Prepared for Andrew Fritch & the NYX leadership team

Separate the ERP from the AI.
Unify the data.

A pragmatic IBM watsonx, Fusion lakehouse, and Planning Analytics roadmap — designed around NYX's AS/400 reality, mid-market budget, and zero-tolerance for AI breaking the business.

5,000
Employees
~$1B
Annual revenue
27
Executive KPIs today
0
AI-caused ERP incidents (target)
Why now

The window is short, and the AS/400 is non-negotiable.

Andrew said it plainly on April 20: push data out to a lake where AI can ingest it; do not let agents make changes directly in the ERP. This site is built around that principle.

The mid-market squeeze is real

Tier-1 automotive suppliers are caught between OEM price pressure and rising input costs. Every dollar spent on AI has to earn its keep — there is no room for science projects.

Agentic AI is moving faster than guardrails

An incoming AI lead is building on Azure. Without a governed data layer, agents will reach directly into operational systems — exactly the failure mode NYX wants to prevent.

DB2 on IBM i is fast — but it's not the AI substrate

Direct queries work for dashboards. They don't work for governed model training, lineage, multi-source joins with Sterling, Excel, SOWs, and Azure-side data, or write-back review queues.

The owner asked for AI at scale

Not just an LLM chat box. That requires a single, governed copy of NYX's data — ERP, payroll, EDI, engineering, and estimating — with policy, lineage, and security baked in.

Current state

What NYX runs today.

An honest inventory. The goal is not to replace what works — it is to give the next wave of AI a safe, unified place to land.

System of record
IBM i (AS/400) — DB2
ERP + payroll system of record. Fast, reliable, and the heart of NYX operations.
B2B integration
IBM Sterling Commerce
B2B / EDI integration with OEM customers and tier suppliers.
Reporting
ASP.NET + SQL Server
Executive dashboard prep layer. 27 KPIs at daily, weekly, and monthly grain.
Shadow IT
Excel — Estimating
Estimators bypass the ERP estimating module. Knowledge lives in spreadsheets.
Unstructured
Engineering documents
900-page SOWs, mold designs, change orders, scanned drawings. Heterogeneous PDFs with no parsing pipeline — today, every read is a human read.
AI in flight
Claude + GitHub Copilot
Already in use for SOW comparison and software development. Ungoverned.
Net-new
Azure (incoming AI lead)
Net-new agentic and LLM workloads being prototyped outside the IBM stack.
Target architecture

One governed lakehouse. Many engines. Zero direct agent writes to the ERP.

Every layer was chosen against a specific NYX constraint: open formats to avoid lock-in, governance as a first-class citizen, and a hard boundary between AI and the AS/400.

01
Sources
IBM i / DB2
ERP + payroll
Sterling Commerce
EDI / B2B
SQL Server
Dashboards
Excel & SharePoint
Estimating
Engineering PDFs / SOWs
Unstructured
Azure data
AI prototypes
02
Ingestion, streaming & document parsing
watsonx.data integration
Batch + change data capture
Confluent (Kafka)
Event streaming from DB2
Unstructured.io
Parses 900-page SOWs, mold specs, change orders into structured chunks
03
Lakehouse on Fusion
watsonx.data
Single open-format copy
IBM Fusion
Storage + compute fabric
Apache Iceberg / Parquet
Open formats, no lock-in
04
Governance & quality
watsonx.governance
Model risk + policy
watsonx.data intelligence
Catalog + lineage
AI security
Prompt-injection, PII, IP
05
AI & orchestration
watsonx.orchestrate
Agentic workflows
GPT (OpenAI)
3rd-party model
Claude (Anthropic)
3rd-party model
IBM Granite
Where IBM models fit
06
Consumption
IBM Planning Analytics
Sales forecasting, S&OP
Executive dashboards
Replaces SQL Server prep
Estimating copilot
Excel-friendly
ERP write-back queue
Human-in-the-loop only
Use cases

Six concrete bets — sized for a $1B manufacturer.

Each one solves a problem Andrew described. None of them require ripping out the ERP or betting the company on a single model vendor.

Executive KPI dashboard, rebuilt on the lakehouse

The 27-KPI executive board moves off the ASP.NET + SQL Server prep layer onto governed lakehouse views. Same daily / weekly / monthly grain — with lineage back to DB2.

Faster refresh, single source of truth

Sales forecasting in IBM Planning Analytics

Connect Planning Analytics to lakehouse-curated order, shipment, and EDI data. Drive S&OP, demand planning, and capacity decisions from a model that learns from history.

Replaces spreadsheet forecasts

SOW & RFQ change detection

Unstructured.io parses 900-page SOWs into clean, chunked sections; watsonx.orchestrate routes the diff to Claude for change detection — versioned, auditable, and tied to engineering review workflows.

Hours saved per 900-page SOW

Estimating copilot for Excel users

Estimators keep Excel. A copilot reads historical job costs from the lakehouse and suggests pricing — meeting people where they are instead of forcing a tool change.

Adoption without disruption

Engineering knowledge mining

Mold designs, change orders, and SOWs flow through Unstructured.io into the lakehouse — indexed, embedded, and searchable. Mold-flow analysis is a multi-year build; we set the foundation now, not the expectations.

Compounds over years

ERP write-back via human-approved batches

When AI suggests a record change, it lands in a review queue — never in DB2 directly. Approved batches flow back through controlled jobs, preserving the AS/400's stability.

Zero direct agent writes
Safety & governance

An air gap between the agents and the AS/400.

The principle Andrew stated — push out, never reach in — rendered as an architecture rule, not a hope.

The boundary

  • Inbound only. CDC streams from DB2 push into the lakehouse. Agents never query the AS/400 directly.
  • Review queue, not write-back. Every agent-proposed change to ERP records lands in a human-approved batch.
  • Policy at the gateway. watsonx.governance enforces what models can be called, on what data, by whom.
  • AI security. Prompt-injection, PII, and IP-exfiltration guardrails on every model call — GPT, Claude, or Granite.

Visualizing the air gap

System of record
IBM i / DB2 — read-only to AI
CDC stream ↓
Governed plane
watsonx.data on Fusion + governance
Model calls ↓
AI plane
Orchestrate + GPT / Claude / Granite
Proposed writes ↓
Human-in-the-loop
Review queue → controlled ERP batch jobs
Phased roadmap

Twelve months, three phases, one governed substrate.

Each phase delivers value on its own. None of them require ripping out the IBM i, Sterling, or the work the AI lead is already doing on Azure.

Phase 1
0 – 90 days

Foundation

  • Stand up watsonx.data on IBM Fusion
  • CDC pipelines from DB2 + Sterling via Confluent
  • Unstructured.io pipeline for SOWs and engineering PDFs into the lakehouse
  • watsonx.governance baseline: catalog, lineage, policies
  • Rebuild the executive 27-KPI dashboard on lakehouse views
  • AI security guardrails for the existing Claude / Copilot usage
Phase 2
3 – 6 months

Forecasting & governed AI

  • IBM Planning Analytics live for sales forecasting and S&OP
  • watsonx.orchestrate pilot: SOW comparison productized
  • Multi-model routing: GPT, Claude, Granite via the gateway
  • watsonx.data intelligence rolled out across business domains
  • First write-back review queue (low-risk records)
Phase 3
6 – 12 months

Scale & agents

  • Estimating copilot in production, Excel-native
  • Human-in-the-loop agentic workflows for production scheduling
  • Engineering knowledge graph fed by Unstructured.io across SOWs and design docs
  • Foundation for mold-flow ML (multi-year program)
  • Joint operating model with the Azure-side AI team
Success metrics

What we'll measure — and what we refuse to compromise on.

Soft-ROI ranges below are directional estimates for a $1B manufacturer; we'll firm them up with NYX finance during the workshop.

< 5 min
Executive dashboard latency
From overnight batch to near-real-time.
$150–250K / yr — analyst time + faster decisions
+15–25%
Forecast accuracy lift
Planning Analytics vs. spreadsheet baseline.
$2–5M / yr — inventory carry, expedites, lost sales
60–80%
SOW review hours saved
Per 900-page document, governed pipeline.
$400–700K / yr — engineering capacity reclaimed
100%
AI actions governed
All model calls policy-checked and logged.
Avoided cost — IP leak, audit findings, rework
0
AI-caused ERP incidents
The hard line. Non-negotiable.
Avoided cost — a single ERP outage runs $250K+/day
3
Production use cases by Q4
Dashboard, forecasting, SOW analysis.
$3–6M / yr — combined run-rate impact
Why IBM, not just Azure

The Azure work continues. The data layer should be IBM's.

This isn't a vendor war. The AI lead's Azure prototypes can — and should — read from the same governed lakehouse. The question is where the substrate lives.

Dimension
IBM watsonx + Fusion
Azure-only
AS/400 integration
Native — IBM i, DB2, Sterling are first-class citizens.
Possible, but always a connector exercise. NYX wears the integration risk.
Data format lock-in
Open: Iceberg / Parquet on Fusion. Query with any engine.
Tendency toward Delta + Synapse-specific stacks.
Governance & AI security
watsonx.governance + AI security as first-class products.
Available, but assembled across multiple Azure services.
Model choice
Orchestrate calls GPT, Claude, Granite — NYX picks per use case.
Optimized for OpenAI; other models possible but not the default path.
Operational posture
On-prem / hybrid via Fusion — fits NYX's manufacturing footprint.
Cloud-first. Good for greenfield AI; less natural for ERP-adjacent workloads.
Best-of-breed where it counts
Pragmatic, not dogmatic — Unstructured.io for document parsing, Confluent for streaming, GPT and Claude alongside Granite.
Tends to pull every workload into the Microsoft stack by default.
Next steps

A two-and-a-half hour working session — then a Phase 1 SOW.

No long sales cycle. The April 20 conversation already established intent. The next meeting designs Phase 1.

Proposed workshop agenda

  • 0:00
    Recap of April 20 conversation & confirm guiding principles
  • 0:20
    Deep-dive: lakehouse on Fusion — sizing, deployment model, Iceberg
  • 0:50
    Governance & AI security walkthrough on NYX-shaped data
  • 1:20
    Planning Analytics demo against synthetic NYX shipment data
  • 1:50
    Joint plan with the Azure-side AI lead — interfaces, not turf
  • 2:15
    Phase 1 statement of work, cost envelope, success criteria

Stakeholders

  • Andrew Fritch
    NYX — IT leadership, sponsor
  • NYX AI lead
    NYX — Azure-side AI program
  • NYX estimating lead
    NYX — pilot owner for estimating copilot
  • Mangat Rai
    IBM — watsonx.data
  • Eyal Abraham
    IBM — watsonx.governance & AI
  • Chris McCool
    Partner — program lead
  • Jack Kaden
    Partner — engagement coordination

Glossary

watsonx.data
IBM's open lakehouse engine. Single copy of data, multiple query engines.
IBM Fusion
Storage and compute fabric watsonx.data runs on. Hybrid-friendly.
Apache Iceberg
Open table format. Lets Spark, Presto, and others share one dataset.
CDC
Change data capture — streams row-level changes out of DB2 in near-real-time.
watsonx.orchestrate
Builds and runs AI agents and workflows across systems and models.
Granite
IBM's family of foundation models. Used alongside GPT and Claude where it fits.

Prepared for NYX Inc. · Based on the April 20, 2026 working session with IBM and partner teams.