AI CAD Modeling in Siemens NX

AI CAD Modeling in Siemens NX: Editable Geometry, Built in Your Live Session

Most AI CAD tools hand back a mesh or a body without history. In a Siemens NX shop, useful AI CAD modeling looks different: sketches, extrudes, holes and blends in the Part Navigator, built in the session that is already open, each one a feature you can edit and a step you can undo.

This page is about the modeling layer itself: which NX operations an AI can run in your session today, how the result stays editable, and how we prove that on your parts. A new part proposed from a description is the job of our Geometry Creation Assistant; work across whole part sets runs as reviewable jobs under AI-Assisted Geometry Creation; several agents that plan, build and check together are agentic CAD for Siemens NX. They all rely on tools of this kind, and the agent-based ones belong to our family of NX AI assistants and chatbots.

AI CAD modeling in NX: the short version

• What it is: a catalog of typed NXOpen modeling tools that an AI model calls in the live NX session. The model chooses the operation and its parameters; NX builds the feature.
• What it builds today: sketches with lines, arcs, circles and rectangles on a standard plane, a planar face or a custom plane; extrudes that create, unite, subtract or intersect; simple, counterbored, countersunk and threaded holes; threads, chamfers, edge blends and blocks; and edits to existing features.
• How it stays editable: every result is an ordinary feature in the Part Navigator. Edits reopen the feature in its own builder or change its expressions.
• What keeps it controllable: one undo step per change, automatic rollback when an NX call fails, and every call shown with its arguments, result and runtime.
• What it is not: generative design, one-sentence text-to-CAD, freeform or Class-A surfacing, or a batch generator for hundreds of variants.

Why AI geometry in NX has to be features, not bodies

The first AI CAD tools most engineers try generate a shape from a description and export it as a mesh or a STEP body. Imported into Siemens NX, that body has no history: no sketch to change, no hole feature to resize, no expression a template can drive. No NX modeling command turns a body back into a feature history, which is why we treat STEP to parametric in NX as a discipline of its own.

AI CAD modeling inside NX avoids the problem at the source. The AI does not draw geometry; it calls the same builders an engineer reaches through the NX dialogs. What comes out is what an engineer would have built by hand: a sketch, an extrude on top of it, holes placed on a face, a blend on the edges that need one. That is also what makes the result reviewable. An engineer can open each feature and see exactly what the AI decided.

The second difference is context. A tool that works in the open session sees the part as it is now, with its bodies, faces, edges and features. It can add a hole to the face you mean on an existing part instead of starting from an empty file, and changing existing parts is a large share of modeling work in a real NX shop.

What the AI can build in NX today

The catalog is deliberately small and grows project by project. These are the modeling tools in our current NX build, each tested in a running session.

Not in the catalog yet: revolve, patterns and datum planes. They follow the same recipe, and we build them when a project needs them, not before. For the operations that are there, our reference test is a plain one. The AI sketched a rectangle, extruded it to 80 × 50 × 30 mm, placed two Ø12 mm holes, added a 5 mm edge blend and a 3 mm chamfer, and ended with the same feature tree as the part an engineer had built by hand.

How an AI CAD modeling request runs in NX

Take a typical request on an open part: two counterbored holes on the top face, then a blend on the long edges. The engineer asks once; the work follows a fixed pattern.

1. Read the work state. Before the model plans anything, the tools report which part is open and which bodies and sketches it contains.
2. Find the faces. The agent identifies the faces and edges the request refers to, for example the planar face on top.
3. Build. The model calls the hole tool with hole type, diameters, depth and positions. NX’s hole builder creates the holes as one undo step.
4. Look. The agent can capture the part in a standard view, so the result is compared with the request.
5. Edit or undo. A follow-up such as a larger blend radius reopens the blend in its own builder; a wrong step is one entry in the NX undo list.

The same pattern runs whether a designer types the request into a panel or an agent plans it as part of a longer task. Single operations triggered from a panel are what our NX Action Assistant does; the protocol layer underneath is explained in how AI agents read and build geometry in NX.

If your team already has NXOpen code, journals or standard features for recurring geometry, those are the natural next tools, and the model learns to use them from their descriptions, not from retraining. For how we write NXOpen code in general, see our page on NXOpen automation.

Editable means editable: the acceptance test for every tool

Parametric is easy to claim. In a project we check it per tool, on a part that looks like yours, before the tool is released to your designers.

Two honest notes on what editable does not include yet. Sketches are created without constraints or dimensions: the feature history is there and editable, but a fully constrained sketch is still the engineer’s step. And edges and faces are selected as fixed sets, not by rules such as all tangent edges, so a blend does not grow on its own when the topology changes.

AI CAD modeling and its neighbors on this site

Several of our pages are about AI that acts in NX. They describe layers of the same stack, and most projects use more than one of them.

What AI CAD modeling does not do

The tools make an AI quick at operations you can name. They do not make it a designer, and they do not remove the limits of NX or of its API. These are the limits we state in the first meeting.

For new parts and variant series driven from tables or ERP data, parametric part and variant generation from master templates is the deterministic route; for the same edit across many existing parts, see AI geometry manipulation.

Control: undo, rollback and every call in view

An AI that changes geometry has to be easier to reverse than the engineer’s own work, not harder.

• Every change is one undo step. Each modeling tool sets its own visible undo mark in NX; a whole sketch counts as one step.
• Failed calls roll back. If NX rejects a value, the tool undoes its own partial work and reports the NX error, so nothing half-built stays in the part.
• Nothing hidden. Every tool call is shown with its arguments, result and runtime, and every model request is logged.
• Read before write. Listing bodies, faces, edges and sketches changes nothing, so new users can watch the AI read their part before it changes anything.
• Confirmation where you want it. In our demo setup the build tools run directly, because every step is one undo. In production, which tools wait for the engineer’s OK is decided per tool with you before the pilot and built in as part of the project.
• Your permissions underneath. The tools run in the engineer’s own NX session, under the engineer’s license and data permissions.

Deployment, data and hosting options

The modeling tools run in the engineer’s own NX session. NX is busy only while a tool executes, and the designer can keep working between calls. Our current modeling tools are built and tested on NX 2606.

Part files stay where they are. What reaches a model is what a tool returns: feature names, IDs, dimensions and, where you allow it, an image of the part in a standard view. Where the models run is decided per project. Claude models are available directly from Anthropic and through cloud providers with EU regions, and models hosted in Germany or on premises can be connected to the same tool layer as part of a project; which modeling tasks a smaller local model handles well is something we test on your parts before we promise it. The background is in our guide to sovereign agentic AI in European manufacturing.

How we build AI CAD modeling for your NX

Most of the effort is not in the AI. It is in the tools: which operations your designers repeat, on which templates, and which part states must never be touched. That comes first.

1. Quick Scan with on-site day. Up to two workshop days, five days in total: which modeling steps your designers repeat, on which parts and templates; the result is a short list of operations and one task to start with.
2. Build the tools. Each chosen operation becomes a typed tool, built on a representative part.
3. Read tools first. The read tools go live first, so the team sees the AI read their parts correctly.
4. Acceptance test. Every build tool passes the six checks above on your parts before it is released.
5. Pilot. A small group of designers on real parts, measured on tasks completed, undos and corrections.
6. Rollout. More operations and more users, with the same tools serving chat panels and agents.

Start with one modeling task in your NX

We do not start with a catalog of forty operations. We start with the modeling steps your team repeats most, the journals or NXOpen code you already have for them, and a part that looks like your real ones.