AI in CAD/CAM: History and Path Forward

The idea of software making machining decisions rather than just recording them goes back much further than the current crop of AI-branded startups. Early numerical-control programming languages like APT, developed at MIT in the 1950s with Air Force backing, were built to formalize how a part's geometry got translated into tool motion, but the actual process planning — which operations to run, in what order, with which tools — stayed almost entirely in the machinist's head. The software calculated; it didn't decide.

That started to change in the 1980s with automatic feature recognition, usually shortened to AFR. Researchers began building systems that could look at a solid model and pick out recognizable manufacturing features — a pocket, a hole pattern, a slot, a chamfer — well enough that the CAM system could suggest a starting toolpath strategy for each one rather than making the programmer define every operation from scratch. Different research groups took different technical approaches to the same problem: some converted the CAD model into a graph and matched patterns within it, some used heuristics tuned to catch features that overlapped or intersected in tricky ways, and others broke the solid down into simpler volumetric chunks and reasoned about those. None of this was "AI" in the sense people mean today, but it was the first real attempt to get software to look at a part the way a process planner would.

Commercial CAM vendors picked the idea up and built entire product lines around it. CAMWorks, first released around 2001, marketed its version under the name Machining Intelligence, and FeatureCAM — originally developed by Engineering Geometry Systems before Delcam acquired it, with Delcam later folded into Autodesk — built its whole identity around automatically recognizing prismatic features and generating operations for them. This feature-based approach is still the backbone of most CAM software sold today: define your stock, let the software extract features, generate an operation plan from a rules-and-templates library, then hand-tune the result. It's a meaningful step up from writing toolpaths by hand, but it works from a library of recognized patterns rather than any deeper understanding of the part as a whole, which is exactly the limitation the newer AI-branded tools are trying to push past.

Parallel to AFR, CAD vendors experimented through the 2000s and 2010s with generative design — using optimization algorithms to produce structural shapes based on load cases and constraints rather than a human sketching them directly, a technique that saw real commercial adoption in tools like Autodesk Fusion and PTC Creo. Generative design lives mostly on the design side of the CAD/CAM boundary, but it matters to this history because it normalized the idea that software could propose engineering solutions a human hadn't thought of, rather than just executing instructions — a mental shift that helped pave the way for the current wave of AI CAM tools, which apply a similar proposal-and-review relationship to the manufacturing side of the process instead of the design side.

The current generation, which is where CloudNC and Toolpath both sit, layers machine learning and physics-based simulation on top of that older feature-recognition foundation rather than replacing it outright. Instead of matching a part against a fixed library of known feature types, these systems are trained on much larger sets of machining outcomes — cutting parameters, tool wear, cycle times, surface finish results — and use that training to generate or score entire machining strategies for a part considered as a whole, not feature by feature. That's a meaningfully different approach from classic AFR, and it's also why this generation of software gets marketed as "AI" in a way that twenty-year-old feature recognition generally didn't, even though the two are clearly part of the same underlying lineage.

CloudNC was founded in London in 2015 by Theo Saville and Chris Emery, at a point when the company's ambitions were considerably broader than CAM software alone. Saville has described the founding motivation as coming out of frustration with how manual and craft-dependent CNC programming still was, even as the rest of manufacturing kept digitizing — a skilled machinist could spend hours or days hand-programming a part that a well-trained algorithm might handle in seconds, and fewer people were entering the trade to do that hand-programming work each year than were retiring out of it. Early on, CloudNC's stated mission was framed around "single-click manufacturing," and the company spent its first several years building not just software but an actual production facility — a 22-machine CNC factory the company still operates and uses to test its own tools against real production work before selling them to anyone else.

That dual identity, part software company and part working machine shop, shaped how CloudNC approached the problem. Rather than trying to sell a from-scratch CAM replacement to an industry that already had deep investments in Fusion, Mastercam, and Siemens NX, the company built CAM Assist as an add-in that plugs into those existing platforms and automates the strategy-generation step within them. CAM Assist first reached the market in 2023, and the company has backing from a mix of financial and strategic investors, including venture firms Atomico and Episode 1 Ventures alongside corporate investment from Autodesk and defense contractor Lockheed Martin — a pairing that reflects the aerospace, defense, and precision-manufacturing customers CloudNC has targeted from early on.

CAM Assist's core pitch is that it looks at a part holistically rather than feature by feature. Classic feature-based machining, the kind built into most CAM software since the AFR era, recognizes an individual feature — a pocket, a hole — and applies a pre-built template or macro to it, which means the software's understanding of a part is really just the sum of whatever separate templates a shop has built up over time. CloudNC's positioning is that CAM Assist instead treats the whole component as one problem, drawing on an internal model of machining physics and shop-floor rules to work out toolpaths and operation sequencing across the entire part at once, in principle catching interactions between features that a purely feature-by-feature system would miss.

In practice, a programmer imports a model into their existing CAM platform, and CAM Assist generates a full machining strategy — operation sequence, toolpaths, and an estimate of cycle time — that the programmer then reviews, adjusts, and posts out to the machine using their existing post-processors, the same way they'd handle a manually built program. CloudNC has extended the underlying capability steadily since the 2023 launch: initial support was limited to 3-axis milling, with 3+2-axis (five-sided, fixed-orientation) programming added in early 2024, and further integrations bringing the tool into Siemens NX and Mastercam on top of its original home in Autodesk Fusion. Beyond pure toolpath generation, CloudNC has built out adjacent features aimed at the quoting and setup side of the job as well — automated cycle-time and cost estimation meant to support faster quoting, a "cutting parameters" engine that sets feeds and speeds from a physics-based model rather than static tables, automatic soft-jaw design for workholding, and pre-toolpath checks that flag missing stock or unsupported geometry before a program gets generated at all.

Toolpath was founded in 2021 by entrepreneur Andy Powell and aerospace engineer and computer scientist Justin Gray, with the company initially operating as Toolpath Labs out of Atlanta, Georgia, and maintaining an R&D machine shop in Bedford Heights, Ohio. Gray, who spent close to two decades working on rocket-related engineering before co-founding the company, built what Toolpath calls its Part Comprehension Engine — the underlying system the company's tools use to analyze a CAD model's geometry, tooling requirements, and machinability. In 2024 the company brought on Al Whatmough, who had spent close to a decade leading the product team behind Autodesk Fusion's CAM capabilities, as CEO, a hire clearly aimed at pairing Toolpath's technical foundation with someone who had already shipped CAM software at large scale.

Toolpath's fundraising history shows a company that's attracted a mix of financial and deeply strategic industry investors. A Leaders Fund-led seed round brought in $10 million alongside participation from former Autodesk CEO Carl Bass and machinist-educator John Saunders, and later strategic rounds added the cutting-tool manufacturer Kennametal and the tooling-software company ModuleWorks as investors — both companies with an obvious commercial interest in AI tools that make CNC programming faster and more accessible. Autodesk itself later made a strategic investment as well, and in early 2025 Toolpath acquired Chatter, a machine-connectivity and shop-floor-monitoring startup, signaling an ambition to extend beyond front-end quoting and programming into closed-loop, connected manufacturing over time.

Where CloudNC's CAM Assist plugs a strategy-generation engine into CAM software a shop already owns, Toolpath positions itself earlier in the workflow, aiming to automate the steps that happen before a part ever reaches full CAM programming. A shop uploads a CAD model, and Toolpath's Part Comprehension Engine analyzes its geometry to flag design-for-manufacturability issues, identify the tooling and machines capable of producing it, and generate a cost and cycle-time estimate — work that traditionally falls to an estimator or senior programmer doing a manual first pass before a job is even quoted. The company frames the underlying problem as a productivity gap rather than purely a skills shortage: with an estimated five million CNC machines in operation worldwide and the number of trained CNC programmers expected to stay roughly flat while machine counts keep growing, Toolpath's stated goal is to let each programmer effectively oversee far more machines than the two or three that's typical today.

From that analysis, Toolpath's CAM Accelerator tool generates a full operation plan and toolpath set using the shop's own tools and parameters, described by the company as evaluating the part's geometry, material, and available tooling the way a chess engine evaluates a board — considering many possible sequences before selecting one intended to balance efficiency, tool life, and safety. A particular emphasis in Toolpath's marketing is setup reduction: because repositioning a part for a new setup is one of the more expensive and error-prone steps in a job, the software specifically tries to minimize the number of setups a plan requires. Rather than trying to compete with Fusion's own CAM environment directly, Toolpath exports its generated plan into Fusion for final review, editing, and posting, so a programmer's existing post-processors and Fusion habits carry over unchanged. Commercially, the platform is cloud-hosted and subscription-based, with a base tier historically priced in the range of $1,500 per year for one user and two connected machines, a price point the company has described as intentionally disruptive relative to the cost structure of legacy CAM licenses.

Both companies, despite different starting points in the workflow, are converging on the same underlying bet: that the shortage of skilled CNC programmers is a bigger long-term constraint on manufacturing capacity than the shortage of machines themselves, and that AI can extend how much output each remaining programmer can produce rather than trying to remove programmers from the loop entirely. That framing matters, because it shapes what these tools are actually built to do — accelerants that sit inside a human-reviewed workflow, not autonomous systems trusted to send a program straight to a machine unsupervised. Every vendor discussed here, including the two profiled above, keeps a human review step in the loop as a matter of both product design and, almost certainly, liability.

The industry-wide limitations are worth stating plainly rather than burying in each company's section, since they apply across the category rather than to any one vendor. Training data for these systems comes from real machining outcomes, which means performance tends to be strongest on the kinds of parts, materials, and operations that are well represented in that training data — typical aerospace and general-precision machining work — and weaker on unusual geometry, exotic materials, or shop-specific practices that fall outside it. Full simultaneous multi-axis machining, arguably the hardest and highest-value programming work to automate, has consistently lagged behind 3-axis and 3+2-axis capability across the category, CloudNC and Toolpath included, suggesting it's a meaningfully harder problem than headline marketing sometimes implies. And because most of the productivity claims in this space come directly from the vendors making the software, shops evaluating these tools are generally best served by running their own trial parts and timing comparisons rather than taking published time-savings figures at face value.

Looking forward, the more interesting trajectory may be less about any single company's roadmap and more about how AI CAM connects to the rest of the manufacturing stack. Toolpath's 2025 acquisition of a shop-floor-monitoring startup, and the broader industry trend toward connecting AI-generated programs with actual machine performance data pulled back from the shop floor, points toward closed-loop systems where a program isn't just generated once from a static model but gets refined using how it actually ran — feeding real cutting results back into the model that generated the plan in the first place. Siemens has already partnered with CloudNC to bring CAM Assist's strategy generation directly into NX, and parallel efforts like Productive Machines' SenseNC, which focuses specifically on toolpath optimization to fight chatter and tool wear using physics simulation, suggest the category is broadening beyond first-pass program generation into ongoing, adaptive process optimization. None of this displaces the need for a skilled machinist's judgment — if anything, the tools that are candid about their current limitations, rather than overselling full autonomy, seem the ones most likely to earn lasting trust on the shop floor.