As organizations explore how artificial intelligence can move beyond chatbots and copilots, one question keeps surfacing: What does AI look like when it’s embedded directly into real business operations?
At ProdCon WI in late 2025, that question was explored through a live demonstration by Kevin Quon, Principal Architect at nvisia, who showcased an agentic purchase order system designed to automate a complex B2B workflow end to end.
The demo - part of nvisia’s AI Lab showcase at the ProdCon conference - wasn’t about replacing a single task. It was about re-imagining how enterprise processes themselves can be decomposed, reasoned over, and orchestrated by AI agents working together.
Purchase orders sit at an interesting intersection of structure and complexity. They’re:
Highly procedural
Dependent on multiple business roles
Backed by deeply nested ERP data
Sensitive to pricing, inventory, logistics, and compliance
In other words, they’re exactly the kind of workflow that breaks simplistic automation - and exactly the kind that benefits from specialized reasoning.
Purchase orders already map cleanly to roles inside an organization. Customer service, inventory, logistics, accounting - each of those responsibilities already exists. Agentic AI just gives us a new way to model them.
Rather than treating AI as a single omniscient system, Kevin’s approach treats agents as business specialists, each responsible for a narrowly defined domain.
Applying agentic thinking with architectural discipline preserves agent Intelligence in grounded enterprise reality
Early experiments with fully open-ended agentic swarms revealed an important insight: not every business process benefits from unconstrained agent autonomy.
For workflows that are well defined and step-driven (like purchase orders) too much freedom can actually degrade reliability. Context windows fill up, reasoning becomes inconsistent, and deterministic steps blur.
The solution wasn’t to abandon agentic thinking - it was to apply it with architectural discipline. Instead of a free-roaming swarm, Kevin implemented a workflow-oriented, multi-agent system, where:
Each agent has a specific role
Each agent has access to a limited, relevant toolset
Context is passed deliberately from one stage to the next
This shift preserves the intelligence of agents while grounding them in enterprise reality.
The demo begins with something familiar: an incoming purchase order email.
From there, the workflow unfolds through a sequence of agents aligned to real business responsibilities:
Order Interpreter: extracts structured data from unstructured email content
Customer Service Agent: identifies customer identity, billing, and shipping details
Product Matching Agent: validates items, substitutes equivalents, and resolves SKUs
Inventory Agent: checks availability and calculates lead times
Logistics Agent: evaluates fulfillment and shipping constraints
Accounting Agent: applies pricing rules, discounts, and payment terms
Each agent contributes context, enriching the purchase order step by step. Rather than asking one system to “figure everything out,” Kevin describes this as programming through delegation.
“If an agent has too many responsibilities, it loses focus,” he notes. “Specialization is how you reduce hallucination and increase trust.”
A critical part of the architecture is recognizing what agents don’t know by default. That’s where tools come in.
Each agent is equipped with specific capabilities - such as SQL queries against ERP-style databases - allowing it to retrieve information that exists outside the model’s training data.
To ensure consistency and predictability, outputs are constrained using structured output, preventing the kind of formatting drift that often plagues LLM-driven systems.
The final result isn’t just a conversational response - it’s a structured purchase order that can be transformed into EDI X12 855, a standard B2B acknowledgment format used across enterprise supply chains.
This is AI operating inside the systems businesses already rely on - not alongside them.
Beyond the demo itself, the system reflects a production-oriented mindset. The agentic workflow is modeled as:
Discrete, object-oriented agent components
Graph-based orchestration defining handoffs and relationships
Containerized services suitable for deployment
Clear separation between domain logic, orchestration, and persistence
In short, it’s not a notebook experiment - it’s an application architecture, built with the same rigor expected of enterprise software.
Kevin describes this as viewing agentic AI not as a new kind of model, but as a new way of programming.
“We’re still writing code,” he explains. “We’re just writing it at a different level of abstraction.”
What this demo ultimately illustrates isn’t just an automated purchase order. It’s a blueprint for how AI can:
Respect organizational structure
Mirror real business responsibilities
Integrate with existing enterprise systems
Scale beyond single-agent interactions
Rather than replacing people, the system encodes how people already work - and allows AI to operate within those boundaries.
This approach is especially relevant for industries where reliability, auditability, and predictability matter just as much as speed.