Three tables. Three answers. Which is correct?

"What was our rooms revenue in July?" P&L (Finance) $12.8M Authoritative total includes all room revenue Room Segments (Ops) $11.9M Breakdown by segment missing segments outside ops domain Reservations (ORS) $12.1M Transaction-level detail missing adjustments outside reservations domain $0.9M drift All three pass every data quality check. An AI picking the wrong one doesn't know the difference.

BI reporting stays in lane. AI has to cross but they don't add up.

Traditional BI

P&L Segments ORS Room rev + other rev + adjustments + tax Monthly, all line items Corporate, Leisure, Group, OTA segments Monthly, ops categories Daily transactional booking records Daily, pre-adjustment SUM SUM SUM Total $12.8M Total $11.9M Total $12.1M Finance Dashboard $12.8M Ops Dashboard $11.9M Revenue Dashboard $12.1M F Sees $12.8M O Sees $11.9M R Sees $12.1M Each dashboard stays in its lane. Each total is correct — for its domain.

AI Analyst

P&L Segments ORS Room Revenue $12.8M By Segment $11.9M Bookings $12.1M "Break down total room revenue by segment" total breakdown AI assembles across sources: Total (P&L): $12.8M Segments: $4.2M + $3.5M + $2.3M + $1.9M = $11.9M ≠ Segments sum to $11.9M but P&L says $12.8M The AI has no way to reconcile them. Which number should it present to leadership?

The map has to be prescriptive, not just declarative.

"Break down total room revenue by segment" AI retrieves from two sources: P&L: $12.8M ≠ Segments: $11.9M Declarative only (catalog / ontology) The AI knows: • "P&L table contains room revenue" • "Segments table contains segment breakdown" • "Both relate to the same property" The AI doesn't know: • Which number is authoritative • Why they differ by $0.9M • Whether the gap is expected or broken • How to present the answer Picks one, guesses, or hallucinates the reason Prescriptive (semantic integrity layer) The AI reads the rule: RECONCILIATION RULE source: P&L (authoritative) segments = P&L − staff bookings expected ∆: ~$0.9M | threshold: <5% last validated: 2026-03-31 | status: ✓ within bounds The AI responds: "Total room revenue is $12.8M (P&L). Segments sum to $11.9M — $0.9M gap is staff bookings not captured in the segment report."

A semantic layer turns disconnected tables into a navigable map.

Without

Lakehouse ERP source ORS source PMS source P&L Segments ORS Bookings Room Types ETL lineage, schema validation, freshness checks, row counts Relationships are technical. No business relationships between tables. AI guesses how to link them.

With semantic layer

P&L Room Revenue $12.8M authoritative = P&L minus staff bookings sum validated ✓ (∆ = staff $0.9M) Room Segments Corp $4.2M Leis $3.5M Grp $2.3M OTA $1.9M = $11.9M maps to P&L via adjustments variance checked (<5%), stored recalculated on each data ingestion ORS Daily Bookings $12.1M pre-adjustment, pre-tax

What the Semantic Integrity Layer does

1 Cross-validate Total: $12.8M $4.2M $3.5M $2.3M + ... ∑ = $11.9M ✓ ∆ $0.9M accounted for Every property Every month Numbers add up. Variances are explained. 2 Maintain continuously ✓ Apply corrections Scale, adjust, reconcile ■ Store variances Track drift over time ! Flag issues Stale data, anomalies, breaks On every data ingestion. Automated, no manual work. Always current. 3 Guide the AI Data map Authority, hierarchy, known variances Business glossary RevPAR = Room Rev / Available Rooms ADR, Occ%, TRevPAR, GOPPAR Semantic understanding Fiscal periods, seasonality, currency AI navigates the map, not the raw tables.

Conventional systems vs an integrated approach

Conventional

Data Engineering Schema, nulls, freshness, dedup Data is clean ✓ No business checks across tables Metadata Catalog Describes tables, columns, lineage Documentation exists ✓ Static. No enforcement. No variance rules. suggestions only AI copilot Has access to catalog and data Treats metadata as a suggestion Can ignore prescribed paths, pick its own tables Each layer works on its own. Nothing guarantees they work together.

Integrated

Data Engineering Schema, freshness + business validation Cross-table checks on every ingestion Maintains the semantic layer continuously Semantic Integrity Layer Active rules, prescribed paths, stored variances Business glossary, data map, authority hierarchy Not documentation — enforcement must go through AI Agent Architecturally bound to the semantic layer Every query reconciled against the map Can't bypass — the layer is the gate One integrated system. The AI navigates the map, not the raw tables.

Generic copilot vs our AI

"How did our RevPAR compare across regions last quarter, and which properties are underperforming vs budget?"

Generic copilot

Scans all available tables Query Engine Ad-hoc pattern matching, keyword similarity Provenance Tracking Generic: "Based on..." / "I found..." No guaranteed link to authoritative source Result • "RevPAR" → averages property RevPARs (wrong) • "region" → picks whichever table has a region field • "last quarter" → calendar Q, not fiscal quarter • "budget" → may pull forecast instead of budget ✕ Closest match ≠ correct match

Our AI

Starts from semantic map Semantic Resolution "RevPAR" → room_revenue_net / available_rooms "region" → property.region via portfolio hierarchy "last quarter" → fiscal Q (Oct-Dec), "budget" → approved_budget table Query Engine Validated SQL from known tables, known joins, known granularity Provenance Tracking Source: P&L authoritative → table.column → row-level citation Result — grounded, auditable Correct metric, correct source, correct granularity ✓ Defaults to the map, not the guess.

Generic copilot vs our AI

"What was our group total revenue last quarter, and how did each region contribute?"

Generic copilot

Scans all available tables Finds regional entity P&Ls Thailand: $14.2M + Japan: $9.8M + Vietnam: $6.5M Also finds Group Consolidated report Sums regional revenues $14.2M + $9.8M + $6.5M = $30.5M Does not know about intercompany transactions Result • "Group revenue is $30.5M" (double-counts interco) • Includes $1.8M shared services + mgmt fees • Thailand HQ charges IT & finance to Japan/Vietnam • Group Consol says $28.7M — copilot doesn't reconcile ✕ Sum of parts ≠ consolidated total

Our AI

Starts from semantic map Semantic Resolution "group total" → Group Consolidated P&L (authoritative) "regional breakdown" → entity P&Ls minus interco eliminations Interco mapped: Thailand HQ shared services, mgmt fees Rule: SUM(regions) - interco = Group Consol ✓ Query Engine Group total from consol; regional split with interco backed out Reconciliation $14.2M + $9.8M + $6.5M - $1.8M interco = $28.7M ✓ matches Group Consol Result — reconciled, auditable $28.7M group total. Regional split nets out interco. Numbers tie. ✓ Parts reconcile to the whole.

Architecture — fully isolated per client

Client Environment Your Data Warehouse Read raw data Write back clean model Dedicated Environment Fully isolated per client Data Engineering AI Analyst Stack Semantic Integrity Layer Hierarchy + mappings + validation Source authority rules Query Engine Guardrails continuously maintains semantic layer Application Layer LLM Provider Model-agnostic Zero data retention API calls only No customer data stored

Layers between AI and data

"What was our revenue last month?" Guardrails Is the question allowed? Does the user have access? Is the scope within bounds? Semantic Layer "Revenue" → room_revenue_net Resolves ambiguity. Enforces metric definitions. Maps to authoritative source. Query Engine Generates validated SQL. Only permitted tables & columns. Parameterized, injection-proof. AI-Optimized Data Model Grounding, Provenance & Quality Every answer traced to a real query result. Citations linked to source data. Compliance checks: did the AI follow the semantic map? Live evals: continuous quality measurement on real answers.

How it fits — nothing changes

Your Existing Warehouse Fabric / Power BI Lakehouse P&L Segments ORS / Dashboard reads from Semantic Integrity Layer ≈ Reconcile ⚙ Auto-fix ⚑ Quality flags ★ Authority rules verified data + context AI Engine Query engine, guardrails, semantic resolution answers + citations Application Layer / Power BI Consistent numbers across every view and answer

Additive, not a replacement.

Your warehouse stays as-is.

Your Power BI dashboards stay as-is.

The layer reads, validates, and serves.

New data sources can be ingested and automatically flow through to the semantic layer.

Still additive — no disruption to existing pipelines.