Takin’ Care of Business

In July we wrote that the atoms are in central Wisconsin. In Own the Means of Production we argued that the region’s manufacturers are defended by things AI can’t download: a made-to-order catalog built over seventy years, a concrete plant’s freight radius, a county’s soil. The bits are cheap. The atoms stay scarce. Our advice was to point AI at the funnel, the one place these companies were weak.

That advice was half the picture. Every plant runs on atoms, but every morning it also runs on a stack of bits that nobody enjoys producing. The production report someone assembles from three exports before the 7:30 meeting. The work order notes a technician typed with gloves on. The nonconformance report that has to cite the right clause of the right drawing. The quote built from a forty-line spreadsheet and a phone photo of a marked-up print. None of that is the moat. All of it decides how fast the moat earns money.

Anthropic now ships a plugin aimed at exactly that layer, called Claude for Manufacturing. We spent a session reading what it does and looking at the plants an hour from our office. This post is what we found. Bachman-Turner Overdrive had the right idea in 1973: some people get up at seven, and somebody has to take care of business.

The useful way to think about it is not “AI for manufacturing.” It is a set of specific jobs that people in a plant already do, each written down as a skill with its own rules. Here is how they group:

Three design choices stand out, and they matter more than the list. First, every number is meant to trace back to its source: material priced at what you actually paid on a named purchase order, a nonconformance cited to the document and clause it violates, a likely cause tied to the specific prior work orders behind it. Second, it proposes and you decide. The data skill classifies every record that fails to match and proposes a fix, and applies nothing without a yes. Third, credentials never enter the conversation. Connections are set up by whoever runs your systems, not pasted into a chat.

Out of the box it connects to Microsoft 365, Oracle NetSuite, Infor M3 and Autodesk Fusion. Anything else goes down what the plugin calls an access ladder: a vendor connector if one exists, then a scoped API wrapper, a read-only database view, and at the bottom a scheduled export file. Keep that ladder in mind. It comes back in section VII.

Greenheck grew from a 1947 sheet-metal shop into the global leader in commercial fans, dampers and ventilation, with more than two thousand people on its Schofield campus and plants beyond it. Its catalog is made to order. That is the moat, and it is also why the morning numbers are hard: when every line runs a different mix every day, “output against plan” is a join across planning, production and maintenance exports that rarely agree on part numbers.

This is the plugin’s most complete workflow. Morning plant pulse runs overnight, unattended. It joins last night’s exports, flags the records that don’t match (and says why: a renamed part, a work order closed in one system and open in another), then writes the report: output against plan by line, OEE with availability, performance and quality shown separately, downtime ranked both by minutes and by how often it happens, and scrap and downtime priced in dollars. It keeps one dashboard per site with 14 and 30 day trends. For each exception that needs a decision, it drafts a specific move, such as a changeover swap or a reschedule, with the consequences traced.

The point isn’t that a computer can make a report. It is that the person who used to make it can walk into the meeting having read it, instead of having assembled it.

Crystal Finishing Systems paints and anodizes aluminum extrusions, and it is growing. It announced an expansion into a vacant factory in the Mosinee industrial park, a $20.9 million project with about 250 new jobs. Its president was clear that it isn’t a move: new jobs, new processes.

New lines have a specific weakness. The maintenance history that tells a veteran technician “this oven did the same thing in 2022” lives at the old plant, partly in the system and partly in people’s heads. The new technicians have neither. Breakdown diagnosis is built for that gap. When a machine goes down it reads the sensor window before the failure, the asset’s maintenance history including the free-text close-out notes technicians actually wrote, photos of what they can see, and the manufacturer’s troubleshooting tree. It ranks likely causes, each with the evidence for it and the cheapest test that would tell them apart. It raises any safety implication before any diagnosis. Then it drafts the maintenance interval change that would stop a repeat, and the downtime lands in the next morning’s report.

Batch mode ranks the machines doing the most damage across every asset, which is the list you want when one plant is training the other.

Kolbe builds premium windows and doors on brand and craft, and was one of five Greater Wausau Chamber members named finalists for the 2026 Wisconsin Manufacturer of the Year awards. Commercial and architectural work is where quoting gets painful. A request can arrive as an architect’s window schedule in a PDF, a spreadsheet export from a bid portal, or a forwarded thread where the glazing requirement changed in the fourth message.

Customer inquiry intake reads all of those and pulls specifications, quantities and dates into one form, with every field traced to the file and page it came from. The detail we like most: it reports what was never stated separately from what was stated two different ways. Those are different problems, and a sales engineer handles them differently. It checks the request against what the plant can actually build before asking the customer anything, drafts the questions back as sendable text, and gives a provisional bid or no-bid that names which unanswered question would change it.

Then quote build prices it: material at a real purchase order price, setup kept separate from run time, margin on its own line, and the promised date checked against actual shop load. When the customer comes back on price, it revises in place and keeps the record of what changed and why.

Wausau Tile has been making architectural precast and terrazzo since 1953; its products have shown up from McDonald’s restaurants to Times Square. County Materials wins on the physics of freight. Both make heavy, specified products where the expensive defect is the one discovered after the truck leaves.

Nonconformance writeup handles the moment of discovery. It records what was found and measured, cites the exact requirement violated to its document and clause, and works out how many pieces are actually suspect and why that scope rather than a narrower one. It finds where every one of them physically is, including the ones already at a customer. It searches past nonconformances at the moment of writing and says whether the last fix held. Where parts are already at the customer, it drafts the escape notification against the contract’s clock. And it refuses to assert a requirement it cannot find, which is the right instinct for a tool that writes fluently.

The longer chain, inspection to CAPA, carries it through ranked candidate causes with evidence for and against, an 8D, and a customer pack that only goes out after you have seen exactly what was redacted.

Gamber-Johnson makes the mounts that hold rugged computers and radios in police cars, forklifts and work trucks, and was another of the 2026 Manufacturer of the Year finalists. A mounting bracket is a classic trade-off: thinner and lighter, or stiffer and heavier, and the answer changes with every vehicle platform.

Design space explore works on an engineer’s existing Autodesk Fusion model. It never designs from scratch and never overwrites the original. First it checks whether the parameters you want to vary actually drive the geometry, and stops rather than generating garbage if they don’t. It confirms the number of variants before making any. Then it builds them, records mass, volume and size for each, and hands the batch to trade study report for the comparison. It does not run the stress analysis itself, and says so.

One honest caveat: this only works if the engineering team uses Fusion. We don’t know what Gamber-Johnson’s engineers run, and many shops this size are on SolidWorks. If that is true, this use case drops off the list.

Here are the five cases side by side, with the column that usually gets left off:

Read the last column. Every case depends on work an AI vendor doesn’t do: turning on a connector, scheduling an export, building a read-only database view, deciding which network the plant-floor historian can talk to. The plugin’s own connector guide says that most of the time a vendor connector already exists and nothing needs to be built, just an administrator turning it on. In a mid-market plant, that administrator often isn’t on the payroll. It is the managed service provider.

In Wausau, one of those firms is The Dirks Group, founded in 1999, with offices in Wausau and Eau Claire and more than four hundred customers. Its clients span manufacturing, financial, legal, government and education, on networks from ten users to five hundred. That is exactly the size of plant this plugin is built for, and exactly the kind of firm that already holds the keys it needs. We haven’t worked with Dirks; we name them because they are the shape of partner this requires, and they are down the road.

There is a security reason this matters, too. In Secret Agent Man we argued that AI agents need the same network zoning IT has used for twenty-five years. A plant is the sharpest version of that problem. The RFQ intake skill reads documents from strangers by design. The breakdown skill wants sensor data from the most sensitive network in the building. Those two should never sit on the same flat network with an open line out. The people who already draw that diagram for the plant should draw this one.

The split we would suggest is simple. The IT partner owns the connectors, the credentials, the network and the exports. The AI work, meaning which reports, which workflows, what the first month should show and how people learn to trust it, belongs to whoever is driving the change. Neither half works without the other.

1. We are guessing from the outside. Every problem in sections II through VI is inferred from public profiles, chamber announcements and news coverage. Greenheck may already have a morning report that runs itself. Kolbe may quote commercial work through dealers in a way that makes intake someone else’s job. The use cases are a starting list for a conversation, not a diagnosis.

2. Your ERP may not be on the list. The built-in connectors are Microsoft 365, NetSuite, Infor M3 and Fusion. Plenty of mid-market plants run Epicor, Plex, SAP or something homegrown from 2004. The access ladder handles that, but the lower rungs are slower to set up and more fragile, and the plugin says so plainly. A scheduled export file works. It also breaks the first time someone renames a column.

3. The real product is clean data, and that is the hard part. Every skill on the list depends on part numbers, work orders and lots that match across systems. The data skill sorts mismatches into eight classes for a reason. Expect the first month to be mostly about fixing keys, and expect a dashboard that looks worse before it looks better. We wrote a whole post, I Can See Clearly Now, about why month two is when good projects get cancelled.

4. This is the alpha question from July. We spent a whole post agreeing with Alex Karp that your proprietary process is the thing to protect. A plugin that reads your maintenance history, your cost history and your customers’ drawings is, by definition, routing that through an AI vendor. Enterprise terms generally say the vendor doesn’t train on it, and the design keeps credentials out of the conversation. That is a real answer. It isn’t the same as owning it yourself, and every plant should decide that deliberately, not by default.

One more, smaller worry. A fluent 8D is easy to nod at. In The Sound of Silence we called that nod debt. Traceability helps only if someone actually checks the trace.

• Pick one report people already build by hand. The morning production report is the usual winner. It runs every day, so you learn fast whether it’s right.
• Find out who owns the exports. If the answer is “our IT provider,” bring them into the first meeting, not the fifth.
• Check which ERP and maintenance systems you run against the four built-in connectors. That one answer decides whether setup takes a day or a month.
• Draw the network before you connect anything. Which skills read outside documents, which read plant-floor data, and which can send anything out.
• Write down what the first month should show, including which numbers will look worse while the data gets cleaned up, before you turn anything on.

This region was built by sawmills and paper mills that turned logs floated down the Wisconsin River into lumber and paper, and it has been turning raw material into finished goods ever since. The atoms in central Wisconsin are still handled. The cheese still ages, the concrete still doesn’t travel, and seventy years of application engineering still can’t be downloaded. What has changed is that the paperwork around the atoms finally has a tool aimed squarely at it.

Somebody still has to take care of business. It just doesn’t have to be done by hand at 6:45 in the morning.

This post sits with Own the Means of Production on why the atoms are the moat, Secret Agent Man on the network zoning agents need, I Can See Clearly Now on surviving month two, and The Sound of Silence on why a fluent report is easy to nod at.