior apply?
Event time, snapshot date, attribution window, and timezone?
Measured, sparse, modeled, or assumed?
Which source produced it, and how fresh is it?
Who can change the definition and approve a new version?
The protected join
Property is the bridge between marketing and economics.
Ad accounts, websites, CRM records, ILS leads, call tracking, and Yardi do not arrive with a universal identity. Leaser resolves them to a canonical property—and protects that key because it is where demand activity becomes an asset outcome.
The shared spine
From lead to NOI without changing the meaning.
- Demand signalCampaign, click, call, ILS inquiry
Source fidelity and a declared capture surface.
- Prospect identityLead and guest card
Tokenized identity with match confidence.
- Leasing progressionTour and application
Canonical funnel stages at a defined event time.
- Outcome truthSigned lease
A PMS record—not an ad-platform conversion.
- Asset stateExposure and occupancy
Computed once at property and floorplan grain.
- Economic valueRevenue and NOI impact
The same outcome expressed in owner terms.
The same metric at the wrong grain becomes a different decision.
A portfolio can appear healthy while a particular floorplan accumulates exposure. A property-level cost per signed lease can hide one channel that creates demand and another that merely captures it. A monthly average can obscure a current pacing failure.
Leaser treats floorplan as the atomic operating grain because that is where inventory pressure becomes actionable. Property is the asset and access-control grain. Portfolio is the learning and capital-allocation rollup. These levels relate, but they are not interchangeable.
The semantic layer makes valid dimensional cuts explicit. It prevents an agent from silently mixing a campaign-day numerator with a portfolio-month denominator and returning a number that looks precise but means nothing.
Define once
One governed metric.
Grain: property or floorplan
Source: canonical unit state
Evidence: measured
Missing data: visible, never silently zero
Consume everywhere
- Operator dashboard
- Decision engine
- Lea analyst
- Revenue forecast
- Action guardrails
- Owner report
Probabilistic intelligence needs deterministic ground.
Generative models are useful precisely because they can interpret ambiguity. Business metrics should not ask them to.
Leaser constrains analytical answers to defined metrics instead of allowing free-form hallucinated SQL. Recommendations consume canonical views. Reports compose the same service outputs rather than recreating formulas. The AI may explain why exposure matters or decide which intervention deserves attention; it does not get to invent a new definition of exposure.
This pattern is increasingly recognized across the data ecosystem. Google describes Looker’s semantic model as a way to define metrics once and use them everywhere. Microsoft’s guidance for AI over Power BI warns that unprepared semantic models can produce low-quality or misleading outputs. Snowflake’s semantic views make facts, dimensions, metrics, descriptions, synonyms, and non-additive behavior explicit in the data contract. The differentiator is not having a semantic layer; it is encoding the right domain semantics and connecting them to action and outcomes.
Meaning includes uncertainty
A number without evidence class is an unfinished sentence.
Real data with adequate sample
Sparse signal, partially pooled
Validated model with an interval
Rule-derived planning input
Leaser classifies a metric by how it was produced—not by how convincing it looks. A signed-lease metric with a small sample remains estimated. A forecast never masquerades as an observation. Evidence travels with the value into decisions, reports, and autonomy thresholds.
Freshness and failure are part of the definition.
A perfectly defined metric built on a broken connector is still wrong. Semantic contracts must include freshness expectations, late-arriving-data behavior, reconciliation tolerances, and what the product does when an input is missing.
Silently replacing missing data with zero changes absence into performance. Carrying forward a stale value turns yesterday into today. Leaser’s staging and mart models test canonical keys, accepted channel values, required fields, uniqueness, and source freshness so uncertainty becomes visible before a system acts on it.
Architecture, not middleware
The language sits in the path of every decision.
The foundation is real. The dictionary is still expanding.
Leaser’s canonical database views already own property and floorplan occupancy, exposure, and CPL behavior. dbt staging models test source contracts; marts express cost per signed lease, channel contribution, the leasing funnel, property state, and preliminary NOI impact. Portfolio reports compose these shared outputs so they agree with the operating surfaces.
The full semantic dictionary is not finished. Forward exposure still needs richer lease-term and renewal data. Multi-touch attribution requires a touchpoint model. Media revenue in the current NOI mart remains a placeholder. Tenant isolation and row-level permissions need to mature with enterprise deployment.
The architecture is designed so those additions extend one contract rather than create new islands of meaning.
The ecosystem advantage
Shared language makes partners composable.
A property system can contribute canonical lease outcomes. A demand partner can contribute touchpoints. A model can produce a forecast with an evidence class. An action provider can execute a governed intervention. Each component becomes more useful because it joins the same decision and outcome spine.
This is how an ecosystem compounds without becoming an integration maze: partners do not merely exchange payloads. They participate in a common operating model.
The semantic layer does more than make numbers agree.
It makes experience reusable.
When every decision carries the same entities, metric definitions, evidence, action record, and outcome, the system can compare like with like. A lesson learned at one property can transfer cautiously to another because the conditions mean the same thing. That is the foundation of portfolio-scale Decision Intelligence.
Before an enterprise can automate judgment, it must define reality.