Executive Summary

For HoseMaster, customer demand does not always arrive as clean, structured data.

Requests for quotes and purchase orders flow into a shared sales inbox through emails containing different formats, terminology, attachments, product references, and varying amounts of information. Turning those communications into actionable order data requires staff to interpret what the customer is requesting before the quoting and order-entry process can move forward.

Working within an engagement led by EIS IT Consulting, Cadensoft developed an NLP-powered quoting agent designed to interpret these inbound communications and transform relevant information into structured order data.

The solution combines large language model processing, retrieval against historical order information, structured output requirements, and a human-in-the-loop review process. High-confidence results can move through an auto-accept workflow while uncertain results are routed for review.

During proof-of-concept operation, the system achieved an approximately 80% success rate identifying orders, demonstrating the potential to automate a significant portion of a workflow previously dependent on manual interpretation.

The result is not simply an AI experiment. It is a working example of how applied AI can begin turning an existing communication channel into a more structured operational workflow.

Client Background

HoseMaster is an industrial hose manufacturer and distributor serving customers whose requirements can involve detailed product specifications and configurations.

Customer RFQs and purchase orders frequently arrive through email, making the shared sales inbox an important entry point into the organization’s quoting and order-processing workflow.

Unlike a structured ecommerce transaction, however, an email does not necessarily arrive in a predictable format.

The information needed to understand an order may exist within the body of a message, attached documents, previous correspondence, customer terminology, or specific product references.

That created an opportunity to determine whether modern natural language processing could interpret those communications reliably enough to reduce the manual effort required to identify and structure incoming orders.

The Challenge

The Inbox Was Designed for Communication, Not Structured Order Processing

Email works extremely well as a communication channel.

It is considerably less predictable as a source of structured operational data.

Inbound communications contained far more information than the quoting process actually required. Relevant product information could be surrounded by signatures, previous correspondence, unrelated text, attachments, customer-specific terminology, and other noise.

The challenge was therefore more complicated than extracting fields from a standardized document.

The system needed to determine what information actually mattered.

It then needed to interpret that information accurately enough to identify the requested products and return the required data in a consistent structure.

Product similarity created an additional challenge. Closely related product variants could produce false-positive matches, requiring the model to distinguish between items whose descriptions or characteristics appeared similar.

The operational objective was not simply to have an LLM read an email.

It was to determine whether AI could become a reliable participant in the quoting workflow.

Why Traditional Automation Wasn’t Enough

Traditional workflow automation works particularly well when inputs are predictable.

An online form, for example, can require a customer to provide specific values in specific fields.

HoseMaster’s existing customer behavior did not provide that structure.

Customers were already submitting RFQs and purchase orders through email, and the content of those communications varied considerably.

Creating value therefore meant working with the information as it actually arrived rather than requiring every customer to adopt a new interaction model.

This made the problem well suited to an NLP and LLM-based approach.

The goal was to create structure from unstructured information while preserving human oversight where the system lacked sufficient confidence.

The Cadensoft Approach

Start With Real Orders

The agent was developed and tested using historical RFQs, purchase orders, sample quotes, and other real-world communications supplied by HoseMaster.

This was important because the variability within actual customer communication quickly exposed challenges that would have been difficult to reproduce with artificially clean test data.

Early testing revealed significant noise within inbound emails, requiring additional filtering and tuning before user acceptance testing.

Rather than treating that as an exception, the workflow was refined around the reality of the data.

Create a Retrieval Layer Around Existing Information

Relevant data was moved from SharePoint into Azure Blob Storage, with Azure AI Search providing retrieval capabilities to support the language model.

This retrieval-augmented generation approach gave the agent access to relevant contextual information when interpreting inbound requests.

The objective was not simply to ask a general-purpose model to guess what a customer intended.

It was to give the model access to relevant business context while constraining its response to a defined operational task.

Define the Output Before Automating the Workflow

Cadensoft worked collaboratively with HoseMaster to define the structured JSON output required from the agent.

This established a consistent data contract between AI interpretation and downstream workflow processing.

Instead of producing free-form AI responses, the model was expected to return specific properties that could be evaluated and used operationally.

That distinction is important.

For AI to participate meaningfully in business operations, its output must ultimately become usable by the systems and people responsible for the next step.

Keep People in the Decision Loop

The solution was deliberately designed around human oversight.

Processed quotes were separated into two paths:

Auto-Accepted

Results with sufficient confidence could progress through the automated workflow.

Pending Review

Items requiring additional validation were surfaced for staff review and correction.

A review interface was also developed to make it easier for HoseMaster staff to inspect processed emails and corresponding quote information.

This created an operational feedback loop.

Instead of requiring the AI to be perfect before creating value, automation could handle higher-confidence work while people concentrated their attention on exceptions.