How Natural Language Input Might Change DfAM and CNC

In manufacturing simulation software, menus and dialog boxes limit possible inputs by design, to prevent setups that could lead to printing or machining failures. But they also intimidate and frustrate new users unfamiliar with the user interfaces and the logic behind the constraints. 

With natural language prompts, many of these hindrances—and the safeguards—are expected to disappear. In this article, we examine the integration of large language models (LLMs) in design for additive manufacturing (DfAM) and computer-aided manufacturing (CAM) software to understand the new possibilities they usher in, along with the concerns they raise.

Natural Language Input is Not a Fad 

“Natural language input in software is becoming a trend. It’s not just an experiment,” says Bart Van der Schueren, chief technology and strategy officer, Materialise. “I’ve been very impressed with how well natural language models can read and interpret our APIs [application programming interfaces].”

Founded in 1990, Materialise offers 3D printing services and software aimed at aerospace, automotive, healthcare, and other industries. Its print preparation software Materialise Magics allows you to prepare, repair, and analyze complex parts for 3D printing, including lattice-rich geometries. It has APIs for automation and integration into broader CAD-to-print workflows.

“I’ve seen how you can use a GitHub copilot with natural language support to write Python code to create a shrink-wrap around a part,” says Van der Schueren. “It might not be doing the job in the most efficient way, but it’s making the correct API calls and writing executable code.” 

Shrink-wrapping, a print preparation step, involves creating a watertight outer shell for a complex part destined for printing. In modern additive manufacturing (AM) software such as Materialise Magics, the process is almost completely automated, driven by algorithms. 

The software can instantly analyze the uploaded part and generate the appropriate setup and estimates. Image courtesy: Toolpath

“In the future, I expect AI-powered copilots will not only propose code but can take action,” says Van der Schueren. “But you need a user with domain knowledge to ask the right questions to get meaningful feedback.” 

Teaching the Chatbots to Speak Manufacturing

Consumer-friendly chatbots like ChatGPT are well-trained in everyday use of natural language, but chatbots for DfAM or CAM software need an additional layer of domain-specific knowledge. Terms like pressure, load, stress, walls, and support mean specific input parameters and geometric features in manufacturing software, quite different from how they might be used in everyday conversation. 

“In our software like CO-AM Brix, the scripting and workflow setup could be increasingly driven by AI. And with our software Mimics, where users create macros to combine certain workflows, large language models would make the job a lot easier. But to do this reliably, you need an orchestration layer around the API, something along the lines of an MCP-style [Model Context Protocol] server to provide the right context and guardrails,” says Van der Schueren. 

Materialise CO-AM Brix is low-code, node-based automation software. In November 2025, Materialise launched CO-AM Brix as part of its open, secure software ecosystem called CO-AM. Materialise Mimics facilitates medical operation planning, analysis, and execution using 3D-printed replicas of anatomical parts. 

The introduction of natural language input is not the death knell for the familiar menu bars and dialog boxes, Van der Schueren assures. Don’t expect a chat window with a microphone icon to become the entirety of your DfAM or CAM software UI. “In many cases, you will still need to dive into the code and change some lines. Your engineering skills are still very important, but LLMs will take on a role of growing importance,” he says.

The integration of LLMs and documenting how they interact with Materialise’s APIs, Van der Schueren believes, is too important to leave to third-party developers. “In a sense, the AI is reading the documentation we provided. It’s our proper documentation that allows the AI to do the job. So if we let others do it, there’s a risk that the documentation is not correct,” he says. 

Natural language input makes manufacturing software easier to use, but it also places new burdens on the user. If you ask a DfAM software to reduce the thickness of a part’s wall, the software will most likely obey. “Then it’s up to you as an expert to know, with the machine you have, what the minimal wall thickness is [so] you can build,” says Van der Schueren. 

A CAM System as a Chessboard

In 2024, Al Whatmough, a veteran of the mechanical CAD software industry, joined the startup Toolpath as CEO. Trained in manufacturing rules, Toolpath’s AI can scan the geometry of a part and recommend the tools appropriate for the job. Whatmough compared automated CNC programming to playing chess. 

“A rook can only move in a certain direction. A knight can only move a certain way. Similarly, we have some default CNC rules baked into the engine,” he says. Based on these known rules, Toolpath can scan a part and tell the user, “This part cannot be machined with the tools you have. This part can be machined if you buy these tools,” he explains. 

To make these assessments accurately, Toolpath needs to know the equipment at the user’s disposal, its strength and limitations. If CNC were a game of chess, the AI needs to know the pieces in the player’s hands—pawns, knights, and rooks—and their positions on the board to offer a winning strategy. And that’s where Whatmough is looking to employ natural language processing. At the time of the interview, he was in the thick of a new feature’s launch. 

“LLMs have gone far beyond text. They can also interpret images. So natural language can help users say, these are the parts I’ve programmed in the past, and these are the pictures of the tools in my shop. So the AI can discover the tools the users have and apply the known rules,” he says.

He thinks of computer-digestible images as part of the natural language of CNC programming in the AI-driven workflow. Therefore, for him, the focus is less on ChatGPT-style text input, but more on 2D and 3D visuals the AI can decode and catalog. He pointed out users would have different preferences, which need to become part of the AI’s rules.

“One user might prefer to use the end mill on his machine; another might prefer to use the end mill in a shop. One might be against borrowing the necessary end mill from the shop; another might not care. One might prefer to buy the tool he needs from Kennametal; for another, it might be ISCAR,” notes Whatmough.

Currently, setting these priorities and preferences in CNC software would require the user to dig several menus deep into the dialog boxes, or write programming code. But Whatmough feels “describing these in natural language, or by dragging and dropping images would make more sense.”

Whatmough believes, in the foreseeable future, AI and natural language will make complicated CNC software easier to learn and use. But the software’s knowledge of the physical equipment and layout is a critical component that makes the AI more powerful. Without that, the AI will be cornered and checked.