Arge Bilişim
tren
Data lines from a machine, a barcode and a sensor feed a neural network whose output reaches rising bars

Factory AI is only as good as its shop-floor data

@rgemas · · 5 min read

When a factory talks about an AI project, the questions usually start with the model: ChatGPT, Claude or Gemini? Which is smarter, which is cheaper? The teams building AI tools point to what surrounds the model: the information, the tools and the setup it is given.

The views below come from software development. In each section we look at what they mean in a factory, and how argemas applies them.

The model is one part of the whole

In an interview with David Ondrej, TypeScript educator Matt Pocock compares it to Formula 1:

Everyone's obsessed with the engine of the Formula 1 car whereas in fact the engine is really only a part of the whole system.

Pocock calls the setup around the model (prompts, tools, the working environment) the “harness”, and says it does as much of the work as the model. It is also the part you control: “…you have much more control of the harness than you do the model.”

Pocock also agrees that a better model makes everything better. The model improves over time; the setup around it is yours to build. In the same interview he adds a warning:

If you're hamstringing your model from day one, then you will need a smart model to get the most out of it.

In a factory, what ties the model's hands is usually the data. If downtime reasons go unrecorded, or scrap is logged in one batch at the end of the day, the AI works with missing information whichever model you choose.

The bottleneck is infrastructure

At Anthropic's Code with Claude event, Jess Yan and Lance Martin summed up what they see from developers putting AI agents into production:

We're seeing that the bottleneck is increasingly infrastructure and not intelligence.

In the user research shared in the same talk, half of developers named infrastructure as their number one blocker to production. Part of that infrastructure is access, security and human approval, which we covered in our post on trusting AI in the factory.

In a factory, the first layer of infrastructure is shop-floor data. An MES collects data from machines, operators, quality, materials and energy on the shop floor in real time. In argemas MES, what is produced on which machine, how productively each operator works, and which downtime happened for which reason are recorded second by second. ERP works at an hourly-to-daily level; we explain the difference in our SAP and MES post.

Every AI function runs on specific data

Each argemas function that uses artificial intelligence or machine learning depends on specific data from the shop floor:

The quality of that data sets the quality of the recommendations. If setup times are not recorded, the planning recommendation rests on estimates; if machine data does not arrive, the maintenance model cannot see a breakdown coming.

The right data at the right time

Jess Yan and Lance Martin's talk also stresses the timing of context:

Context at the right time can be incredibly powerful. It's the knowledge that the agent needs to do work. But context at the wrong time can be a huge distraction.

In a factory, that means defining in advance the data each decision needs. In the argemas AI Integration Layer, technical discovery defines each connection's data objects, direction, frequency, authorization, failure behavior and acceptance test. The system of record remains in place. Sources that can be connected include enterprise applications such as ERP, MES and WMS, as well as PLCs, RFID, barcodes, IoT sensors, SCADA and energy monitoring systems.

Check the context first

In her guide to choosing a model and effort level in Claude Code, Anthropic's Lydia Hallie gives this advice:

When Claude gets something wrong, your first instinct shouldn't be to adjust a knob, but to examine the context you have provided. Is your prompt too vague? Is Claude connected to the right tools?

The guide also notes that a larger model helps when a smaller one stays confidently wrong however much context it is given. The order is still clear: context first, then the model. When a forecast misses in a factory, the first questions are the same: Can the AI reach the relevant data? Is the data current? Are the records complete?

The model side can stay flexible. The argemas AI Integration Layer is provider-independent: it works with OpenAI, Anthropic, Google or a local model. Queries are not stored by the provider or used for training. The data collected from the shop floor, and the setup built on it, stay as a lasting investment, and the model can be changed as needs change.

The data behind each number should be visible

When Anthropic introduced live dashboards in Claude, it made the source of every number visible:

Click any number to see the query behind it, or ask Claude to explain it. Each chart shows when its data was last refreshed.

When a manager in a factory asks “Will we meet this month's production target?”, they should also be able to see which data the answer rests on. In the argemas AI Layer, users can inspect the evidence behind a recommendation. Every output is tied to transaction data the user can inspect and to defined sources, with live transactions and approved enterprise sources as the evidence.

Where to start

Before choosing a model, it helps to answer these questions:

  • Which decision do we expect the AI to support?
  • What data does that decision need, and is it being recorded today?
  • How often is the data updated: by the second, or by the day?
  • On whose authority will the AI access that data?

The argemas AI Layer follows the same order: define the decision before selecting the model. For details, see our What is MES?, AI Layer and AI Integration Layer pages.

Sources

Share this post

XLinkedInWhatsApp
All posts