Enterprise-Scale GenAI and Agentic AI Deployment in B2B Supply Chain Operations

wesco · E-commerce · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: B2B supply-chain and industrial-distribution AI operations

Wesco needed to move enterprise AI from proofs of concept into governed, measurable production systems for internal and external supply-chain workflows, including everyday productivity tools and higher-ROI agentic applications such as accounts-payable fraud detection and inventory management.

Constraints:

Enterprise AI work had to connect to real ERP, CRM, data, and workflow systems rather than remain isolated experiments.
Use cases needed governance, legal/compliance involvement, security, and model-risk controls before scaling.
ROI had to be translated from technical metrics into business KPIs such as productivity, engagement, cost, margin, and time to value.
Autonomous agent decisions needed traceability and auditor-facing explainability.
Human oversight still mattered for fraud decisions and feedback loops while the system matured.

Architecture

Wesco described an enterprise AI platform that layers generative and agentic capabilities on top of mature data and analytics foundations. A governed, air-gapped sandbox supports experimentation; reusable architectures, prompt banks, MLOps/LLMOps standards, observability, and model-management policies move use cases toward production. The detailed fraud example uses a LangGraph-style state graph with planner/executor agents, receipt extraction, entity resolution, anomaly detection, routing to auto-approval or human review, and an investigator agent backed by memory and trace logs.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  erp-and-data-feeds["ERP and data feeds"] --> planner-and-executor-agents["Planner and executor agents"]
  composable-ai-platform["Composable AI platform"] --> planner-and-executor-agents["Planner and executor agents"]
  composable-ai-platform --> llmops-governance-layer["LLMOps governance layer"]
  llmops-governance-layer["LLMOps governance layer"] --> observability-and-audit-tracing["Observability and audit tracing"]
  planner-and-executor-agents --> receipt-extractor["Receipt extractor"]
  planner-and-executor-agents --> entity-resolver["Entity resolver"]
  planner-and-executor-agents --> anomaly-detection-agent["Anomaly detection agent"]
  planner-and-executor-agents --> decision-making-router["Decision-making router"]
  receipt-extractor --> entity-resolver
  entity-resolver --> anomaly-detection-agent
  anomaly-detection-agent --> decision-making-router
  decision-making-router --> human-or-auditor-review["Human or auditor review"]
  decision-making-router --> investigator-agent["Investigator agent"]
  human-or-auditor-review --> investigator-agent
  investigator-agent --> observability-and-audit-tracing

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`ERP and data feeds`	Operational inputs such as invoices, receipts, accounts-payable data, and enterprise source systems.	`Planner and executor agents`
`Composable AI platform`	Governed platform layer with YAML-driven chatbot/agent configuration, model-management policies, and reusable architectures.	`Planner and executor agents`, `LLMOps governance layer`
`LLMOps governance layer`	Guardrails, intake, compliance, prompt banks, model management, monitoring, drift checks, and production promotion controls.	`Observability and audit tracing`
`Planner and executor agents`	State-graph agents that coordinate task execution across specialized fraud-detection agents.	`Receipt extractor`, `Entity resolver`, `Anomaly detection agent`, `Decision-making router`
`Receipt extractor`	Extracts invoice/receipt information with OCR, LLM, RAG, memory, and cache capabilities.	`Entity resolver`
`Entity resolver`	Deduplicates and normalizes extracted entities before anomaly checks.	`Anomaly detection agent`
`Anomaly detection agent`	Combines ML and LLM methods to detect duplicate, anomalous, or potentially fraudulent transactions.	`Decision-making router`
`Decision-making router`	Applies policy rules to route anomalies to auto-approval or human/auditor review.	`Human or auditor review`, `Investigator agent`
`Human or auditor review`	Provides oversight, additional information, and reinforcement-learning feedback.	`Investigator agent`
`Investigator agent`	Explains why a fraud decision happened using long-term memory, policy rules, and stored JSON traces.	`Observability and audit tracing`
`Observability and audit tracing`	Captures model, data, and agent-level traces for drift monitoring and later audit review.	—

Data flow

In the fraud-detection case, ERP and invoice data enter a planner/executor workflow. A receipt extractor applies OCR, LLM, RAG, and cache capabilities, emits structured JSON, sends it to entity resolution for deduplication and normalization, then to anomaly detection and a decision router. The router either auto-approves, asks a human or auditor for more information, or lets an investigator agent explain the decision from stored traces and memory.

“imagine you have your ERP and data feed coming in. Uh you know like the way we started uh is you know we have a planner and an execution agent or an executor agent.”transcript, stated

“once that received extraction extractor kicks in and it sends back what it finds as a JSON format for instance it kicks in uh you know like the the executor agent kicks back to the entity resolver.”transcript, stated

Why this architecture

Wesco chose a platform-and-governance architecture because isolated pilots do not prove enterprise value. The system pairs reusable architecture, an air-gapped experimentation path, business-owned prioritization, LLMOps observability, and human oversight so that AI use cases can move from sandbox experiments into production workflows with measurable ROI and auditability.

“you have to involve uh your your ancillary teams like legal and HR and marketing to make sure that it's it's not only ROI but you're also thinking about compliance and ethics”transcript, stated

“proper governance to go from P to production uh as as we uh you know like look at these experiments and kind of make that call to push it above the line to scale”transcript, stated

Models used: LLMs · OCR models · RAG · LangGraph · Isolation Forest · Open-source models · Managed-provider LLMs · Self-hosted models · Domain-specific language models

Scale

Wesco is described as serving more than 50 countries.
“Wesco, a B2B supply chain and industrial distribution company serving over 50 countries”summary, stated
The production portfolio includes more than 50 use cases.
“Results include deployment of 50+ use cases, significant improvements in employee productivity through "everyday AI" applications”summary, stated
The fraud-detection example uses five specialized downstream agents.
“we have like five different agents here starting from a receipt extractor which is based on that intelligent document crossing”transcript, stated

Alternatives considered (roads not taken)

Move individual AI pilots directly into production.

Rejected because: Wesco framed production scaling as requiring governance, reference architectures, observability, monitoring, and MLOps/LLMOps pipelines.

“what does scaling comprise like includes uh architecture it includes like reference architectures it includes observability and monitoring pipeline”transcript, stated

Measure generative AI with only technical NLP metrics.

Rejected because: The organization intentionally translated metrics like BLEU into business KPIs such as engagement, cost, time savings, and ROI.

“we were just looking at a blue score to start with uh because this is text generation heavy on NLP but then the way we are translating that to auto business metrics”transcript, stated

Let the fraud system auto-approve every anomaly decision.

Rejected because: The router keeps a human or auditor in the loop when more information is needed.

“it should be auto approved or it should kick in uh or it should kick back to a human in the loop like either the user or an auditor.”transcript, stated

Patterns

`multi-agent-orchestration` (agent architecture)

The fraud-detection workflow uses a planner/executor plus specialized receipt-extraction, entity-resolution, anomaly-detection, decision, and investigator agents.

“we have a planner and an execution agent or an executor agent. So these are the ones that you see on the top with multiple faces. Uh and these things feed off to multiple agents.”transcript, stated

`human-in-the-loop-gate` (control)

Fraud decisions can be routed to a user or auditor instead of automatically approved when the policy rules call for review.

“there it decides if it if it should be auto approved or it should kick in uh or it should kick back to a human in the loop like either the user or an auditor.”transcript, stated

`observability-tracing` (operations)

The platform traces agent decisions so auditors can later inspect why an autonomous decision was made.

“if an agentic AI uh acts on a decision uh autonomously you know like a few days few weeks few months down the lane you your your auditor wants to see why that decision was made”transcript, stated

`eval-driven-development` (measurement)

Wesco links technical metrics to business KPI measurement, including A/B testing and pre/post deployment comparisons.

“are we doing AB testing and see that you know pre uh pre-eployment of this capability and post deployment do we see that incremental improvement”transcript, stated

`fallback-degradation` (risk control)

The decision router can fall back from autonomous approval to human/auditor intervention when confidence or policy requires it.

“based on the policy rules uh set within the company uh it kind of um you know like kicks uh out to a decision- making agent uh using a router logic”transcript, stated

Tradeoffs

Platformize AI use cases before scaling them.

Chose A composable platform with YAML configuration, reference architecture, governance, and LLMOps over One-off low-code or pro-code pilots.

Rationale: The platform path lets teams experiment quickly while keeping a governed route to production.

“We've also built our own platform uh which has a composible design and composible architecture and you know using YAML based configuration files we can quickly spin up a chatbot”transcript, stated

Keep experimentation isolated but production governed.

Chose An air-gapped sandbox plus a governed promotion path over Unrestricted experimentation directly in production-like environments.

Rationale: The sandbox enables rapid iteration without skipping the risk-management step needed for production deployment.

“this sandbox environment which is completely airgapped. uh so we can quickly test and iterate. So it helps us it it you know like that there is an ease of uh experimentation and testing”transcript, stated

Use a hybrid model strategy.

Chose Open-source, managed-provider, self-hosted, and domain-specific models over A single generic external LLM for all enterprise tasks.

Rationale: Different tasks need different cost, privacy, control, and enterprise-context properties.

“We have multiple policies whether that is a hybrid AI model strategy you know or to support that hybrid AI model strategy uh both for open source models”transcript, stated

Gotchas (what bit them)

POCs do not automatically become enterprise value.

Root cause: Scaling requires governance, ROI mapping, architecture, observability, and MLOps/LLMOps rather than only a working model.

Resolution: Use business-driven intake, prioritization, reusable architectures, and production promotion controls.

“the challenges with AI projects especially when you take it to scale uh or especially as you move beyond P. Now every company faces a mix of these challenges”transcript, stated

Technical model scores can fail to prove business impact.

Root cause: Data-science metrics such as accuracy or generation quality do not map directly to revenue, margin, cost-to-serve, productivity, or customer experience.

Resolution: Translate technical metrics into business KPIs and pre/post deployment measurements.

“how do we kind of transition from ROI from a pure play technical metrics to business KPIs and how you make that connection”transcript, stated

Agentic systems need more than infrastructure monitoring.

Root cause: Auditable autonomous decisions require model, data, and agent-level traceability, not only platform uptime.

Resolution: Add LLM and agent observability, trace stores, drift checks, and feedback loops.

“It's not just infra level monitoring or even like data level observability but you need to have multiple layers of this observability and monitoring.”transcript, stated

Evaluation

Evaluation combines technical metrics, business KPIs, A/B testing, drift monitoring, human feedback, and auditor-facing traceability.

Metrics: BLEU-like text-generation score · A/B-tested click-through or engagement changes · Cost and time savings · ROI and business KPI impact · Data/model drift · Traceability of autonomous decisions

“we look at drift you know like for the data and for the traditional ML we look at both the data the model drift as a part of our MLOps”transcript, stated

“there is also continuous learning from human feedback so we have put that reinforcement learning in the loop for most of the products”transcript, stated

Outcomes

Wesco reports a portfolio of more than 50 production AI use cases.
The organization connected everyday AI to employee productivity and transformational AI to higher-ROI supply-chain work.
The fraud-detection example demonstrates an auditable end-to-end multi-agent workflow.

Wesco's case is a broad enterprise AI program across a B2B supply-chain company serving more than 50 countries, not a single isolated app.

“Results include deployment of 50+ use cases, significant improvements in employee productivity through "everyday AI" applications”summary, stated

“So we have put this together and um you know like as I said um the the the goal here is like as um as we continue to mature this”transcript, stated