Industrial AI buyer's toolkit · updated 23 September 2026
IndustrialProcessAI
Home / Guides / LLMs and AI assistants for manufacturing
Buyer's guide · plant, operations, CI and IT leaders

Which LLM or AI assistant fits your manufacturing operations in 2026?

The short answerPick the assistant that connects to the systems already holding your production data, then check that data is complete. Claude for Manufacturing reads ERP, CAD and Microsoft 365 data; Anthropic sees job-based shops as its best current fit, as reported by Unite.AI. Microsoft and Google tie their agents to their own data platforms. None of them captures machine stops and counts by itself.

Last updated: 21 November 2026 · Sources listed at the end of the page

What do the main AI assistants for manufacturing actually do?

Since mid-2026, the large AI providers have shipped manufacturing-specific packaging around their general models. The products differ less in model quality than in what they plug into. The table compares what each provider has published.

ProviderProductWhat it connects toCaptures machine data itself?Best fitSource
AnthropicClaude for Manufacturing (Team and Enterprise plans), 17 skillsAutodesk Fusion, Infor, Microsoft 365, Oracle NetSuite; other systems through regular exports saved to Microsoft 365No. Works from connected systems and exports; does not operate equipmentJob-based manufacturing where each job has its own quote, drawing and routing (Anthropic's view, as reported by Unite.AI)Anthropic, Unite.AI
InforInfor connector for ClaudeInfor ERP, read and writeNoInfor customers who want approved transactions executed from the conversationInfor, The Manufacturer
MicrosoftFactory Operations Agent (Copilot Studio, Microsoft Foundry); Azure IoT Operations and Microsoft FabricMES and QMS data; edge telemetry routed through Azure IoT Operations into FabricThe agent does not; Azure IoT Operations carries telemetry from the edgePlants whose MES and QMS data already runs on Microsoft's stack (our reading)Microsoft, Microsoft
Google CloudGemini Enterprise with Manufacturing Data EngineMachine telemetry, PLC and SCADA streams contextualised into BigQuery, alongside ERP recordsIngests and contextualises it; still depends on the machines and controllers producing signalsPlants building a unified data layer on Google Cloud (our reading)Google Cloud
AWS“Built for Industrial AI” programme shown at Hannover Messe 2026Not specified as a single assistant; agentic AI and physical AI demonstrationsNot statedNot enough published detail to judgeAWS
Siemens + NVIDIA“Industrial AI operating system” (stated ambition)Siemens Electronics Factory in Erlangen as the first blueprint for AI-driven, adaptive sites from 2026Not statedLarge Siemens-equipped sites following the Erlangen model (our reading)NVIDIA, Manufacturing.net

Anthropic's 17 skills are grouped under sales, quality, operations, supply chain, design and setup. Its production report compares output against target and shows equipment downtime, with the cost of that downtime and of missed output. Unite.AI notes that the core plugin's Infor connector is limited to record lookups and that the plugin does not write to systems of record. Writing back is the job of Infor's own connector, which supports actions such as creating customer orders, releasing purchase orders and issuing raw materials to production, and can recommend rescheduling when a production risk cannot be resolved. The Manufacturer reports that shifting production lines after unplanned downtime is among the actions run after user approval, and that connector actions are recorded in Infor.

Microsoft's Factory Operations Agent in Copilot Studio gives production insights in natural language from MES and QMS data, covering bottlenecks and metrics such as OEE and yield, and supports root-cause analysis. Google's Manufacturing Data Engine ingests and contextualises machine telemetry, PLC and SCADA data into BigQuery, the data layer Google places under Gemini Enterprise agents. One company-reported example: GE Appliances says it has deployed more than 800 AI agents across manufacturing, logistics and supply chain, and that agents analyse shift data in minutes rather than hours. Siemens reports, for Erlangen, up to 85% less engineering effort for new PCB assemblies, up to 40% shorter time-to-market and up to 20% cost savings from agent-based production planning. These are vendor figures, not independent measurements.

What data do they all need from the shop floor?

Every product in the table reasons over data that some other system has already recorded. Anthropic cites a 2025 Manufacturing Leadership Council survey in which 79% of manufacturers said they enter data manually. An assistant asked about downtime cost on a line whose stops are written on paper at the end of the shift will produce a confident report built on incomplete numbers.

Gartner's March 2026 Market Guide for MES points the same way. In Siemens' summary of its key takeaways, AI “requires modern, connected MES environments”, and interoperability (open APIs, composability, MCP) plus a strong data fabric are described as essential to scale AI.

The minimum the assistants need from the floor, for the use cases vendors are demonstrating:

  • Machine states with timestamps: running, stopped, slow. Without them, “equipment downtime” in a production report is an estimate.
  • Stop reasons: coded consistently enough that a model can group them.
  • Good and scrap counts per order or shift: the basis of “output against target”.
  • A link to the order: so machine events can be tied to the ERP job, quote and routing.
  • An access path: an export, an API or an MCP server the assistant is allowed to call.

Several shop-floor data vendors now expose this layer to assistants. ENLYZE offers an MCP server usable inside Claude, Cursor and VS Code to ask about OEE, downtime, machine performance, production runs and machine status. MachineMetrics announced Max AI in November 2025 as an AI layer for its MES platform that combines machine data with ERP context. TEEPTRAK captures stops and part counts automatically on machines of any age and offers Excel export and APIs. The list is illustrative, not a ranking.

How should you choose an AI assistant for manufacturing?

Use these five checks before any pilot. Each one can be answered in a vendor demo.

  1. Map the connectors to your systems of record. List where orders, routings, quality records and machine events live today. Prefer the assistant with a native connector to the system that holds the data you most need; exports are a workable fallback, not a live feed.
  2. Audit the machine data before the model. If stops and counts are typed in by hand, fix capture first. Ask the vendor which report breaks when a machine has no automatic signal.
  3. Separate read from write. Decide which actions, if any, the assistant may execute in the ERP. Confirm that existing roles, permissions and approval controls apply, and that every action is logged in the system of record.
  4. Match the product to your production model. Anthropic's view, as reported by Unite.AI, is that the best current fit is job-based work with quotes, drawings and routings, while Unite.AI also notes that process manufacturers can use some skills immediately. Check skill by skill rather than at product level.
  5. Buy a measurable use case, not a platform. MIT NANDA's 2025 study found that for 95% of companies in its dataset generative AI was falling short, while tools bought from specialised vendors succeeded about 67% of the time. Fix one KPI, such as time to produce the daily production report, and measure it before and after.

Where does human review stay mandatory?

Every vendor in this guide keeps a person in the decision. Anthropic designs its manufacturing skills so that the person responsible reviews and approves the work before it reaches a customer or supplier. Infor's position is that agents do the legwork while people make the calls: when an action could affect another order or customer, the agent returns to the user for approval. Gartner lists human-in-the-loop controls and explainability under trust.

In practice, keep review mandatory for any output that leaves the plant, such as quotes, supplier messages and customer delivery dates, and for any action that changes the plan, such as a purchase order or a schedule change. Treat a report built on machine data you have not verified as a draft. Querying and summarising can run with lighter oversight once the source data has been checked.

What changes next?

Two developments are worth watching, with caution about timing.

Agents that operate equipment. Anthropic opened a research preview of the Model Hardware Standard (MHS) on 27 August 2026, described as a shared specification for AI agents to safely operate physical devices. It works with any device that has a programmable interface, is model-agnostic and can be reached through standard protocols such as the Model Context Protocol. It is a preview for an initial group of labs and manufacturers, not a production feature, and Claude for Manufacturing itself does not operate equipment.

MCP as the connection layer. Shop-floor vendors are publishing MCP servers, and Gartner names MCP among the interoperability requirements for scaling AI on MES. If that continues, the choice of assistant may matter less than whether your machine-data system exposes a well-governed MCP or API endpoint. Ask current and prospective vendors for their MCP plans in writing.

FAQ

Which AI assistant is best for manufacturing in 2026?
There is no single best one. For Claude for Manufacturing, Anthropic sees the best current fit in job-based manufacturing where each job has its own quote, drawing and routing, as reported by Unite.AI. Microsoft and Google tie their agents to their own data platforms. The deciding factor is usually which systems already hold your production data.
Can an LLM read machine data directly from the shop floor?
Not on its own. The assistants read what a connected system holds: an ERP, an MES, a data platform or an export. Machine states, stops and counts have to be captured first by a shop-floor system or an industrial data layer, then exposed through a connector, an API or an MCP server.
Does Claude for Manufacturing control machines?
No. Anthropic states that Claude for Manufacturing does not operate equipment. Its skills end with a person reviewing and approving the work before it reaches a customer or supplier. Anthropic's separate Model Hardware Standard for agents operating devices is a research preview.
Can these assistants make changes in the ERP?
Some can, with approval. Infor's connector for Claude supports write actions such as releasing purchase orders, subject to the user's existing roles, permissions and approval controls. Claude for Manufacturing's own skills do not write to systems of record.
Why do so many generative AI pilots fail to show results?
MIT NANDA's 2025 study found that for 95% of companies in its dataset, generative AI implementation was falling short, with most pilots showing little to no measurable P&L impact. Tools bought from specialised vendors succeeded about 67% of the time, roughly twice the rate of internal builds.

Sources

  1. Anthropic, “Claude for Manufacturing”, 12 November 2026
  2. Unite.AI, “Anthropic launches Claude for Manufacturing with 17 workflow skills”
  3. Infor, “Claude in manufacturing operations”, 12 November 2026
  4. The Manufacturer, “Infor teams up with Anthropic to integrate Claude AI into manufacturing operations”, 16 November 2026
  5. Anthropic, “Model Hardware Standard: research preview”, 27 August 2026
  6. Microsoft Adoption, scenario library: “Improve factory operations”
  7. Microsoft, “Industrial intelligence unlocked: Microsoft at Hannover Messe 2026”
  8. Google Cloud, “The agentic factory”, 10 September 2026
  9. Google Cloud press release, GE Appliances and Gemini Enterprise, 22 April 2026
  10. AWS for Manufacturing at Hannover Messe 2026, “Built for Industrial AI”
  11. NVIDIA newsroom, Siemens and NVIDIA expand partnership, 6 January 2026
  12. Manufacturing.net, “Siemens expands AI at Erlangen plant”, October 2026
  13. Gartner, 2026 Market Guide for Manufacturing Execution Systems, March 2026
  14. Siemens, key takeaways from the Gartner 2026 Market Guide for MES
  15. Yahoo Finance on MIT NANDA, “The GenAI Divide: State of AI in Business 2025”
  16. MachineMetrics, “MachineMetrics unveils Max AI”, November 2025
  17. ENLYZE, ENLYZE MCP product page
  18. PerfTrak product page (teeptrak.com)

Company-reported figures are labelled as such. In the comparison table, “best fit” for Anthropic is Anthropic's view as reported by Unite.AI; entries marked “our reading” are editorial inferences from what each product connects to.

Before you shortlist
Most failed pilots fail on data, not on the model
Free, no login. One follow-up from TEEPTRAK, then nothing.
Score your data readiness →
Industrial AI Vendor Evaluation Scorecard + RFP
The weighted scorecard and RFP question bank that keep vendor comparisons honest: with a POC plan built in.
Download the Scorecard + RFP →