Every new program arrives with a model and a launch date, and the work instructions are the last thing ready. They wait on the one or two engineers who know the product well enough to write them, and those engineers spend their hours rebuilding what engineering already created. They retype torque values from the model, capture and label a view for every step, and redo all of it when the design changes.
AI looks like the way out, and for good reason. A large language model (LLM) turns rough notes into a clear step in seconds and gives a procedure written by five people one consistent voice. Put it in front of a real work instruction, though, and the questions start.
What about the model? A step has to name the right bolt the way the parts list names it and carry the torque value engineering attached to it. That information lives in the CAD model, in the assembly hierarchy, and in the product and manufacturing information (PMI) on each feature.
What about the sequence? Before anyone writes step one, someone works out which subassembly goes in first, which fastener disappears once the housing closes, and which check catches the mistake people actually make on this build.
What about the visuals? AI won't take the photographs for you. Someone still has to capture the screenshot with the right components showing and the housing out of the way, shoot the photo on the floor, or record a technician doing the job.
What about your format? The instruction has to look like every other one in your library, with the warning where your warnings go and the parts list laid out your way.
What about review? A draft becomes a work instruction when an engineer has checked it and released it under version control, and the floor builds to the released version.
Put those together and the real problem shows. The model sits in Product Lifecycle Management (PLM) or on a file share, the screenshots in a folder, the video on someone's phone, the format in a Word template, and the review in email. You build each patch by hand, feed the patches to an AI, and it hands back one more file to manage. AI can help reassemble the patches, but you still made every one of them, and that's where the hours were going in the first place.
A work instruction platform is the one place where all of those sources come together. The CAD model comes in from PLM with its structure and data intact, your PDFs and videos come in beside it, and the instruction is built, formatted, reviewed, and published in the same environment. The visuals come from the model itself, so a view is something you set up rather than something you capture.
That's also where AI belongs. AI helps manufacturing engineers create work instructions faster when it works inside the place that already holds the model, the visuals, the format, and the review, because it can reach everything the instruction needs.
In Canvas Envision that place is Envision Creator, the authoring application, with Envision Portal holding what gets published and controlling who can open it. Envision Creator imports more than 30 CAD formats, from SolidWorks, CATIA, Creo, and NX to STEP, and carries the assembly structure, part metadata, and PMI through with a secure, lightweight version of the geometry. You turn the assembly to the angle the step is done from, hide the housing, explode the parts, cut a section to expose a hidden fastener, or animate the motion, and labels come straight from the part names engineering gave them. Nobody takes a screenshot. The manual steps are still there, but now they happen in one place with the model, the documents, and the layout all within reach, and that's where AI can finally help with them.
Inside the platform, AI has everything the instruction needs, and that's only the start. It also has to understand what you asked for, what's in front of it, and what the result is for. It has to know when a job needs something it can't do alone, such as reading geometry, breaking a video into steps, or pulling procedure text from a PDF, and call on that capability. When the request is unclear, it has to ask what kind of document, for whom, and to which standard, rather than guessing. An AI without guardrails takes whatever is in the prompt and produces something plausible, whether or not it's the document you needed. An AI with guardrails checks what it's been given, asks for what's missing, and works within the rules you've set before it produces anything.
Evie, the AI built into Canvas Envision, is built on those guardrails. It interprets the request, breaks the job into the work that needs doing, and routes each piece to specialist sub-agents for 3D, video, and images. In its AI Agents mode, a workflow asks for the source, the intent, the audience, and the standard before anything is generated. Evie recommends an assembly sequence from the model and any context you add, you review and adjust it, and Evie builds the instruction from the sequence you settle on. Evie is trained on the evCreate SDK, the programmatic surface of Envision Creator, so what it produces lands as real content in the document.
Evie also works the 3D content directly. It's trained on the operations an author performs to build a visual, hiding and showing parts, moving them, coloring them, and labeling them, which is the work an engineer would otherwise do by hand for every step. Ask for the view a step needs and Evie sets it up on the model, so the visual comes from engineering data instead of a screen capture. A general-purpose LLM has none of that training, and it's what lets Evie produce the visuals as well as the words.
The floor reads every instruction faster when they all look the same, and auditors expect them to. Spelling out the format in a prompt holds until a second author writes the prompt differently, or the standard changes and every saved prompt has to change with it.
A sturdier approach defines the format visually, in the layout itself. Each region of the page carries what belongs there, so the AI can read where a step, a warning, or a parts list goes, and when the standard changes you edit the layout rather than a set of prompts.
Envision Creator does this with master layouts, the templates behind every page, and the prompt objects placed inside them. A prompt object is an instruction carried in the template that Evie reads when the template is applied. It tells Evie what belongs in each area and what it can and can't do there, so the guardrails travel with the template instead of living in somebody's prompt. Change the master layout and every page built on it follows.
The work that built the patchwork stops. Nobody retypes values from the model, captures and crops screenshots, rebuilds the format in a Word template, or chases review through email. The model, the source documents, the views, the layout, and the review sit together in Envision Creator, and Envision Portal holds every published version and who can open it, so there are no loose pieces left to manage.
Evie works on top of all of it. It reads the actual model and source documents, follows the prompt objects in your layouts, and runs in the environment it was trained on, while your expertise goes into the sequence, the checks, and the review. In our internal tests, Evie turned a 20-page PDF into a structured draft in under 10 minutes against 6 hours by hand, a 10-minute video into a step-by-step draft in about 5 minutes against 2 to 4 hours, and a 50-part assembly into an instruction in minutes against 2 days.
In a patchwork, change isn't impossible, but it's exhausting. Every revision means finding each affected screenshot, retaking it, retyping the values that moved, and reassembling the document, and engineering changes arrive faster than that work can absorb them. So teams cut corners. Instructions stay general, and experienced operators on the floor fill the gaps from what they know.
Those operators are retiring, and the people replacing them need instructions that are clear, accurate, and current from their first day. That makes the speed of change the real test of any authoring process.
In Canvas Envision a view is a camera position on the model, not a screenshot, so when the model changes the view updates with it. Envision Creator carries the change through every page and view in the authoring document that uses the changed part, and Envision Connector brings released changes in from PLM. New parts are where Evie helps most. It helps you place added components where they belong and updates the views that follow so every step stays consistent, in minutes rather than days. You review the update and publish a new version, and the revised instruction reaches the floor only after that review.
AI is already changing how work gets done in engineering, sales, service, and nearly every other department. Manufacturing work instructions ask more of it, because they demand engineering accuracy, constant change, controlled delivery, and the same format every time. AI can meet those demands when it has a home built for them, a work instruction platform that holds the model, the layout, and the review, and that's what Canvas Envision provides.
Explore Evie to see how it builds instructions from a model, a PDF, or a video, or start with the Canvas Envision platform to see how authoring, delivery, and integration fit together.
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.
