When manufacturers talk about AI on the factory floor, the conversation usually starts with what workers see. Smart glasses, tablets, sensors, dashboards. But the quality of execution on the floor is determined long before a worker picks up a device. It is determined at the moment someone sits down to turn engineering knowledge into instructions a person can actually follow.
That moment is the least modernized step in most manufacturing operations, and it has quietly become the bottleneck between engineering and execution. My career began almost thirty years ago, first helping bring 3D CAD into the mainstream, then bringing visualization everywhere in engineering, then using 3D as a communication tool to broadcast engineering data across the enterprise. Yet through all of that progress, the instruction that actually reached the factory floor stayed a PDF for twenty years. The reasons why are worth understanding, because for the first time in my career, with the arrival of AI, the technology has caught up to the problem.
The bottleneck nobody budgets for
Consider how a typical work instruction gets made today. A technical writer or manufacturing engineer takes CAD models and drawings, captures screenshots, photographs a prototype, and stitches everything together in a document editor. A single complex procedure can take days or weeks. Every engineering change order restarts some portion of that cycle.
The stakes of this slow process are rising fast. Deloitte and The Manufacturing Institute project that manufacturers could need as many as 3.8 million new employees by 2033, and that roughly half of those skilled positions could go unfilled. At the same time, experienced workers are retiring and taking decades of practical knowledge with them, most of it never formally captured. The people who could write it all down have never had a fast way to do it.
Why a language model alone was never going to fix this
When large language models arrived, it was tempting to believe the authoring problem would solve itself. Point an LLM at your documentation and let it write procedures. But a work instruction is not a block of text. It is a structured, visual, step-based artifact built around 3D geometry, part data, safety requirements, and company standards, and it has to live inside systems that govern how work gets done.
A general-purpose LLM can describe an exploded view. It cannot build one. It cannot classify the parts in an assembly, position a camera, label components, or place the result into a delivery system connected to engineering data. Generating words was never the hard part of this problem.
What actually works is a vertical AI architecture. An LLM serves as the natural language interface, agentic orchestration coordinates specialized tasks, and every one of those tasks is grounded in domain knowledge of manufacturing content, data structures, and workflows. The language model provides the interface. The vertical layers do the work.
What this looks like in practice
Our team built this architecture as Evie, the AI built into Canvas Envision. The details matter here, because they are what separate vertical AI from a chatbot wrapper. Evie is trained on the platform’s authoring SDK, which means it operates directly on the content rather than describing it. Ask for an exploded view and Evie analyzes the assembly structure, calculates part distances and directions, and builds the view.
The same applies at document scale. Evie manipulates existing 3D content across an entire instruction set, recoloring part categories, ghosting background components, and setting camera angles. Multi-step requests execute as coordinated sequences, with orchestration agents handling each specialized operation.
The conversion agents are where decades of trapped knowledge come loose. Purpose-built agents take a training video, a legacy PDF manual, or a 3D model and generate a structured draft of step-by-step instructions. In our experience, assembly instructions for a fifty-part model drop from roughly two days to under an hour, and a ten-minute training video becomes structured instructions in about five minutes instead of a few hours.
Capabilities like these are what I wished for when we were pioneering 3D communication two decades ago. The manual effort they eliminate is the very thing that kept work instructions trapped in the paper era, and watching it fall away is more than satisfying, it is remarkable. Once you watch Evie take legacy content and 3D data, ask you a few questions, and merge it all into a rich, interactive, multi-step work instruction right before your eyes, you cannot unsee it. Work that would take hours or days unfolds in minutes, you wonder where this has been all your life, and nobody who sees it can imagine going back to the old way.
The missing piece was vertical SaaS
There is a reason this problem outlasted twenty years of otherwise remarkable progress. The industry solved 3D authoring long ago, and rich interactive content has been technically possible for decades. What never existed was a governed home for it. Deploying that content meant building custom applications, writing code to host a viewer, and stitching approvals together by hand across disconnected tools.
The missing piece was vertical SaaS, a platform built for this specific workflow from end to end. In digital thread terms, it has to span the first mile, where knowledge is captured from engineers and experienced workers and structured into instructions, and the last mile, where those instructions reach the person doing the work at the moment of execution. It also has to carry
everything in between, the review and approval workflows, revision control, and live connections to engineering source data. Content gets authored, reviewed, approved, and delivered in one governed environment, instead of being assembled in one tool and parked in another for sign-off.
This is exactly what we set out to build with Canvas Envision over the past few years, a governed environment where teams author, review, approve, and deliver in one place, running the entire workflow themselves. Then along came AI to supercharge it. That structure is precisely what makes the workflow ripe for vertical AI, because AI can only augment and automate a process that has somewhere to put its output, rules to follow, an approval path to respect, and a delivery mechanism to feed. The right stack has to exist before the intelligence can matter.
From authoring to delivery and back
Solving authoring is the start, but the goal is authoring to delivery and back. Instructions authored with Evie’s help reach the worker through the platform’s interactive delivery environment, where each step presents live 3D content the worker can rotate and inspect. The same interface captures confirmations, inspection results, and feedback, and that operational data flows back upstream to engineering. The loop closes.
Not every organization is ready for interactive delivery on day one, and that is fine. The same authored content can be saved off as a PDF for teams that need a familiar format while they transition. They still get better instructions produced in a fraction of the time, with the interactive path waiting when they are ready.
The expert stays in charge
None of this removes the human from the process, and it should not. Manufacturing procedures are safety-critical and regulated, and an incorrect torque spec or a skipped warning has real consequences. The AI produces the draft at remarkable speed, and the platform’s built-in review and approval workflow ensures an expert decides what actually reaches the floor. Speed without governance would be a liability, and the two arrive together here.
There is a deeper reason the expert matters, beyond checking for errors. AI systems generate information by analyzing documents, models, and videos. What they cannot generate is wisdom, the understanding of why a step is done a certain way, which shortcuts are safe, and which are catastrophic. That knowledge lives in the heads of experienced engineers and technicians, and the AI’s real job is to make capturing it fast enough to happen before those experts walk out the door.
How faster authoring shows up on the floor
The payoff arrives downstream, at the point of execution. When authoring takes minutes instead of days, instructions keep pace with engineering changes, so the worker is always executing against the current revision rather than a stale printout. When conversion is cheap, the backlog of undocumented procedures finally gets documented, and tribal knowledge becomes structured guidance available to everyone.
Adoption research points the same direction. PwC and The Manufacturing Institute found that AI adoption in manufacturing hinges on the readiness of frontline leaders, and BCG’s research on AI at work shows regular AI use among frontline employees stalling at about half the workforce. Putting AI in the authoring seat sidesteps that gap entirely. The frontline worker does not need to adopt anything new, they simply receive better instructions, faster, with a feedback channel that makes their experience count.
The manufacturing industry has spent two decades wiring up machines, systems, and data streams, yet the connection that needed the most work was always the one between what engineering knows and what the workforce can act on. What finally closes it is not any single technology but a convergence, the best of large language models, deep manufacturing domain knowledge, and a mature vertical SaaS platform all arriving at the same moment.
That convergence is what we built with Canvas Envision and Evie, and it means no manufacturer has to cobble these pieces together from general-purpose parts. After twenty years, this problem is no longer waiting to be solved, and that is worth celebrating.
