For years, "AI in manufacturing" was a slide-deck genre: a stock photo of a robot arm, the word disruption, and nothing you could install. Sometime in the last eighteen months that changed. I write code that touches real machines for a living, and 2026 is the first year I can point at AI systems doing useful work on actual shop floors — just mostly not the work the keynotes promised. Here's my honest mid-year read.
The year agents got real
The biggest structural shift is boring-sounding: a protocol. MCP (Model Context Protocol) gave LLMs a standard way to call tools, and that turned "chatbot" into "agent that can query your machines." The corporate players moved fast — Tulip's frontline copilot, Siemens' Industrial Copilot, Bosch's shop-floor agent work all point the same direction: the assistant sits on top of your existing data plumbing and answers questions a human would otherwise dig for. I've written about what MCP is and how to build a server for your own CNCs, and I'll say the quiet part: the DIY version is a weekend project, not a procurement cycle.
The pattern that works
Every deployment I've seen succeed shares one shape: the AI gets read-only access to data that already exists (machine states, alarms, manuals, quality records), and a human keeps the keys to anything that moves. Write access is where pilots go to die.
AI moved into CAD
Autodesk shipped an assistant that can script the Fusion API from plain language, and SolidWorks' Aura push is aiming at the same space. Having used both styles of tool: they are genuinely good at the boilerplate layer — "write me a script that exports every component as STEP" — and genuinely bad at design intent. Which is fine! Boilerplate was the barrier. If AI writes your first Fusion script for you, you still need to know what to ask for and how to check it — the API knowledge just stopped being the gate.
Vision stopped needing 10,000 images
Machine-vision projects used to stall on one question: "do you have thousands of labelled defect images?" Nobody does. The few-shot tooling and vision-language models that matured this year train on dozens to a couple hundred samples, which moves whole categories of inspection from "not feasible" to "two-week project." For dimensional gauging, though, boring rule-based OpenCV still wins on repeatability — deep learning is for judgment defects (scratches, porosity, finish), not measurement.
The quiet winner: local models
The least-hyped trend is the one I see most often in real shops: local LLMs on air-gapped networks. Aerospace and defense suppliers were never going to send machine data or drawings to a cloud API, and in 2026 the open-weight models you can run on a single workstation GPU crossed the "good enough for RAG over manuals and log summarization" line. Security-constrained shops went from excluded from AI to quietly ahead, because their use cases were always retrieval, not creativity.
What still doesn't work
- LLM G-code straight to the spindle. Models write plausible G-code with confident, subtle errors — wrong modal state, wrong offset, wrong direction. Useful as a drafting aid, dangerous as an autopilot.
- Predicting failures from no data. No model rescues you if you never instrumented the machine. Collection first, prediction second.
- Agents with write access. Every horror story I heard this year starts with an agent that could change something. Read-only is not a limitation; it's the design.
- Replacing the person who knows the shop. The AI answers questions; deciding which questions matter is still the job.
What I'd actually do this half
- 1Get your data out of the machines first. FOCAS, MTConnect or OPC-UA — every AI use case downstream depends on this being done.
- 2Stand up RAG over your manuals and quality docs. Highest value-to-risk ratio in the whole space.
- 3Let AI write your automation scripts — CAD macros, report generators — where a human reviews before anything runs.
- 4Keep a human gate on anything that moves steel. Not forever, maybe. But definitely now.
The shops winning with AI in 2026 aren't the ones with the biggest models. They're the ones that spent the last two years making their data reachable.
If you want a sober second opinion on where AI actually fits in your shop — or someone to build the read-only data layer it all depends on — get in touch.
