S&B Filters: From Prototype to Production AI in 4 Weeks

S&B Filters is a U.S.-based manufacturer of high-performance air filters for the automotive aftermarket. With 700+ employees and a complex, multi-channel business, the company runs its core operations — order management, internal workflows, customer data — on NetSuite.

As the business scaled, so did the volume of customer inquiries — especially around order status, delivery timelines, and product details — putting increasing pressure on the support team.

The company attempted to solve this with AI — but the results fell short. 

S&B's CEO, Berry Carter, is the kind of leader who builds before he buys. Rather than wait for a vendor pitch, he wired up Claude's MCP (Model Context Protocol) connector to NetSuite himself, wrote his own prompts, and stood up an internal AI assistant for order status.

For a CEO-built prototype, it worked. It proved the use case. It showed the team what was possible.

Then the limitations of an experiment-grade architecture surfaced:

• Latency: Response times reached 4–6 minutes, making it unusable in real customer interactions.
• Prompt architecture: The system relied on a 40-page prompt, making it difficult to maintain, hard to control, and inconsistent in outputs.
• Data variability: PO numbers arrived as "PO-1234," "1234," and other formats. Orders spanned Shopify, phone, and email with different conventions
• No path to production: Behaved as an internal tool, not a customer-grade system

We didn't extend the prototype. We rebuilt the architecture from the ground up — a custom AI agent stack engineered for S&B's data, channels, and operational realities. 

The result is not a chatbot. It's an AI layer on top of NetSuite that serves internal teams and external customers from a single foundation.

We started with what delivers the most immediate value: instant access to NetSuite data and order status. The AI assistant now:

To maximize impact, we deployed the solution across two key environments:

1. Internal AI Assistant (for support teams)

The most critical use case for operations:

👉 What used to take minutes inside NetSuite now takes seconds.

2. Customer-Facing AI Assistant (website)

A self-service layer for end users:

We extended AI beyond transactions into product knowledge. The assistant now: 

This complements NetSuite by bridging the gap between data and customer understanding.

Previously, agents had to manually navigate the NetSuite interface to locate order details — a time-consuming and fragmented process. With the AI assistant, they can now simply enter an order or PO number and instantly retrieve all relevant information.

What used to take minutes of manual search is now reduced to seconds — significantly improving both response speed and agent efficiency.

NetSuite is structured. Real-world inputs are not. We built a layer above the ERP that:

👉 This allows the system to consistently deliver accurate results — even when inputs vary across channels.

Product knowledge is continuously evolving — with new products, updated installation guides, and varying levels of access requirements. While NetSuite provides a robust foundation, enabling this level of flexibility and content control benefits from an additional, purpose-built layer.

We built a dynamic knowledge layer connected to — but not limited — NetSuite:

👉 The AI always delivers accurate, up-to-date answers — without exposing sensitive internal information.

The previous solution technically worked — but with 4–6 minute response times, it was unusable in real workflows. Support agents still had to manually search NetSuite, and customers experienced delays.

We optimized for instant access to NetSuite data, not just answer quality:

👉 What used to take minutes now takes seconds — enabling both faster support and a seamless customer experience.

• LLM Orchestration — Streaming inference, tool calling, and context pipelines that keep the model grounded · OpenAI API

• Java Spring Boot — AI tool webhook execution, attribute resolution strategies, and cross-service business logic

• Docker · GitHub Actions · AWS ECR — Fully automated CI/CD; containers built and shipped to the cloud on every push

• Widget UI — Streams model responses into the interface in real time · React

• Admin Platform — Prompt management, analytics, and system observability · React · Redux · React Router

• Pinecone — Vector store for top-k semantic retrieval, grounding the model in relevant knowledge and reducing hallucinations in production

*Methodology note: Estimates are based on pre-implementation ticket volume, average handling time reduction driven by AI-assisted responses, and typical support agent cost benchmarks. Automation rate reflects the share of inquiries fully resolved by AI without human intervention, while ROI accounts for both operational savings and implementation costs.

For S&B Filters, this is no longer a support project. It's the foundation for an AI-first operation:

For other companies running on NetSuite, S&B is a reference point. The same approach — production AI layered on top of the ERP — applies across operations, finance, support, and sales.

Already on the roadmap: