"Manufacturers Want AI Answers From ERP Data. Read-Only Is Only the First Control"
Manufacturers are beginning to connect generative AI tools directly to inventory, production, purchasing and compliance systems. The promise is straightforward: employees can interrogate operational data in plain English instead of assembling reports, searching documents or waiting for a specialist. The risk is equally direct because an answer can sound authoritative while relying on incomplete records, misunderstood definitions or data the user should never have accessed.
Wherefour, a manufacturing ERP and traceability provider serving small and midsized producers, launched four AI capabilities on October 6: a read-only Model Context Protocol connector, document auto-tagging, AI-generated control points and AI-drafted quality-test definitions. The company says the tools remain disabled by default, allowing customers to decide when and how they should be introduced. This positions the launch as an attempt to bring AI into manufacturing workflows without immediately giving it authority over live transactions.
Read-Only Access Limits Damage, Not Bad Decisions
The connector allows compatible AI tools to query Wherefour data covering inventory, production, purchasing, sales and traceability. Users could ask which products are most profitable, where excess inventory is accumulating or what demand should be forecast next. Matthew Brown, CEO of Wherefour, said the approach is intended to give smaller manufacturers access to analytical capabilities they may not have internally. “Many small to mid-sized manufacturers don't have an operations data scientist on staff, and this gives them something close to one,” Brown said in a written response to The Supply Chainer. “They can ask which products are most profitable, where they’re carrying too much inventory, or what they need to forecast next and get an answer in plain English from their own data.”
The connector is read-only, meaning the AI can retrieve and interpret information but cannot alter inventory, orders or production records. Brown recommends that manufacturers begin there before considering any system capable of executing transactions. “If AI can view inventory and orders but can’t change them, wrong answers never turn into transaction errors,” he said. “That’s how we built our MCP connector so that Wherefour remains the system of record.”
This is a meaningful safeguard, but it does not make the output correct. A misleading profitability calculation could still influence purchasing, pricing or production planning even if the AI never writes directly to the ERP. Read-only access reduces the system’s technical authority, but it does not eliminate its influence over human decisions. Manufacturers will therefore need to distinguish between preventing unauthorized transactions and preventing employees from acting on inaccurate analysis.
Human Approval Needs More Than a Button
Wherefour’s other features move closer to operational workflows. Its document tool can read a certificate of analysis and suggest tags such as lot number, vendor or barcode, while its AI can draft process control points and quality-test configurations. Users must review and approve those drafts before implementation. Brown said user-level permissions determine who can access AI functions, while the company’s Control Points can require reviews or signatures at defined stages. “Every approval is logged, so manufacturers get the speed of AI without sacrificing quality control,” he said. “They know who signed off at each step, and if something goes wrong, they can see exactly where.”
That approach aligns with broader guidance, but an approval record proves only that someone approved an output. It does not establish whether the reviewer had sufficient context, understood how the answer was produced or checked the underlying source data. The National Institute of Standards and Technology warns that generative AI may require “additional human review, tracking and documentation, and greater management oversight.” Precisely Chief Product Officer Matt Waxman made a similar point when discussing AI agents connected to enterprise systems: “AI agents need access to the same trusted data management capabilities and established controls organizations already rely on.”
The Reporter’s View: Governance Starts With the Question
Wherefour’s design places useful boundaries around a new interface to manufacturing data. Read-only access, disabled-by-default functions, user permissions and recorded approvals all reduce obvious operational risk. The harder problem sits upstream, where manufacturers must define which data an AI may combine, which calculations require validation and which recommendations demand specialist approval. A purchasing manager, quality lead and finance director may all ask what appears to be the same question about product profitability but require different data, assumptions and access rights.
Auto-tagging introduces another risk because an incorrectly labeled certificate, lot number or supplier record may become easier to search and reuse precisely because the system has organized it so efficiently. The first weeks of customer use will therefore not establish whether the technology saves time at scale. The stronger test will be whether manufacturers can identify incorrect answers, challenge recommendations and trace each conclusion back to reliable operational records. Governed AI is not simply AI with an approval screen; it is a system in which access, definitions, evidence and accountability remain visible even when the answer arrives in seconds.
