
Insights
November 20, 2026
Claude for Manufacturing: office first, machines later

Insights
November 20, 2026
Claude for Manufacturing: office first, machines later
Anthropic’s Claude for Manufacturing launch covered three layers: a plugin for office work, an Infor connector inside the ERP, and a standard for AI to operate equipment. In office work, AI pays off most when it is built around the business’s own processes. On the factory floor, it needs a shared standard that the industry has tested.
Anthropic released Claude for Manufacturing on 12 November, a plugin of seventeen skills that drafts quotes, quality reports, breakdown diagnoses and supplier requests from records a plant already keeps. The technical webinar that followed on 17 November presented three separate pieces of work sitting at three different distances from the shop floor. Anthropic described an adoption path in which manufacturers start with code and office work, move into design and engineering, and reach the running of the factory floor last, and the products on show lined up with that sequence closely.
Read in that order, the office layer is ready to use now, the enterprise layer depends on which ERP a business runs, and the equipment layer is still in a limited research preview that manufacturers can apply to join. In our view, the lesson changes once AI moves from paperwork to equipment. In office and planning work, the stronger results come from businesses that build AI around their own processes and data from the ground up, using packaged tools as components, whereas on the factory floor no single business should be inventing its own way for AI to control machines, and that is where an industry-wide, tested standard matters most.

The Claude for Manufacturing plugin
The plugin covers sales, quality, operations, supply chain, design and setup, and connects to Autodesk Fusion, Infor, Microsoft 365 and Oracle NetSuite. Anthropic opened the session with a 2025 Manufacturing Leadership Council finding that 79% of manufacturers still enter data manually, which is the problem the plugin is built to reduce.
The demonstration followed a single request for quote for a pump mounting bracket, 50 pieces with an alternate quantity of 200. Claude read the RFQ form and the drawing, pulled past orders for the customer from the CRM and material prices and machine rates from the ERP, and listed the gaps that could change the price before doing anything else. An estimator accepted one assumption and left two others open. Claude then modelled the part, generated a two-setup machining plan at 13.46 minutes per part, and built a cost sheet that reached $84.06 a part at 50 pieces and $69.74 at 200. The quote stopped at approval with three issues still attached to it: whether the surface could be made flat enough, the remaining quality risks, and how the part would be held during machining.
Two caveats belong next to that sequence before anyone repeats it. The demo carried a label calling it an illustration of the workflow, not live software, and AI-assisted quoting is already common in most industries that receive RFQs. The detail worth keeping from the demonstration is that open issues travel with the quote to the person who signs it, so an unresolved assumption stays visible at the point where a price becomes a commitment.

Infor Industry AI for Claude
Infor presented Infor Industry AI for Claude, a certified connector and a plugin that let Claude read and act on manufacturing data inside Infor, with single sign-on and role-based entitlements deciding what each person can see and do. Infor also described Claude as its preferred knowledge-work tool across its own workforce.
The demonstrations were organised by role. A production planner asked which orders were at risk, and Claude traced a late plasterboard partition panel two levels down its bill of materials to a timber rail that was ten pieces short, then offered an alternative timber frame assembly that was already in stock. A customer service rep reviewed rush orders and saw options such as transferring stock from another warehouse to meet the requested date. A buyer dealing with a late shipment was shown which orders it affected and which alternative suppliers could cover the gap. Every change waited for the user to confirm it before anything was written back to Infor.
This is useful to Infor customers and of limited direct use to anyone else. The lesson that travels beyond Infor's customer base concerns data quality: the planner's answer depended on bills of materials, orders, stock and capacity sitting in one system accurate enough to trace through, and a business whose planning still lives partly in spreadsheets would receive a confident answer built on incomplete records.
A different approach: ALUCOR's in-house AI tools
ALUCOR, a Dubai-based EPC contractor and one of our clients, did not start from a vendor product. At WETEX on 21 October, Rahul Bhatia, its Vice President of Sales and Marketing, presented on deploying AI to improve construction productivity, arguing that productivity on site has barely moved in decades while project schedules keep getting shorter.
His examples were tools that the company had built in-house rather than bought. PulsePoint lets workers report site issues from their phones by voice, in their own language, and AI translates each report, categorises it and turns it into an assigned action that management sees the same day, at a rate of 50 to 100 reports a day on a single project. Alberio breaks the piping scope into spools and welds so that every joint is tracked from fit-up through to the test package at handover. The company also uses AI agents to extract material data from complex isometric drawings, and its fabrication shop runs robotic welding and 3D scanning.
Neither PulsePoint nor Alberio is the kind of thing a contractor buys as a package, because each one encodes how ALUCOR runs its sites: who reports a problem and in which language, where the report goes, and what a joint has to pass before handover. Packaged tools such as Claude for Manufacturing can supply pieces of that work, the translation, the drawing extraction, the drafting, but the value sits in the process designed around them, which only the business itself can do. Each tool started with one process that mattered, captured data at the point where the work happens, and left the decisions with people.

The Model Hardware Standard
Building your own stops being a sensible option once AI is expected to act on equipment, and the third talk addressed exactly that step, which Anthropic says AI cannot yet take: running a process safely and at scale on physical equipment. A typical cell mixes PLCs, robot controllers, vision cameras, weld controllers and torque tools from different vendors, each reached through its own protocol, and Anthropic put the timing problem plainly: a machine acts in about a millisecond, a person needs about 250 milliseconds to react, and a physical mistake cannot be rolled back.
The Model Hardware Standard proposes one driver per device, sitting on top of existing protocols such as OPC UA, EtherNet/IP and GigE Vision, which describes what an agent can read, set and run on that device. Limits are enforced inside the driver, so a request outside them is refused, and the equipment keeps its own logic and safety systems. The standard began as a project between Anthropic and HHMI Janelia Research Campus and is in limited research preview, with access by application and a stated aim of testing the standard and building safety evaluations before it is open-sourced.
The strongest evidence in the talk came from a laboratory setting at QuEra. There, Claude wrote a controller overnight that recovers a dropped laser lock in 0.9 to 14 seconds with a 99.3% success rate, against roughly 150 seconds and 58% for a hand-written script that had taken two to three months to build. The design choice behind that result matters more than the numbers, because the agent wrote ordinary code and then left the loop, so no AI is running at the moment the equipment acts. The plant walkthrough that followed used a fictional plant with simulated data, and Anthropic's recommended starting point for a manufacturer is read-only access on a single cell or bench.
The plugin and the Infor connector will make paperwork faster for any manufacturer whose records already sit in systems Claude can read, and the businesses that gain most from them will be those that have already designed the process the tools slot into. Equipment is a different matter, because a plant cannot afford to invent its own rules for what an agent may do to a robot or a press, and a shared standard with limits enforced in the driver is work that has to come from the industry and its suppliers. The least finished part of Anthropic's announcement may therefore be the one that matters most to manufacturers over the next few years.

Cactix Editorial Team
The Cactix Editorial Team is a crew of curious minds who write, edit, and shape ideas across marketing, communications, and the web. We care about clarity, substance, and telling stories that move people and businesses forward.
Anthropic’s Claude for Manufacturing launch covered three layers: a plugin for office work, an Infor connector inside the ERP, and a standard for AI to operate equipment. In office work, AI pays off most when it is built around the business’s own processes. On the factory floor, it needs a shared standard that the industry has tested.
Anthropic released Claude for Manufacturing on 12 November, a plugin of seventeen skills that drafts quotes, quality reports, breakdown diagnoses and supplier requests from records a plant already keeps. The technical webinar that followed on 17 November presented three separate pieces of work sitting at three different distances from the shop floor. Anthropic described an adoption path in which manufacturers start with code and office work, move into design and engineering, and reach the running of the factory floor last, and the products on show lined up with that sequence closely.
Read in that order, the office layer is ready to use now, the enterprise layer depends on which ERP a business runs, and the equipment layer is still in a limited research preview that manufacturers can apply to join. In our view, the lesson changes once AI moves from paperwork to equipment. In office and planning work, the stronger results come from businesses that build AI around their own processes and data from the ground up, using packaged tools as components, whereas on the factory floor no single business should be inventing its own way for AI to control machines, and that is where an industry-wide, tested standard matters most.

The Claude for Manufacturing plugin
The plugin covers sales, quality, operations, supply chain, design and setup, and connects to Autodesk Fusion, Infor, Microsoft 365 and Oracle NetSuite. Anthropic opened the session with a 2025 Manufacturing Leadership Council finding that 79% of manufacturers still enter data manually, which is the problem the plugin is built to reduce.
The demonstration followed a single request for quote for a pump mounting bracket, 50 pieces with an alternate quantity of 200. Claude read the RFQ form and the drawing, pulled past orders for the customer from the CRM and material prices and machine rates from the ERP, and listed the gaps that could change the price before doing anything else. An estimator accepted one assumption and left two others open. Claude then modelled the part, generated a two-setup machining plan at 13.46 minutes per part, and built a cost sheet that reached $84.06 a part at 50 pieces and $69.74 at 200. The quote stopped at approval with three issues still attached to it: whether the surface could be made flat enough, the remaining quality risks, and how the part would be held during machining.
Two caveats belong next to that sequence before anyone repeats it. The demo carried a label calling it an illustration of the workflow, not live software, and AI-assisted quoting is already common in most industries that receive RFQs. The detail worth keeping from the demonstration is that open issues travel with the quote to the person who signs it, so an unresolved assumption stays visible at the point where a price becomes a commitment.

Infor Industry AI for Claude
Infor presented Infor Industry AI for Claude, a certified connector and a plugin that let Claude read and act on manufacturing data inside Infor, with single sign-on and role-based entitlements deciding what each person can see and do. Infor also described Claude as its preferred knowledge-work tool across its own workforce.
The demonstrations were organised by role. A production planner asked which orders were at risk, and Claude traced a late plasterboard partition panel two levels down its bill of materials to a timber rail that was ten pieces short, then offered an alternative timber frame assembly that was already in stock. A customer service rep reviewed rush orders and saw options such as transferring stock from another warehouse to meet the requested date. A buyer dealing with a late shipment was shown which orders it affected and which alternative suppliers could cover the gap. Every change waited for the user to confirm it before anything was written back to Infor.
This is useful to Infor customers and of limited direct use to anyone else. The lesson that travels beyond Infor's customer base concerns data quality: the planner's answer depended on bills of materials, orders, stock and capacity sitting in one system accurate enough to trace through, and a business whose planning still lives partly in spreadsheets would receive a confident answer built on incomplete records.
A different approach: ALUCOR's in-house AI tools
ALUCOR, a Dubai-based EPC contractor and one of our clients, did not start from a vendor product. At WETEX on 21 October, Rahul Bhatia, its Vice President of Sales and Marketing, presented on deploying AI to improve construction productivity, arguing that productivity on site has barely moved in decades while project schedules keep getting shorter.
His examples were tools that the company had built in-house rather than bought. PulsePoint lets workers report site issues from their phones by voice, in their own language, and AI translates each report, categorises it and turns it into an assigned action that management sees the same day, at a rate of 50 to 100 reports a day on a single project. Alberio breaks the piping scope into spools and welds so that every joint is tracked from fit-up through to the test package at handover. The company also uses AI agents to extract material data from complex isometric drawings, and its fabrication shop runs robotic welding and 3D scanning.
Neither PulsePoint nor Alberio is the kind of thing a contractor buys as a package, because each one encodes how ALUCOR runs its sites: who reports a problem and in which language, where the report goes, and what a joint has to pass before handover. Packaged tools such as Claude for Manufacturing can supply pieces of that work, the translation, the drawing extraction, the drafting, but the value sits in the process designed around them, which only the business itself can do. Each tool started with one process that mattered, captured data at the point where the work happens, and left the decisions with people.

The Model Hardware Standard
Building your own stops being a sensible option once AI is expected to act on equipment, and the third talk addressed exactly that step, which Anthropic says AI cannot yet take: running a process safely and at scale on physical equipment. A typical cell mixes PLCs, robot controllers, vision cameras, weld controllers and torque tools from different vendors, each reached through its own protocol, and Anthropic put the timing problem plainly: a machine acts in about a millisecond, a person needs about 250 milliseconds to react, and a physical mistake cannot be rolled back.
The Model Hardware Standard proposes one driver per device, sitting on top of existing protocols such as OPC UA, EtherNet/IP and GigE Vision, which describes what an agent can read, set and run on that device. Limits are enforced inside the driver, so a request outside them is refused, and the equipment keeps its own logic and safety systems. The standard began as a project between Anthropic and HHMI Janelia Research Campus and is in limited research preview, with access by application and a stated aim of testing the standard and building safety evaluations before it is open-sourced.
The strongest evidence in the talk came from a laboratory setting at QuEra. There, Claude wrote a controller overnight that recovers a dropped laser lock in 0.9 to 14 seconds with a 99.3% success rate, against roughly 150 seconds and 58% for a hand-written script that had taken two to three months to build. The design choice behind that result matters more than the numbers, because the agent wrote ordinary code and then left the loop, so no AI is running at the moment the equipment acts. The plant walkthrough that followed used a fictional plant with simulated data, and Anthropic's recommended starting point for a manufacturer is read-only access on a single cell or bench.
The plugin and the Infor connector will make paperwork faster for any manufacturer whose records already sit in systems Claude can read, and the businesses that gain most from them will be those that have already designed the process the tools slot into. Equipment is a different matter, because a plant cannot afford to invent its own rules for what an agent may do to a robot or a press, and a shared standard with limits enforced in the driver is work that has to come from the industry and its suppliers. The least finished part of Anthropic's announcement may therefore be the one that matters most to manufacturers over the next few years.

Cactix Editorial Team
The Cactix Editorial Team is a crew of curious minds who write, edit, and shape ideas across marketing, communications, and the web. We care about clarity, substance, and telling stories that move people and businesses forward.
Resources


