How Will Generative AI Change Design Work? The Mechanical Design AI Timeline and What Manufacturers Should Prepare Now

Let us start with the conclusion. AI automation of mechanical design is already within technical reach, and roughly three years of preparation time remain. That does not mean designers will lose their jobs. What will actually happen is that the gap between you and the competitors who have mastered AI will widen beyond recovery.

This article summarises a talk given on 10 July 2026 to the Kansai Engineering Administration Committee, titled How Generative AI Will Change Design Work and the Timeline Manufacturers Should Prepare For. Starting from the automation that has already happened in the software industry, it lays out the order in which mechanical design AI will become practical in manufacturing, and the preparations manufacturers and designers should begin now.

What happened first in the software industry

Before discussing AI entering design work, it is worth looking at an industry where automation has already arrived: software development.

A few years ago, competitive programming, where contestants solve hard problems under tight time limits, was considered out of reach for AI. Today generative AI scores alongside the very best humans. At the same time, companies have begun stating publicly that AI writes more than half of the code produced in house. What would have sounded like a joke a few years ago is now ordinary.

Why software first? The reason is simple: a machine could judge the answer automatically. Run the code and the tests immediately tell you whether it is good or bad. In other words, a verification loop, where AI produces output, a machine scores it, and the AI learns, could be run at high speed without human involvement. AI advances first in the domains where such a loop can be built. That principle holds regardless of industry.

The flip side is that domains where a verification loop is hard to build come later. That is precisely why CAD and engineering drawings have long been considered territory AI is bad at.

Why AI has struggled with CAD and drawings

Compared with text or images, design data is hard to score mechanically. Whether a given 3D model is a good design depends on manufacturability, cost, assemblability, and design standards specific to each company. Unlike code, which either passes or fails when executed, checking the answer is hard to automate. If the learning loop does not turn, progress is slow.

There is also far less publicly available 3D data to learn from. Text and images on the internet are close to inexhaustible, whereas corporate 3D models and drawings essentially never leave the company.

That picture is now starting to break down. One reason is learning from video. Approaches that teach AI the physical behaviour of the world through video have advanced, giving AI a foundation for handling three-dimensional shapes and spatial relationships. The other is understanding semantic structure: research has progressed on capturing geometry not as a mere cloud of points or polygons, but in units closer to design intent, such as this is a shaft, this is a mounting face, this hole is for a bolt.

The clearest evidence is the difference between 3D generative AI a year ago and today. A year ago it produced only vaguely plausible lumps; now visible breakdowns in geometry have clearly decreased. The distance between CAD and generative AI is moving from cannot be done yet to will be possible soon.

The mechanical design AI timeline

So when will what become possible? Ordering things by how easily a verification loop can be built, we expect AI-driven design automation to reach design teams in manufacturing on roughly the following timeline.

The key point is that the document and knowledge domain is already practical. Recording design decisions, searching internal standards, and running routine checklists can be supported well by generative AI as it stands today. By contrast, Physical AI, which handles assembly and adjustment on the floor, faces vastly higher costs to acquire training data and is unlikely to take off for some time. Work that is completed inside the digital world changes first is closer to the reality.

Two weaknesses of generative AI, and where people remain

So far we have discussed progress, but generative AI has two structural weaknesses. Those weaknesses are exactly why the designer role remains.

Weakness 1: it cannot get smarter outside the digital world

AI can only learn from information that has been turned into data. This material warps in this season, this supplier can hold this tolerance, this jig gets fine-tuned on the floor every time: most such insight is written down nowhere. Site-specific knowledge that has never been digitised stays outside the reach of AI. Conversely, only the companies that put that knowledge into words and data can teach AI their own design practice. That is why handing down design know-how has suddenly become a serious management issue.

Weakness 2: accuracy will never reach 100 percent in principle

Generative AI assembles its output probabilistically, so it always contains some level of error. In design, a single oversight can lead to a product failure or large-scale rework. As a result, the value of people who can verify AI output actually rises. Judgement and verification, rather than the work itself, become the main arena for designers. We also examine the link between rework cost and design verification in Rework Costs on the Manufacturing Design Floor, and the Arguments Around 3D and 2D Design Verification.

What manufacturers should prepare now

What to do as a company

First, create an environment where generative AI can actually be used. A blanket ban in the name of security will not close the gap three years from now. Deciding which information may be handled, and making the tools usable in daily work, is the starting point.

Second, document your design know-how within a year. To teach AI your own design practice, your judgement criteria first have to exist in written form. That is why you should start without waiting for veteran engineers to retire. A mechanism for making scattered drawings and design documents usable across the organisation can be put in place first, in the form of Drawing and Design Knowledge AI.

Third, do not adopt too many niche AI tools. Tools built for one narrow task may be overtaken by general-purpose AI within six months to a year. Investing in a general-purpose foundation and in organising your own data tends to last. Processes with clear evaluation criteria, such as design verification, are the exception and are well worth automating as a system, as with Automated Design Verification.

What to do as an individual designer

Start with the habit of asking AI first whenever you do not know something. It costs nothing, you can start today, and it produces the largest difference a year from now.

Next, put your design decisions into words. People who can explain in writing why they chose a dimension or a material can give AI the right instructions. Finally, become a designer who can verify. The ability to judge whether a model or drawing produced by AI is sound will become the scarcest skill of all.

Frequently asked questions

Can mechanical design be automated with generative AI?

Partly, yes, already, and full automation is within technical reach. What is practical today is document drafting support, design knowledge search, and automation of routine checks. Practical 3D modelling and partial automation of design verification are roughly two years out, and end-to-end automation including reading 2D drawings roughly three years out.

Why does AI struggle with CAD and drawings?

Because a machine cannot easily judge what the correct design is, so no verification loop for learning could be built. Unlike code, which passes or fails when run, quality depends on manufacturability, cost, in-house standards and many other conditions. Scarce public 3D data was another constraint. With progress in learning from video and in understanding semantic structure, that wall is beginning to fall.

Will AI take away the jobs of designers?

The jobs will not disappear, but their content will change. Generative AI cannot learn undigitised site knowledge, and its accuracy will never reach 100 percent in principle. Designers who know the shop floor and can verify AI output therefore become more valuable. The gap that opens is not between people and AI, but between companies that master AI and those that do not.

Where should manufacturers start?

Two things: create an internal environment where generative AI can be used, then document your design know-how within a year. Rather than rushing to adopt tools specialised for a single task, prioritise the ability to use general-purpose AI and the organisation of your own data. For individuals, the habit of asking AI first is the fastest preparation available at zero cost.

Summary

Generative AI advances first in the domains where a verification loop can be built. What happened in software will reach the world of CAD and drawings after a delay. Roughly three years of preparation time remain. Whether you can use that time to build an environment for generative AI, and to put your design know-how into a form AI can learn from, will decide your competitiveness afterwards.

The reasoning behind each stage, the actual demo examples, and a checklist of what to prepare are covered in detail in the deck.

This article was written by WOGO Inc., a 3D and AI startup founded out of the University of Tokyo, based on the talk How Generative AI Will Change Design Work and the Timeline Manufacturers Should Prepare For, given by WOGO CEO Jingchao Qin at the 530th regular meeting of the Kansai Engineering Administration Committee, AI and Design, held on 10 July 2026.