"Move these labels clear of the symbols" sounds simple to a drafter, because the drafter fills in the missing context without thinking about it. An AI system has to work out which labels are meant and which symbols matter. It then has to pick an editing operation and check that the result is still readable. Each of those steps can go wrong, and that's what separates the tools that turn plain English into AutoCAD work.
natural language to AutoCAD commands AI vendors
Several vendors now let you type a request in plain language and have AutoCAD carry it out. They get there in different ways.
CADshift connects an outside AI assistant to AutoCAD. The assistant runs separately, and CADshift gives it tools for reading and changing the open drawing. You talk to the assistant, and it acts on the DWG.
cadssistant lives inside AutoCAD as a palette. You type a request beside the drawing, and it runs AutoCAD commands and AutoLISP to carry it out. The conversation and the drawing sit in the same window.
ArchiLabs Desktop is a Windows application that works with native AutoCAD and Revit files, and it's built for requests that depend on other project documents. You can attach a PDF markup, a spreadsheet, a screenshot or a narrated screen recording, select the elements in question, and describe the change. It starts at $19 a month plus usage credits (pricing).
The differences matter less in a demo than in daily use, where the hard part is usually getting the right context to the tool.
Explanation, execution or code
It helps to separate three kinds of tool that all get described as "AI for AutoCAD."
A chatbot can explain how a command works without ever seeing your drawing. That's useful when you've forgotten the options for ALIGN or want to know why a dimension style isn't updating, but someone still has to do the work.
A connected agent can inspect the current document and run supported actions on it. This is what most people mean by natural language to AutoCAD commands: you describe the edit, and the drawing changes. It's the version a production drafter can use directly, because the drafter stays in the conversation and can answer questions as they come up.
A code generator writes an AutoLISP routine or script for someone else to review and run. That suits a CAD manager building a tool for the team. It's less suitable for a drafter in the middle of a revision, because it adds a review step and asks the user to judge code they didn't write.
The lines blur in practice. cadssistant, for example, generates and runs AutoLISP from its palette, so the drafter gets execution while code does the work underneath. What matters is knowing who is responsible at each handoff. If a routine runs, the person running it should be able to see what it changed, even if they couldn't have written it.
Where plain-English requests break
Most failures come from ambiguity. Ask an agent to move "the left-hand label" in a drawing with two equipment bays, and there are at least two reasonable answers. A good tool responds by asking you to select one, or by listing the candidates it found. A tool that guesses may pick the right label this time, which is worse in the long run, because it teaches you to trust a guess. Selecting the object, or naming it by a tag or layer, removes most of this problem.
Units are the other common trap. "Offset it by 600" means one thing in a millimeter drawing and something very different in a drawing set up in inches, and a drawing's units don't always match what its title block suggests. The safest habit is to state units explicitly and have the agent repeat back the distance it plans to apply. Our DWG units and scale guide explains how to establish a drawing's measurement context before asking for any geometry edit.
Batch edits and generated routines
Repetitive changes are where plain language saves the most time and where weak tools are easiest to spot. Take renumbering a set of equipment blocks. A good result changes the selected blocks, leaves unrelated blocks alone, and reports any selected block that lacks the expected attribute, without skipping it silently or treating it like the others. Run the same request again on the updated drawing, and numbers that are already correct should stay as they are.
Those two behaviors, reporting exceptions and being safe to run twice, separate a routine you can hand to the whole team from one that only works when its author is watching.
Choosing the interaction your team can sustain
The right vendor depends on who will make the requests. If your CAD manager wants to build tools, a code-generating workflow may be the best fit. If production drafters will be asking for edits all day, look for an agent that works on the live drawing, asks when it's unsure, and reports what it changed. If your requests usually start in another document, such as a consultant's markup or an equipment list in Excel, the tool needs to read those files too.
That last case is what ArchiLabs Desktop is designed for. Try it with a few ordinary requests from your own week: one precise edit, one ambiguous edit and one repeated operation. Those three will tell you more than any list of supported commands.
Confirm drawing units and reference dimensions before a CAD batch edit to avoid consistent-looking geometry at the wrong scale.