Modernization Without Meltdown
The following blog is a summary of our discussion on the podcast, produced with the assistance of AI.
Prefer to watch?
Fraser Paine, Head of AI at HSO, and CTO Jourdan Templeton share how to build a modern, AI-ready data platform without disrupting the business.
Learn how to:
Separate data and AI platforms for clarity and control
Turn data into an AI engine through governance and consistency
Use generative AI to clean and enrich unstructured data
Modernize without a big-bang rewrite - and make this your last migration
Due to your cookie preferences, we cannot show you this content. Accept the cookies to be able to view this content.
The Big Idea: Separate - but Connected - Platforms
Most enterprises have invested in a single “data platform” and assume it’s ready for AI. We argue for two complementary platforms:
Data Platform: the governed backbone for ingesting structured and unstructured sources (databases, APIs, files, transcripts, images), shaping them into consistent, documented models, and publishing to consumers (analytics, apps, services).
AI Platform: the operational layer for models, prompts, agents, evaluation, cost controls, deployment standards, security, and guardrails.
Treating these as one often leads to inconsistent results, runaway bills, and compliance risk. Treating them as partners creates speed with safety.
Why It’s Hard (and Expensive) - and How to Fix It
At first glance, “just store the data” feels simple. In reality, enterprise data is:
Voluminous & sensitive (PII, commercial IP)
Heterogeneous (tables + PDFs + audio + images)
Regulated (policies, retention, access)
A modern data platform solves this with standardized ingestion, uniform structuring, and consistent publishing—all documented in a data catalog so humans and agents know what data means. With consistency in place, every downstream consumer (BI, agents, apps) gets more accurate and more reliable.
Generative AI Belongs Inside the Pipeline
AI isn’t just a chatbot at the edge. Embedding GenAI in ETL/ELT is a force-multiplier:
Clean & normalize: unify dates, IDs, phone formats, product names
Enrich: classify sentiment and intent in call transcripts; extract entities from free-text reviews; summarize long documents into structured fields
Multimodal: work directly from audio or images stored in the lakehouse
Result: unstructured data becomes analytics- and agent-ready with far less manual effort.
Trusted Self-Service: Agents Over Cataloged Data
With well-described datasets (via a catalog and business glossary), AI agents can safely:
Generate queries on trusted, governed tables
Perform iterative analysis (“why did we lose these opportunities?”)
Feed results into next-best actions for sales, service, or operations
Leaders get faster answers to follow-on questions without spinning up new dashboards for every scenario.
Sync on-prem SQL to the cloud lakehouse to unlock modern AI/data tooling—without migrating every gnarly business rule on day one.
Move to a building-block architecture so you can upgrade components (e.g., file formats, compute engines) without re-platforming.
Adopt opinionated reference architectures for data & AI to standardize security, cost, identity, and deployment from the start.
Start Small. Prove Value. Scale with Confidence.
Don’t throw “all the data” at AI. Instead:
Pick one question and a contained dataset
Apply AI over the trusted data to answer it
Measure adoption and impact, then expand
This builds capability inside the team, wins stakeholder trust, and creates a repeatable pattern for scale.
Why This Might Be Your Last Migration
Traditional platforms forced you into a painful cycle every five years. A modern, modular architecture lets you evolve in place, swapping parts without rebuilding the whole. That shifts investment from “maintenance” to continuous innovation—crucial as AI and data tooling evolve monthly.
governance, catalogs, lineage, and contracts make AI safer and smarter.
treat models, prompts, eval, and cost controls like a product with SLAs.
enrich unstructured data at ingestion to multiply downstream value.
mirror → modularize → optimize; avoid big-bang rewrites.
start with a focused pilot, measure hard outcomes, and scale patterns that work.
If you’re mapping your next data or AI move - whether mirroring legacy systems to the cloud, embedding GenAI in your pipeline, or enabling agents over governed data - we’d love to compare notes and help you connect architecture to outcomes.