Off the peg.

Manufacturing’s new AI skills come with someone else’s method written in. The connectors are welcome. The judgement has to be the plant’s own.

Macro of embroidery threads woven in a tight basket weave, every colour of the spectrum

On 12 November Anthropic released Claude for Manufacturing. It is a plugin of 17 skills, with connectors to Autodesk Fusion, Infor, Microsoft 365 and Oracle NetSuite. The skills draft quotes, write up non-conformances, rank root causes, help diagnose breakdowns and compile the morning production report. Outputs can be traced back to the source records. The skills are designed so that a person reviews each output before it reaches a customer or supplier. A launch webinar promises to put a plant’s know-how to work.

The direction is right. Engineers spend too much of the week putting together documents that the records already contain. A tool that does that job hands the time back to the shop floor. What matters is what else arrives in the box.

The three layers

A skill is a set of written instructions: what to look for, in what order, and what counts as finished. That makes it a method, and every method has an author. The plugin bundles three kinds of thing that deserve different treatment.

The three layersWill a generic version do? Less at every layer down.
LayerExampleWill a generic version do?
PlumbingConnectors and data joins
NetSuite orders matched to Infor work orders

Yes. A plant gains nothing by writing its own.

PaperworkDocument formats
8D report, RFQ pack, NCR

Mostly. The customer usually sets the form.

The plant’s methodJudgement
What counts as a cause, what to contain, which jobs to quote

No. This is where one plant differs from the next.

The bars restate the three answers, Yes, Mostly and No. They are not measured shares.

The first layer is plumbing and the second is paperwork. The third is the plant’s method, and a generic skill supplies it by default.

A skill file is somebody’s judgement written down, and installing one adopts it.

Ford’s 8D form

The 8D shows how this tends to go. Ford’s powertrain organisation codified it in its 1987 Team Oriented Problem Solving manual. Ford later made Global 8D its own standard, and the method became standard across the automotive supply chain. One company’s problem-solving method became the form that much of the supply chain fills in to satisfy a customer.

The form travelled well. The habit behind it travelled less reliably, and in many plants the 8D is now written at a desk to close a complaint. A skill that drafts the report from NCR records and maintenance logs makes the desk work faster. The cause is still found at the machine, by someone who knows what normal looks like on that machine. An 8D written without a walk to the machine is the old problem with better grammar.

Lidl’s SAP ledger

Lidl has already made a costlier version of this mistake, outside manufacturing. In 2011 it set out to replace Wawi, its in-house merchandise system, with eLWIS, which was built on SAP for Retail. In July 2018, seven years and a reported €500m later, it was reported to have halted the project and gone back to the old system.

The reported point of collision was inventory valuation. Lidl valued stock at purchase price, while SAP’s retail standard assumed retail price. Lidl chose to customise the package rather than change its accounting, and the customisation proved slow and costly.

Lidl’s SAP ledgerTwo ways to value stock, and seven years spent making them one

2011Lidl sets out to replace Wawi, its in-house merchandise system, with eLWIS, built on SAP for Retail.

July 2018Reported halted after seven years and an estimated €500m. Lidl goes back to the old system.

Lidl’s method · stock valued at purchase price

Customise the package rather than change the accounting

SAP for Retail’s standard · stock valued at retail price

As reported. Inventory valuation is the reported point of collision, and the €500m is a reported figure.

Some post-mortems file this under resistance to change. That reading lets the software off. A discounter arguably knows its margin through exactly what each item cost, so Lidl had good reason to defend its valuation. The error came earlier. Lidl bought a system whose built-in method contradicted its own, then spent years bending one to fit the other.

Ask which of a package’s assumptions are your business before the invoice, not after it.

The quiet convergence

Plants that adopt the skills unaltered face two consequences.

The first is homogenisation. If many job shops rank causes with the same instructions, their 8Ds will converge on the same reasoning and the same phrasing. Supplier quality engineers at the large customers will learn to recognise it.

The second matters more. Many small manufacturers have no written method for quoting, containment or fault-finding. It lives with the estimator and the maintenance lead. For them a generic skill beats a blank page, and because no written method competes with it, it will quietly become the method. That is a realistic place to start, and the hazard is staying there. The sensible route is to take a copy of the skill and rewrite it with what the plant actually does.

The expensive part

Writing a skill takes an afternoon. Getting an estimator to explain how a job is really priced takes weeks. So does getting a maintenance lead to explain why one spindle gets attention before another. Most of it surfaces only when a draft is wrong and the expert corrects it. No vendor can ship that in a plugin, which is why the plugin ships the generic version.

One test separates the useful skills from the templates: does the skill encode a decision this plant makes differently from the one down the road? If it does not, it is a template, and templates are cheap.

The connectors are commodity plumbing. A plant’s method has to be captured before anyone can encode it.

The discipline

Manufacturing Engineers know what a plant’s method looks like at the machine, and what it costs when someone else’s is written in over it. The hard part is knowing which of a package’s assumptions the business depends on.

Kaipability works where a plant’s people meet the systems now arriving to read their records. An operator’s own method belongs in the tools before someone else’s is written in for them. Operators adopting these tools can start that conversation with us.

Sources

Q&A

Questions this dispatch answers.

Written to be quoted by AI assistants and search engines. Self-contained answers, verdict first.

What are the three layers of an AI skill for manufacturing?
Connectors and data joins, document formats, and judgement. A generic connector will do: a plant gains nothing by writing its own. A generic document format mostly will: the customer usually sets the form. Generic judgement will not: what counts as a cause, what to contain and which jobs to quote is where one plant differs from the next.
Why did Lidl cancel its SAP project?
Because the package’s built-in method contradicted Lidl’s own. Lidl valued stock at purchase price, while SAP’s retail standard assumed retail price. Lidl customised the package rather than change its accounting, and in July 2018, seven years and a reported €500m after it began, it was reported to have halted eLWIS and gone back to its old system, Wawi.
What is the future of AI skills in manufacturing?
Rewritten, not installed. Generic skills can carry the connectors and the paperwork. The value sits with plants that take a copy and rewrite it with what they actually do, because a skill that encodes no decision the plant makes differently from the one down the road is a template, and templates are cheap.