Every CAD vendor now ships something with "AI" on the box, and most of the coverage is written by people who have never priced a job on a Friday afternoon. I run a small design engineering and signage studio. What follows is what has genuinely landed in the last two years, what it is good for, and which parts a shop our size can use this year without a data scientist on the payroll.
The CAD copilots are real, but they are assistants, not designers
The big CAD packages have all grown an assistant. One turns a text prompt into a rough shape, another translates plain language into commands, a third reads the model and flags problems early. Read the feature lists carefully and a pattern appears: almost all of it is guidance, retrieval and command translation. How do I do this, what does this constraint mean, where is that setting.
That is worth real money when you are the only person in the office who knows a package well, or when you touch a tool twice a year and forget it in between. It is not the software producing a manufacturable part. A generated shape still has no tolerances, no fixing detail and no idea what your router can hold.
The genuinely useful direction is different: models that regenerate because an input changed rather than because someone typed a request. Field-driven and parametric systems where geometry, simulation and outputs update together were already automation before anyone called it AI, and they remain the closest thing to a tool that does engineering work.
The most practical development of the cycle is not a model at all. CAD vendors have started opening their software and project data to agents through standard interfaces, so an assistant can query what is in a project or drive a live session on your machine. For a five-person shop that matters far more than text-to-CAD, because it automates the admin around the engineering, not the engineering itself.
Quoting: the first place AI pays for itself
Pricing fabrication work means reading a drawing, spotting the expensive detail, and remembering what the same part cost last time. That is a document problem, and documents are what these models are best at.
Quoting tools now analyse drawings, match them against past jobs and pull the historical price into the estimate. Others extract tolerances, finish notes and geometric callouts and turn them into a checklist before anyone commits a price.
The value is not the number the tool produces. It is the callout nobody read. Most of the money lost on a fabrication job goes to a tolerance, a finish note or a fire classification that sat on sheet three. A model that reads every sheet every time is cheap insurance, and it fails safely: the worst case is a false alarm.
On the floor: vision, toolpaths and bearings
Machine vision is the most mature of the lot, and also the most capital-hungry. Cloud platforms for training and deploying inspection models have cut the setup of a repeating inspection from months to days, and the case studies are impressive. Note what they describe: multi-site manufacturers with a repeating part and a camera already on the line. Nobody inspects one-off letter trays this way.
CAM is closer to us. Copilots for CAM programming now propose several candidate machining strategies for a feature and let the programmer pick and refine, starting with 2.5- and 3-axis work. The realistic saving is in reaching a workable program fast enough to quote from, and the honest framing from the vendors themselves is that this does not replace the manufacturing engineer.
Nesting is the other floor-level optimisation everyone asks about, and it is largely a solved, non-AI problem that people still get wrong for organisational reasons. I wrote it up separately in six rules that actually cut sheet waste: the sheet is usually lost before the nester runs.
Predictive maintenance is the quiet one. Modern controls watch motor speed and torque during normal cutting and pick up spindle bearing wear, ball screw wear and guide damage without extra sensors. Across a fleet of hundreds of machines that catches a handful of crashes a year; on a single router it will probably never trigger.
Where it still breaks
Three limits, in order of how often they will bite you.
Data. Most enterprise AI pilots show no measurable effect on profit, and the reason is rarely model quality; it is integration and the absence of a defined outcome. The small-studio version is simpler: if your past quotes live in a spreadsheet with inconsistent part names, no tool can find what the last one cost.
Dimensional trust. This is the one that matters for engineering. General-purpose models reading drawing callouts still hallucinate dimensions and tolerances often enough that a small model trained on real drawings beats them by a wide margin, and generating CAD from text is nowhere near a number you would send to a laser. Genuine progress, and still a research topic.
Liability. If a dimension is wrong on a production file, the studio that stamped it is responsible, not the model vendor. No assistant changes who signs the drawing. In practice: AI output never goes downstream unchecked. It reads, summarises, flags and drafts, but a human approves every number that reaches the shop.
What we are actually adopting this year
The sector we sit in is not, by and large, ready. A large share of print and signage businesses use no AI at all, most of the rest limit it to design and colour, and nearly half report no automation of any kind. Three quarters have fewer than fifty employees. That is our sector, and it means the bar for being ahead is low.
Adopt now, at our scale:
- Drawing and specification review. Run every incoming client PDF and drawing set through a model that lists tolerances, finishes, fire classifications and anything unusual, then check it yourself. Cheapest win available.
- Agentic access to your own CAD and project data. Start read-only: let an assistant answer questions about the project before it is allowed to change anything.
- CAD assistants for the packages you use rarely. They earn their keep on the workflow you touch twice a year.
- Quote intake and historical pricing, if your past jobs are structured enough to be searchable. If they are not, fix that first; it is the actual project.
Still enterprise-only, and fine to ignore:
- In-line machine vision inspection. It needs volume, a fixed part and a line.
- Fleet predictive maintenance. The numbers make sense across hundreds of controls, not one.
- Text-to-CAD as a production route. Watch it, do not build on it.
The pattern is the same everywhere. AI is currently good at reading what nobody has time to read, and bad at being accountable for a number. Point it at the first job and keep it away from the second. Where that fits a specific workflow is most of what we do in CAD automation.
