An engineer building a work instruction works with their hands. They open the model, expand the assembly tree to find the right subassembly, turn the view to the angle the step is done from, hide the housing, and type the step. An AI can't do any of that by hand. It can't look at the screen to read the model, and it can't click to change the document.
So it needs another way in. It has to read the CAD model as data, and it has to act on the instruction through defined operations that do what an author does with a mouse and keyboard. Evie, the AI built into Canvas Envision, is built on both. The overview is in how AI helps manufacturing engineers create work instructions faster. Underneath, it comes down to how Evie reads the model, how it acts on the document, and where you stay in control.
An AI working from a screenshot or a pasted description sees what the model looks like and none of what it holds. The assembly hierarchy that records which parts belong to which subassembly, the part names engineering assigned, the metadata, and the product and manufacturing information (PMI) never reach it. Read as data, the model gives the AI the same structure an engineer works from in the assembly tree.
Acting on the document needs the same care. Without a defined set of operations, an AI that writes straight into a file leaves you unable to see exactly what it did, repeat it, or fence it in. So the useful design separates the two. The AI reasons about what should happen, and the platform carries it out through a defined set of operations that you can inspect, constrain, and repeat.
It starts with the import. Envision Creator brings in more than 30 CAD formats, from native SolidWorks, CATIA, Creo, and NX files to STEP, and converts each one into a secure, lightweight version of the geometry. The conversion keeps what engineering put into the model, the assembly hierarchy, part names, metadata, and PMI, so Evie works from the engineering data itself without a CAD seat or a screenshot.
Evie reads that structure directly. It uses the hierarchy to know which parts belong to which subassembly and the part names and metadata to know what each part is. For the geometry itself, Evie uses patent-pending techniques that capture the model's positions and analyze them, so it can work out how the parts sit and relate to one another the way an engineer would when studying the assembly. That's what it draws on to recommend an assembly sequence and to set up the view each step needs.
Evie keeps the reasoning and the acting separate. A general-purpose large language model (LLM) does the reasoning. It interprets your request, breaks the job into the work that needs doing, and hands each piece to a specialist sub-agent, one that analyzes the 3D model, one that turns video into steps or segments, or one that processes photos and diagrams. It then assembles what comes back into a plan for the instruction.
The model never touches the document. Every action that lands in an instruction goes through the evCreate SDK, the programmatic surface of Envision Creator, and Evie is trained on that SDK. When Evie decides what the instruction needs, it writes the code that calls the SDK, and that code authors and edits the content in Envision Creator. Placing text, translating it, and formatting it to your standard all happen this way. So do placing an image or a specific segment of a video, building a step, switching a layout, and, on the 3D side, hiding, showing, moving, coloring, and labeling parts to set up the view a step needs.
Because every change is a defined operation, what Evie produces lands in the document as ordinary content. You can see it, edit it, re-prompt to extend it, or delete it, the same as anything an author placed by hand. It goes through the same review in Envision Creator, with comments on the page and a version history for every revision, and nothing reaches the floor until someone publishes it.
That's also where the engineer stays in charge. Evie recommends an assembly sequence from the model and any context you add, such as a video of the job or an existing PDF procedure. You review it, adjust it, and Evie builds the instruction from the sequence you settle on. The fixture that has to move first, the check that catches the mistake people actually make, and the judgment about what a new hire needs are still yours to add.
Yes, and that's what makes the design practical at scale. The evCreate SDK that Evie drives is the same one your developers can program against. Where a job needs a deterministic result, such as ingesting a process plan that has to be handled the same way every time, a program written against the SDK produces it, and it runs alongside Evie on equal footing. A person, a program, or Evie can drive the same document, depending on what the job calls for.
The same operations run in batches, so anything that works on one page of an instruction works on many pages at once.
A template defines where things go, but an AI needs to read that intent, not guess at it. Spelling the rules out in a prompt breaks as soon as a template has something different or unexpected in it.
Envision Creator puts the rules in the template instead. Master layouts, the templates behind every page, carry prompt objects in their regions. A prompt object is an instruction that Evie reads when the template is applied. It says what belongs in that area and what Evie can and can't do there, so a step lands where steps go and a warning lands where warnings go. Change the master layout and the guardrails change with it, on every page built from it.
Evie doesn't run on the consumer AI tools people use on their own computers. It uses secured commercial model deployments inside Amazon Bedrock or Google's Gemini Enterprise Agent Platform, and it can't search the internet. That means the model providers can't use your prompts or outputs to train their models, your data stays inside an enterprise cloud environment rather than a public service, and with no internet access, nothing Evie works on goes out and nothing outside can steer it. Evie itself doesn't train on your data either. Your models, documents, and instructions are used to do the work you ask for and aren't used to train the AI. Organizations that need to can also turn Evie off entirely. Our Security page covers certifications, deployment options, and the rest of the review a security team will want.
The pattern underneath Evie is simple to state. The model thinks, the platform acts, and you decide what ships. That's what lets an AI work on engineering data without becoming a black box, and it's why the same surface serves an engineer, a developer, and an IT reviewer at once.
Request a Demo and watch Evie work on your own CAD model and template.
Garth Coleman is CEO of Canvas Envision, where he is changing how manufacturers capture, share, and execute the knowledge their best people carry. He brings nearly three decades in enterprise software spanning CAD, Product Lifecycle Management (PLM), and 3D communication, including senior leadership at Dassault Systèmes.
Book a demo with your own CAD data. Most teams are authoring in minutes.
