On 12 November, Anthropic released Claude for Manufacturing, a set of skills and connectors for the people who run plants. Read from the COO’s chair, it is a good start. It will make engineers, buyers and quality teams faster, and bring Claude into many more corners of the factory. It will not kill the AI startups in manufacturing, and it will not, on its own, change a company’s P&L. This piece explains why, sets out where we think the stack is heading, and what we would do about it as a COO or a CTO.
AI use : 25%.
Ford Madox Brown spent thirteen years on Work, the painting above, which he finished in 1865. Navvies dig up a road in Hampstead while, on the right, two men stand and watch: the writer Thomas Carlyle and the theologian F. D. Maurice. In the catalogue he wrote for its exhibition, Brown called them the brainworkers: they seem idle, he explained, but they are the cause of well-ordained work in others. I know of no better job description for a COO, and it is the right chair from which to read this launch. Claude for Manufacturing hands the people doing the work a better shovel. The COO’s question is a different one: how the work itself is ordered.
What we are talking about
Set the headlines aside and this is what shipped. Claude for Manufacturing is a plugin for Claude’s Team and Enterprise plans. The announcement says seventeen skills; the public repository counts twelve, plus five chains that string them together, such as request-for-quote to quote, or failed inspection to corrective action. There are four connectors: Autodesk Fusion for design, Oracle NetSuite or Infor for the ERP (a plant connects one of the two), and Microsoft 365 for documents and email.
| Area | Skills | What they do |
|---|---|---|
| Sales | Inquiry intake, quote building | Read a customer request, list what is missing, cost it against similar past jobs, draft the quote |
| Quality | Non-conformance write-up, root cause and CAPA | Document a failed inspection, check past reports, rank likely causes, draft the corrective action report |
| Supply chain | Supplier quote request | Prepare a request for quotes and compare the bids that come back |
| Operations | Breakdown diagnosis, daily production report | Rank the likely causes of a stopped machine from sensor exports, repair notes and manuals; report output against target and the cost of downtime |
| Design | Design-space exploration, trade study | Generate variants of a Fusion model, score them against targets, draft the decision memo |
| Set-up | Data consolidation, onboarding, custom connector | Gather the plant’s records and exports, set the plugin up, connect a system that has no connector yet |
The design choices are sound, and a COO should notice them. Every skill that would change something stops for a person’s approval before anything is sent or written. Without any connector, eleven of the twelve skills run on scheduled exports dropped into a folder, which the documentation calls “the designed path, not a fallback”. Writing back into Infor, to release a purchase order or reschedule production, goes through Infor’s own plugin and follows the user’s Infor permissions. Nothing here runs a machine: that is a separate research track, the model–hardware standard. And Anthropic is candid about the status of the skills: a reference implementation, a starting point that each company will need to adapt.
Read as a COO, this is good news. A large share of the week of an estimator, a quality engineer or a buyer goes into gathering information before any judgment is made: the last similar job, the old 8D, the supplier’s last three emails. Anthropic cites a 2025 Manufacturing Leadership Council survey in which 79% of manufacturers still enter data by hand. Twelve well-written procedures, run by a strong model on the company’s own records, will take hours out of those weeks. We also expect the launch to bring Claude into the daily habits of engineering, quality and supply chain teams.
Why it does not kill the startups
What it will not do is kill the AI startups in manufacturing, whatever the hot takes say. Since an Anthropic release sent software stocks sliding in February, each vertical launch from a lab tends to be read as a death sentence for the companies in that vertical. We argued then that the view from the field was more nuanced, and this launch shows why.
Look at what a skill is: a folder of instructions and reference material, the procedure a good engineer would follow, made legible to a model. That is useful. It is also, by construction, the part of the stack anyone can copy, and Anthropic has published these ones for anyone to read. A startup that a folder of instructions can replace was a prompt with a sales team. The companies worth backing were never selling the procedure. They sell what the procedure runs on.
Hours saved are not the P&L
Sit in the COO’s chair for a moment. A COO does not manage hours saved. A COO manages throughput, on-time delivery, cost per unit, working capital, quality escapes, and the time it takes to launch a product. A day a week saved by an estimator moves none of those by itself. Left alone, the time is reabsorbed, as Parkinson would have predicted. It reaches the P&L only when the process around it changes: who approves the quote, which margin floor applies, how fast the plant commits capacity, and what the freed time is spent on.
That is the gap between individual productivity and company productivity, and every general-purpose technology has had to cross it. Electric motors did little for factory productivity until, decades later, plants were redesigned around them: the story of the driveshaft. Skills and connectors are the new motor. The redesign is the COO’s job.
Our reading at OSS: the moat is everything around the model
At OSS Ventures we build and back software companies for industrial operations, and with them we have shipped or co-shipped more than eighty AI systems into factories. Seen from there, the durable value sits in four places, and none of them comes in a plugin.
- Deployment. Getting a system to work inside a real plant, with its exceptions, its shifts, its works council and its twenty-year-old ERP, is most of the work. It is done by engineers in the building, over months, not by a download.
- Context engineering. Deciding what the model sees, when, and in what form. A skill tells Claude how to write an 8D. The context tells it that one press drifts when the shop is cold, that this customer changes its order every Thursday, and which of the 300 pages of design rules apply to this part. In our deployments, around 40% of the context an agent needs exists in no system at all: it is the contextual layer, and it has to be earned.
- Knowledge management, in a graph. What we call the ontology: the 50 to 200 objects that describe a domain (machines, parts, orders, people, rules), the links between them, and the rules of what a good decision looks like, kept in a knowledge graph that models can query and people can inspect and correct.
- Proprietary data, and fine-tuning on it. Years of non-conformities, setters’ notes, sensor histories, quotes won and lost: data no lab has ever seen. It is what models are evaluated against, and what smaller models are fine-tuned on for narrow, high-volume tasks. The dataset is the moat. The fine-tune is how you cash it in, and it has to be redone as base models improve.
This is, as far as we can tell, Palantir’s view too, and Palantir was early and right about it. Its platform treats language models as interchangeable components and puts the value in what it calls the Ontology: a live model of a company’s objects, the links between them, and the actions that people, and now agents, are allowed to take on them. It delivers that through forward-deployed engineers who stay in the client’s building until the system works. The labs are reaching the same conclusion from the other side: they are now building forward-deployed teams of their own.
The model is the intelligence, not the product
The thesis popular in the labs is that the model is the product: each new generation absorbs the software built around the previous one. There is truth in it. Much of what was sold as an AI product in 2023, a chat over a PDF or a meeting summary, is now a feature of the assistants themselves.
We do not buy it for operations. The model is intelligence: an extraordinary input, improving fast and increasingly interchangeable. The product, in a plant, is everything that turns that intelligence into a decision the plant can act on: the enterprise context and its curation, the proprietary data sources and the way they flow, the permissions, and the written rules of what a good decision is here. In the terms of an earlier essay, the model is the heating element; the value is in the thermostat and the dial.
Claude for Manufacturing makes the point by itself: its skills are useful only once fed the plant’s records, its approval gates hand each decision back to a person, and its writes go through the ERP’s own plugin and permissions. Anthropic has shipped a better element and a few good dials, and left the context, the governance and the feedback loop to the customer. That is the right division of labour, and it shows where the hard work is.
A simple test, for a vendor or an internal project: if the next model were ten times better tomorrow, would this be worth more or less? If less, the model was doing the work, and you hold a wrapper. If more, you own something a better model makes more valuable.
Where we think the stack is heading
- Power users, 30 to 50% more productive, under conditions. Engineers, buyers and planners working with Claude or an equivalent, including open-weight models such as GLM that a company can run on its own servers, will use curated skills and connectors to raise their own productivity by 30 to 50%. The conditions matter. In the best-known field experiment, consultants using AI completed 12% more tasks, 25% faster, when the task sat inside what the model does well, and were 19 percentage points less likely to get it right when it sat just outside (Dell’Acqua and colleagues, 2023). Professionals took 40% less time on writing tasks (Noy and Zhang, 2023). Support agents resolved 14% more issues an hour on average, 34% more for novices, and barely more for the most experienced (Brynjolfsson, Li and Raymond). The gain is real where the task is well scoped and the context is right. It belongs to the person, not yet to the company.
- The AI stack becomes everything around the model. First, permissions: which person or agent may read what, and which declarations each may make, from releasing a batch to scrapping a part or placing a purchase order (the argument of The factory as institutional fact). Then context engineering, data management, and the proprietary sources a company does not want to share with any model vendor. This is where engineering effort, and budgets, will move.
- Systems of record shrink in scope and grow in data. The ERP, the MES and the PLM keep the ledger: auditable, deterministic, legally defensible. Fewer people log into them, fewer screens are needed, and much of the process logic moves out. Meanwhile their data capabilities have to grow tenfold, because their main users become agents reading and writing through governed interfaces, which is how Claude reaches NetSuite and Infor in this launch. The stack rearranges; it does not disappear.
- The stack must be balanced and sovereign. Balanced, so that no single vendor, whether a model lab, an ERP vendor or a startup, holds the keys, and the model can be swapped when a better or cheaper one appears. Sovereign, so that the company controls its data, including the new data this era creates: the agents’ decisions and their traces, the corrections people make, the know-how captured from people’s heads. That new data is the most valuable asset most manufacturers will create this decade. It should land in systems they own, which is also why the autonomous enterprise will not be vendor-locked.
- About one manufacturer in eight will rewire, and take a disproportionate share of the margin. Over the next five years, we expect 10 to 15% of manufacturers to rewire their processes, job descriptions and IT specifications around this reality, starting with a few key processes: new product introduction, manufacturing excellence, procurement, supply chain. They will capture a disproportionate share of the margin. The others will be busy with proofs of concept; with chocolate-covered broccoli, big AI budgets poured over failing systems of record for small gains; or with bottom-up experiments that never scale because the architecture underneath is wrong. Early data points the same way. BCG counts 5% of companies as “future-built” for AI, with 1.6 times the EBIT margin of the laggards, while 60% report little or no value (BCG, 2025). In McKinsey’s 2025 State of AI survey, the 6% of “high performers” were about three times as likely as the others to have fundamentally redesigned their workflows.
Put together, layer by layer:
| Layer | What Claude for Manufacturing brings | What is still yours to build, and own |
|---|---|---|
| Intelligence | Claude, on Team and Enterprise plans | The ability to swap models, open-weight ones included, as better or cheaper ones appear |
| Procedures | Twelve skills and five chains, published as a starting point | Your quoting rules, your 8D standard, your design rules, written down and maintained |
| Connections | Four connectors, and scheduled exports for everything else | Live, governed access to the rest: MES, quality system, PLM, maintenance, historians |
| Context and knowledge | Whatever the connectors and exports carry | The graph of your plant: objects, links, rules of good, and the knowledge in people’s heads |
| Permissions | Approval before any change; ERP writes follow ERP permissions | Who, person or agent, may declare what, across every system, with an audit trail |
| Proprietary data | Out of scope | Your datasets, evaluations and fine-tunes, and the decision traces of the new work, kept sovereign |
| Process and roles | Out of scope | The rewired process: job descriptions, decision rights, KPIs, IT specifications |
What it means for a COO and a CTO
If we ran operations or technology in a manufacturing company, this is what we would do.
- Pick three to five step changes, not fifty use cases. The few changes that would alter the business: for instance, new product introduction in half the time, a quote out the same day with its true margin known, procurement that follows every supplier rather than the top twenty, a supply chain that re-plans at every event rather than every month. Write each one down in P&L terms.
- Work out, honestly, what each one takes technically. Spoiler: it is not a set of Claude skills. It is very likely a new data stack, serious context engineering, and a redesign of the process itself. Skills will be part of it, at the top.
- Partner with people who are serious about results. Look for teams that have put systems into production in plants like yours, that will send engineers into the building, and that will accept being paid on outcomes. In MIT’s 2025 study of enterprise AI pilots, buying from specialised vendors and partnering succeeded about twice as often as building alone, as Fortune reported.
- Plan for a quarter of it to fail. If you launch four step changes, budget, staff and communicate as if one will not work. Set the stop criteria before you start, so that killing a project is a decision, not a scandal.
- Scale fast when it works, and change the process. When a step change works, make it the default within a quarter, and rewrite what must change around it: job descriptions, decision rights, KPIs, IT specifications. That is where the P&L moves, and it is the step most companies skip.
One test to keep at hand: if the project succeeded completely, would it change who does which job, who makes which decision, and what the team produces? If not, it may be useful, but it will not show up in the P&L.
Better shovels
We think Claude for Manufacturing will significantly increase the use of Claude by engineers and their colleagues in plants, and create real individual productivity. Every COO should want that for their teams. The prize, productivity at company level and a P&L that looks different, will come through harder efforts: the data, the context, the permissions, and above all the redesign of the work. Brown’s brainworkers would recognise the job. Better shovels are welcome. The work still has to be well ordered.
Let’s build.
