Why someone still signs off the first part
CNC programming happens at a screen, so more of it sits inside AI’s reach than most production work. CAM software already generates toolpaths for common features, and language models can draft and edit G-code. That is why shops keep asking whether AI will replace CNC programmers. The honest answer is task erosion rather than a job disappearing: the drafting speeds up, the responsibility stays with a person.
Look at what the work involves. A programmer reads a part drawing or CAD model, decides how the part will be held, picks tooling, sets feeds and speeds, writes and posts the program, then simulates it to catch collisions. None of that is finished until the first article is cut and measured. Proving out a program at the machine is where the plan meets a real fixture, a worn insert and one machine’s quirks.
Our Can AI do it? figure for this job is 45 out of 100, the share of task time software can handle today. The share left over is small in volume and large in consequence. A bad post-processor output crashes a spindle; a tolerance call made wrong scraps a batch. The US Bureau of Labor Statistics counts about 28,500 people in this occupation, with median pay of $68,120 and projected growth of 5.9% from 2025 to 2035 (BLS).
What AI handles, what it assists, and what it leaves
Work AI can do on its own accounts for 13% of task time here. That is the repeatable end: generating toolpaths for standard pockets, holes and profiles from a clean model, and running a program through a post-processor for a known control. These are rule-bound steps with a checkable output, which is why software moved into them first.
Assisted work accounts for 71% of task time. Here the tool suggests and the programmer decides: proposing cutting tools, feeds and speeds for a material, flagging collisions in simulation, drafting setup sheets and operator notes, and rewriting sections of a program after a test cut. The suggestion is often right. Confirming it against the machine on the floor is still a judgment call.
Tasks that need a person come to 16% of task time. Proving out the first article sits there, along with deciding how an awkward casting gets fixtured, judging whether a surface finish passes, and coaching the operator who will run the job for the next three shifts. Those tasks carry the cost of being wrong.
What has actually been tested
No head-to-head test of AI against qualified CNC programmers has been graded for this occupation yet. Our Is it better than a person? evidence grade is D, which means not measured, so we publish no parity number at all. Vendor claims about automated programming are not the same thing as a measured result.
What would settle it is straightforward to describe. Run the same set of parts on the same machines, half programmed by software and half by experienced programmers, then compare cycle time, tool life, first-article pass rate and scrap. Add shop records showing how many AI-drafted programs reach the spindle without edits. Until figures like that exist, the grade stays where it is, and the way we score each question keeps the number blank instead of guessing.
When this could change
Most likely between 2044 and 2055 (8 in 10 of our scenarios). The chart above shows the spread, and the replacement-year method explains how that window is built.
Two things could pull it earlier. First, this job needs no robot: the robotics panel on this page puts the physical share of the work near the bottom and lists no hardware tier, so progress depends on software alone. Second, the cost panel shows AI tooling priced far below a programmer’s annual cost, which gives shops a reason to trial automated programming on simple parts.
Two things hold it back. Liability is the first: nobody signs off a crash risk on a $400,000 machine from a simulation alone. The second is fragmentation. Shops run mixed fleets, legacy post-processors, homegrown tooling libraries and low-volume or one-off parts, so a model trained on clean geometry meets messy reality fast. Our tracker of AI mentions in job postings shows how quickly employers start asking for these skills by name.
How to stay needed in a CNC programming job
Lean into the parts of the job that carry risk and context. Own the prove-out: be the person who takes a program from simulation to a measured first article. Own fixturing and workholding for difficult parts, where geometry, access and rigidity have to be traded off. Own operator handover, so the setup sheet, the tool list and the training match what the machine actually does.
Two skills compound from there. One is metrology and quality judgment: reading GD&T, running a CMM, and knowing when a part passes. The other is using CAM automation as a reviewer, not a typist, which means checking machine-generated toolpaths for tool engagement, chatter risk and tool life rather than retyping code by hand.
What to do: take one repeat part, let the software generate the program, and log every edit you make before it runs clean.
Nearby roles are worth a look if you want to see how the task mix shifts. Computer Numerically Controlled Tool Operators sit closest, with more time at the machine. Machinists and Tool and Die Makers carry more hands-on fabrication. You can put any two of them side by side on our compare any two jobs page, or browse the rest of the metal and plastic workers family and the wider manufacturing sector.