Practitioner Intelligence

No vendor whitepapers.
Real patterns from real engagements.

POVs, frameworks, and case studies written by practitioners who have shipped Data and AI in production — not consultants who have only sold it.

Featured POV

Why Your Semantic Layer is the Most Important Investment Before Your First LLM

Every enterprise racing toward GPT-4 wrappers is missing the foundational point. The bottleneck is never the model — it is the absence of a shared business vocabulary that both humans and machines can trust. At its core, a large language model is a statistical reasoning engine. Feed it ambiguous, inconsistently named, semantically disconnected data and it reasons beautifully — about the wrong thing.

The enterprises winning at GenAI in 2026 are not the ones with the best models. They are the ones that built their semantic layer 18 months ago and are now compounding on that foundation.

LLMs hallucinate in proportion to semantic ambiguity — not model size or prompt quality
A unified metrics layer (dbt Semantic Layer or similar) reduces NL-to-SQL error rates by 60–80%
Canonical entity model is prerequisite: one Customer, one Product, not 12 versions across systems
The semantic layer becomes the AI's tool registry — the structured knowledge it can reliably query
Teams that skip this step rebuild it after their first production AI failure — at 3× the cost
Discuss This With The Curator →8 min read · May 2026
// What happens without a semantic layer
"What is our customer churn rate?"
3 different "churn" definitions
"Show Q3 revenue by region"
Revenue field exists in 5 tables, all different
"Which products are at-risk?"
No canonical product entity
"Summarize top client portfolios"
AUM computed differently per system
Result
Confident, wrong answers at scale
Case Studies

Engagements with measurable outcomes

Not anonymized composites — real engagement patterns with real numbers. Company names are withheld by agreement.

Case StudyFinancial ServicesApr 2026

From Data Swamp to RAG-Ready in 12 Weeks

The Problem

A mid-market wealth management firm had invested $2M in an AI-powered client advisor tool. Six months after launch, hallucination rates were at 34% on factual queries about client portfolios. The model was GPT-4. The problem was not the model.

The Diagnosis

Our SIGNAL assessment revealed: Governance score 18/100. No column-level lineage. Metadata coverage under 12%. Three conflicting definitions of "client AUM" across four source systems. The RAG pipeline was retrieving semantically incorrect documents because the underlying data had no shared meaning.

The Engagement

12-week engagement: Weeks 1–3 data inventory and SIGNAL scoring. Weeks 4–7 unified entity model and metadata layer. Weeks 8–10 data contracts and lineage. Weeks 11–12 RAG pipeline rebuild on governed data corpus.

// Results
Hallucination rate
34%6%
-83%
Query accuracy on portfolio data
51%94%
+84%
Time to answer regulatory audit queries
3 days4 hours
-94%

"We spent six months blaming the model. Turns out the model was fine. Our data was not."

Case StudyHealthcare & Life SciencesMar 2026

Unstructured Data Governance for Clinical AI

The Problem

A regional hospital network was building an AI-assisted clinical documentation system. The model performed excellently on benchmark datasets. In production, it generated clinically incorrect summaries 22% of the time — not because of model failure, but because unstructured clinical notes had never been governed.

The Diagnosis

Clinical notes, radiology reports, and discharge summaries existed in 7 different formats across 4 EMR systems. No PII classification. No concept normalization (ICD-10 codes used inconsistently). No lineage between the training corpus and the source documents — making retraining on failure cases impossible.

The Engagement

Implemented AI-era governance for unstructured data: automated PII detection and redaction pipeline, clinical concept normalization layer mapped to ICD-10/SNOMED, column-level lineage from source EMR to training corpus to model output, and data contracts between the clinical data team and the AI engineering team.

// Results
Clinical documentation error rate
22%3.8%
-83%
Regulatory compliance readiness
Not auditableFull audit trail
Model retraining cycle time
6 weeks4 days
-95%

"We could not explain why the model was wrong. After implementing lineage, every error was traceable to a specific data quality issue in 3 minutes."

Points of View

Practitioner positions on contested questions

Not balanced takes. Actual positions, backed by field evidence. Agree or disagree — these are the conversations worth having.

FrameworkGovernance · May 2026 · 6 min

The Data Contract Manifesto for the GenAI Era

Data contracts are not a technical nicety. They are load-bearing walls for AI reliability. A data contract is a formal, versioned agreement between a data producer and a data consumer that specifies: schema, quality thresholds, freshness SLA, ownership, and breaking change policy. In the GenAI era, every RAG corpus, every feature store, every training dataset needs one.

Key insight: Without data contracts, your AI quality degrades silently. A schema change in an upstream system propagates downstream into your RAG index, your model outputs start drifting, and you have no audit trail to diagnose why.

01Schema contract — field names, types, nullable constraints, enumerated values
02Quality contract — completeness %, uniqueness, freshness window, anomaly tolerance
03SLA contract — update frequency, latency guarantees, incident response SLA
04Change contract — deprecation notice period, breaking vs non-breaking change policy
POVArchitecture · Apr 2026 · 7 min

The Modern Data Stack is Dead. Long Live the AI Data Stack.

The Modern Data Stack thesis — Ingest → Transform → Serve — was correct for the analytics era. It assumed the consumer of data was a human analyst running SQL. GenAI changes the consumer to an LLM that reasons in natural language, retrieves by semantic similarity, and fails silently when context is ambiguous. The MDS needs a new layer: Govern → Semanticize → Productize.

Key insight: The four tools that define the AI Data Stack in 2026: dbt (semantic layer), a vector database (retrieval), a data catalog with lineage (governance), and a data contract framework (trust). Everything else is plumbing.

01Replace batch-first with context-first — AI needs freshness windows, not snapshots
02Add semantic layer between transform and serve — LLMs query meaning, not columns
03Govern unstructured data with the same rigor as structured — it feeds your RAG
04Productize data before exposing to AI — APIs, not direct table access
POVStrategy · Mar 2026 · 4 min

Stop Buying AI Tools. Start Buying Data Credibility.

The enterprise AI tool market hit $47B in 2025. Most of that spend will produce no lasting value. The diagnostic question every CDO should ask before the next AI procurement: "Do we have a semantic layer?" If the answer is no, the tool budget is premature. You are buying a racing car before building a road.

Key insight: The organizations compounding on AI in 2026 are not the ones who moved fastest. They are the ones who invested in data credibility — governance, lineage, semantic clarity — while their competitors were buying AI tools.

01Audit: can your AI outputs be traced to source data? If not, stop and build lineage.
02Audit: do you have one authoritative definition of Customer, Product, Revenue? If not, build it.
03Audit: do your data pipelines have quality SLAs? If not, your AI has no trust foundation.
04Only after those three: evaluate AI tools for your specific use case.
Industry Benchmarks

The numbers every CDO should have memorized

73%
Orgs where AI is blocked by data, not model capability
Gartner 2026
57%
CDOs who have not adapted data strategy for GenAI
AWS CDO Study 2025
2.8×
Faster AI deployment with mature data product teams
McKinsey 2025
60–80%
Reduction in hallucinations with proper semantic layer
TheDatapedia 2026
7%
Enterprise orgs that have scaled AI to production
McKinsey 2025
< 11%
Orgs where unstructured data governance is mature
Gartner 2026
Weekly Intelligence

The AI-Ready Data Brief

One email, every Friday. The 3 most important developments in Data and AI that week — with practitioner commentary, not press release summaries.

No spam · Unsubscribe anytime