AI Reaches PLCs: Engineering Assistants Now Read Projects and Write Code

Evolution: from chatbots to built-in agents

Developers working in Python, C++, and JavaScript are used to AI assistants. For PLCs, the situation was different. Three years ago, the first experiments looked promising but were practically useless: ChatGPT could generate a PID controller in ST or a TCP transfer block in TwinCAT, but the results had to be checked with suspicion. There wasn't enough context, training examples were scarce, and integrating the output into an existing project was a significant challenge.

Today the approach has changed. Instead of general-purpose models, specialized tools are emerging that understand project architecture, contain platform documentation, and can read a program's structure directly from the IDE. The shift from assistant to agent is concrete: from suggestions to executing a chain of actions — read the project, make a change, compile it, fix any errors found.

Five tools, five approaches

PLC Copilot was the first tool in this category to gain advanced capabilities. It supports Rockwell Studio 5000, Siemens TIA Portal, and AutomationDirect Productivity Suite (CODESYS is on the waitlist). It can explain logic, find connections between tags, prepare documentation, and generate code for import. Pricing: Pro $99/month, Ultra $499/month, a perpetual per-device license at $720 (with $270/year support).

Copia AI/Copilot stands out for its source of context. It works with projects from Copia Source Control and DeviceLink backups, which lets it compare versions and changes, generate code fragments, and convert between LAD and ST. Pricing isn't public — you need to contact Copia's sales team.

The Structured Text assistant (unnamed in the text) focuses on local IDE connections. A Python bridge for CODESYS and a separate application for TwinCAT read project structure, apply changes, and pass compilation results to a cloud agent. It supports CODESYS V3.5 SP18+ and TwinCAT 3.1+ (TIA Portal support is coming soon). It shows changes made and fixes errors based on compiler diagnostics. Pricing: Pro €24/month, Power €49/month, free trial available.

Siemens Eigen Engineering Agent is built directly into TIA Portal V19, V20, and V21. It can generate PLC code, create HMIs, and configure devices. The company claims it can read electrical schematics in XML and AML formats and generate project structure from a machine description. Being embedded in the IDE gives it an integration advantage, but the price is steep: $2,100/year for a single license (it can be assigned to three users, though only one can use it at a time). The trial period is one month.

Beckhoff TwinCAT 3 CoAgent for Engineering is described as an assistant for PLC development, I/O configuration, and HMI work. It promises suggestions based on project structure and access to the Beckhoff Information System documentation. Its status is coming soon; timelines should be confirmed with the vendor.

Other solutions

These five aren't the only ones. Also in development: • Factory Agent from Software Defined Automation • FactoryTalk Design Studio Copilot from Rockwell Automation

All of them offer a similar set of features: logic parsing, documentation, and coding help. Which features are actually available depends on the product and the environment it supports.

Architecture: four components

What an AI tool delivers isn't determined by its interface ("project — query field — answer"), but by what happens underneath. The architecture consists of four parts:

1. Reading the project: extracting blocks, networks, variables, types, calls, and configuration. For text-based ST, this is one task; for graphical LAD or FBD, it's another.

2. Selecting context: finding related blocks, calling programs, data types, and library documentation. Simply feeding the model more text isn't always the right approach. This is where a dependency graph and a RAG (Retrieval-Augmented Generation) approach come in: the model receives the question along with retrieved material.

3. Translating the response: turning the output into a change to an engineering object or an import file while preserving the project's structure.

4. Feedback: compilation, diagnostics, and error fixing. A successful build only confirms syntax — the logic of the new block still has to be verified separately.

The practical value of a tool depends on how well these four stages are implemented, not on how loud its marketing claims are.