Claude for Manufacturing drafts quotes and 8D reports from NetSuite and Infor data, with approval before any connector changes
The useful part of Anthropic's manufacturing launch is not the factory. It is the pattern: read from the system of record, draft, link every number to its source, and wait for a human.
What happened
Anthropic released Claude for Manufacturing, a plugin for Team and Enterprise customers. According to Anthropic's product page, it ships ready made skills for quoting, quality, operations, supply chain and design, and connects to Oracle NetSuite, Microsoft 365, Autodesk Fusion and Infor. The Decoder counts 17 skills and says other systems can be brought in through file exports to Microsoft 365.
The tasks are practical: drafting quotes from actual inventory, comparing suppliers, writing quality reports such as NCRs and 8Ds, generating design variants in Fusion, and ranking likely causes when a machine goes down. With the Infor plugin, The Decoder reports, Claude can also order from backup suppliers when deliveries slip and reschedule production, after human approval.
Anthropic's page is direct about control. Vendors' systems remain the system of record. Drafts can link to the drawing, requirement or record behind each figure. On Team and Enterprise plans, owners can require approval before a connector changes anything, or block changes entirely. The Decoder notes it does not control physical machines yet.
My take
You do not need to run a factory to copy this design. Swap NetSuite for HubSpot or GoHighLevel and the same rules make a safer sales or ops agent:
- The CRM or ERP stays the source of truth. The agent reads and proposes. It does not keep its own copy of the data.
- Every figure in a draft links back to the record it came from. Reviewers check sources, not vibes.
- Writes go through an approval gate by default. Read access is wide, write access is earned.
- Drafts say what they do not know. Anthropic's sample 8D marks a root cause as not validated and lists the test that would settle it. That honesty is worth putting in your prompts.
The other lesson is scope. Anthropic shipped skills for specific documents: quotes, 8Ds, supplier requests. Narrow, named jobs with a clear reviewer are where agents earn trust first. Broad "do everything" assistants are where they lose it.
More posts
- Open OCR models now read 1,000 pages for under $0.10, and better serving saved more than switching modelsNov 20, 2026
- Gemini 4 Neon costs $3 in and $15 out per million tokens and tops Zapier's AutomationBench at 56.3%, though it uses more output tokens per taskNov 20, 2026
- Gemini API's Tool Retrieval loads tool schemas only when the model asks for them, so agents with hundreds of tools send fewer input tokensNov 20, 2026
