Demonstration environment — synthetic Paddywax operating data. No production data is displayed.
The AI layer sits on top of NetSuite — it never replaces it.
NetSuite stays the single system of record for financials, orders, inventory, procurement, and support. The Paddywax AI layer connects through four sanctioned integration channels, reasons over synced data, enforces guardrails, and writes every action back into NetSuite — including its own audit trail.
System Architecture
Select an integration channel to trace which AI services and NetSuite modules it connects.
Experience Layer
What Paddywax leadership and operators seeCommand Center
Executive briefing, KPI grid, approvals
Agent Workspaces
11 role-based agent consoles
Governance & Audit
Guardrails, activity log, controls
Paddywax AI Layer
Vercel-hosted · never stores data outside the sync boundaryAgent Orchestrator
Routes tasks across the 11-agent fleet (A01–A11)
KPI / Insight Engine
Compiles SuiteQL, materializes briefing metrics
Policy & Guardrail Engine
Confidence thresholds, approval routing, PII rules
Sync Cache (Read Replica)
15-min delta sync of 14 record types
Vector Store / RAG
Embedded NetSuite history, fragrance taxonomy
LLM Gateway
Model routing, prompt logging, cost controls
NetSuite ERP — System of Record
Paddywax Demo Sandbox · Synthetic data · NetSuite 2026.1Financials / GL
transaction · account · posting periods
Order-to-Cash
salesorder · itemfulfillment · invoice · rtnauth
Inventory & WMS
inventoryitem · locations · reorder points
Procure-to-Pay
purchaseorder · vendor · itemreceipt
CRM & Support
customer · supportcase · campaign
SuiteCloud Platform
SuiteScript · SuiteFlow · Saved Searches · custom records
OAuth 2.0 M2M
Client-credentials flow, certificate-based. No user passwords or TBA tokens in the AI layer.
Least-Privilege Role
custrole_pw_ai_layer scopes every call — record-level and field-level permissions enforced by NetSuite.
Full Audit Trail
Every AI write-back lands in custrecord_pw_ai_audit with agent ID, confidence, and before/after values.
Governance Limits
SuiteCloud API governance respected: request throttling, concurrency pooling, and nightly bulk via Connect.
Connection Options Compared
Each channel is chosen for what it does best. Hot paths use REST, analytics use SuiteQL, events use SuiteScript, and heavy history moves through SuiteAnalytics Connect so the API governance budget stays untouched.
| Channel | Best For | Direction | Latency | Volume | Governance Cost |
|---|---|---|---|---|---|
| SuiteTalk REST | Real-time record CRUD | Read + Write | ~200–400ms | Single records | Per-request governance units |
| SuiteQL via REST | Cross-record analytics queries | Read only | ~300–800ms | 1,000 rows/page | Query concurrency pool |
| RESTlets + User Event scripts | Custom logic & event push | Bi-directional, event-driven | <5s from record save | Per-event | SuiteScript usage units |
| SuiteAnalytics Connect | Bulk historical extract | Read only | Nightly batch | Millions of rows | No API governance impact |
What Is Connected Today vs Planned
The MVP scope is NetSuite and its native modules. Other systems referenced in agent workspaces are labeled below so there is no ambiguity about what the initial implementation includes.
| System | Status | Note |
|---|---|---|
| NetSuite ERP (SuiteTalk, SuiteQL, RESTlets, SuiteAnalytics) | Connected in MVP | System of record — all reads and writebacks |
| NetSuite WMS / Support Cases / CRM modules | Connected in MVP | Native NetSuite modules, same connection |
| SuiteCommerce (web + InStore POS) | Connected in MVP | Order, session, and POS data via NetSuite |
| Manhattan Associates WMS (Portland DC) | Connected in MVP | Warehouse system of record — feeds the WMS Agent (A11) reconciliation and network views |
| NL / UK / Australia node feeds (Manhattan + 3PL) | Planned integration | Configuration pending validation — counts advisory only |
| Klaviyo (email/SMS campaigns) | Available via existing API | Campaign sync scoped for phase 2 |
| GA4 / web analytics | Planned integration | Session-level analytics enrichment |
| EDI / retailer portals (SPS, national accounts) | Planned integration | Chargeback and ASN automation |
| All figures shown in this demo | Demonstration data | Synthetic dataset modeled on Paddywax operations |
Designworks Collective — Multi-Brand Path
The agent layer is subsidiary-aware by design. Every query filters on NetSuite subsidiary and channel segments, so onboarding Sweet Water Decor, DesignWorks Ink, or any sibling brand means adding a subsidiary filter and brand-specific KPI targets — not new integrations. The orchestrator, guardrails, audit trail, and approval workflows are shared platform infrastructure across the portfolio.
How a Recommendation Becomes a NetSuite Transaction
- 1
Sync
Every 15 minutes the sync service issues SuiteQL deltas (lastmodifieddate > last cursor) across 14 record types and refreshes the read cache. User Event scripts push urgent changes — a new escalated support case, a PO status flip — in under 5 seconds.
- 2
Reason
The Analytics & BI agent (A05) compiles KPI queries against the cache, joins SuiteAnalytics history for trend context, and drafts the executive briefing. Task-specific agents (A01–A10) pull the records they own — cases, POs, work orders, customers — through the orchestrator.
- 3
Check
Before any write-back, the Policy & Guardrail Engine scores confidence against the scenario threshold. Low-confidence or high-impact actions are diverted into the NetSuite approval queue instead of executing autonomously.
- 4
Act
Approved actions write back through SuiteTalk REST (case updates, sales orders) or the custom RESTlet (custscript_pw_ai_writeback). NetSuite remains the system of record — the AI layer never forks the data.
- 5
Audit
Every action — autonomous or approved — is logged to custrecord_pw_ai_audit inside NetSuite itself with agent ID, confidence, latency, and before/after values. The Governance page reads this record type directly.