Detailed Analysis
S&B Filters, a U.S. manufacturer with over 700 employees, became an instructive case study in the gap between AI prototyping and production-grade deployment when its CEO personally built an internal AI assistant using Anthropic's Claude MCP connector integrated with NetSuite, the company's enterprise resource planning backbone. The prototype successfully handled order status lookups and demonstrated core viability, but quickly revealed critical weaknesses when subjected to real operational conditions: response times stretched to four to six minutes, a single 40-page prompt was holding the entire system together, and purchase order numbers arrived in incompatible formats from three separate sales channels — Shopify, phone, and email — with no reconciliation layer in between. The CEO, having proven the concept himself, then engaged an external development team to rebuild the system from the ground up rather than patch the existing prototype.
The engineering team's reconstruction centered on NetSuite as the authoritative data source and prioritized an input normalization layer as the primary technical challenge, accounting for roughly 80% of total development effort. This layer validates inputs across formats, cascades through identifier fallbacks from Sales Order numbers to PO numbers to customer references, and leverages conversational context when input data is ambiguous or malformed. From a single backend, the team deployed two distinct interfaces — one internal tool for support staff and one customer-facing assistant embedded on the company's website — with differentiated access controls governing what each interface could expose. The knowledge base was connected to OneDrive, allowing client staff to update content without requiring redeployment, a critical operational feature for a business with ongoing product changes.
The quantified outcomes are significant for a deployment of this scale: approximately 50% of support requests now resolve fully through automation, first response times improved by a factor of 24, and the company projects roughly $140,000 in annual savings with a 250% return on investment in the first year. These figures reflect a broader pattern emerging across mid-market manufacturers and distributors, where ERP-connected AI agents are proving particularly high-value because the underlying transactional data is already structured, authoritative, and query-amenable — the obstacle has never been data availability but rather the normalization and orchestration layer sitting between messy real-world inputs and clean database records.
The case illustrates a recurring dynamic in enterprise AI adoption: executive-led prototyping accelerates organizational buy-in and validates feasibility faster than traditional IT procurement cycles, but the prototype itself typically cannot survive contact with production traffic, format diversity, or multi-user concurrency. The CEO's willingness to build the initial version personally compressed the discovery phase and produced a concrete artifact around which engineering requirements could be specified. Claude's MCP architecture enabled that rapid prototyping by abstracting NetSuite connectivity into a manageable interface, but the production system required an entirely different engineering discipline — one focused on fault tolerance, identifier disambiguation, and separation of concerns across user personas.
The expansion roadmap S&B Filters is now pursuing — full order management, dealer identification, and personalized discounting through the same system — reflects a strategic pattern in which a narrow, successful AI deployment becomes the foundation for progressively broader process automation. Each new capability leverages the same normalization infrastructure and AI layer already validated in production, reducing marginal implementation cost and risk. For Anthropic, cases like this serve as evidence that Claude's MCP tooling can support genuine enterprise integration beyond consumer or developer use cases, particularly when paired with the kind of domain-specific engineering that transforms a compelling demo into a scalable operational system.
Read original article →