What AI already replaced in programming — and what it still can’t touch in 2026

Casey Holt

Casey Holt

August 27, 2026

What AI already replaced in programming — and what it still can’t touch in 2026

I stopped arguing about whether AI “replaces programmers” sometime last spring. The argument was too large to be true and too useful to vendors. What I can defend, after another year of shipping with models in the editor, is a smaller split: a set of tasks that are already cheaper to draft by machine, and a set of jobs that still fail when you take the person out. 2026 did not dissolve the second set. It made the first set obvious enough that pretending otherwise is nostalgia.

I led platform teams before I started writing about agentic systems. I still read diffs for a living. This is the inventory I would give a hiring manager, a junior, or a CFO who watched a keynote and came back with a headcount theory.

Already cheaper: the work that was always a draft

Boilerplate is gone as a craft. Not “gone” as in it vanished from the repo. Gone as in I will not pay a person to type a repository interface, a DTO, a client stub, or the fifth CRUD handler that looks like the fourth. An agent drafts that in the time it used to take me to find the house style. My job is to refuse the extra abstraction it also drafted because it saw a tutorial.

Test scaffolding is in the same bucket. Generating a table of cases, a factory, a mock that matches a signature — that used to be an afternoon and a sour mood. It is now a first pass I expect. What is not cheaper is deciding which test is a compliment and which test would catch a lie. Models will happily encode the bug. They have been doing that all year. The replacement is the typing, not the judgment about what must stay red if the code is wrong.

Docs that describe what the code already does are cheaper. A README that matches the flags, an OpenAPI sketch, a changelog bullet. Docs that describe why we will not do the obvious thing are not cheaper. Those docs are still a person who remembers the outage.

Mechanical refactors inside a well-tested boundary are cheaper: rename, extract, update call sites the compiler can see. Mechanical refactors across a boundary the compiler cannot see — a queue contract, a stored procedure, a mobile client on an old build — are not. The model does not have the old build. It has confidence.

Regex, SQL that is not the query plan, CSS that is not the layout budget, a migration that is only additive: drafts. I still read them. I do not start them from a blank buffer if I can avoid it.

Empty meeting room whiteboard with erased architecture marks

Already cheaper, with an asterisk: the ticket that was a recipe

A well-specified ticket with an acceptance check is the sweet spot. “Add a timeout, fail this test if it hangs.” “Export the CSV with these columns.” “Bump the library and keep the suite green.” That work has been replaced in the sense that the first implementation is no longer a scarce skill. The scarce skill is writing the ticket so it cannot be satisfied by a lookalike, and reviewing the lookalike.

I will not pretend this is a small asterisk. Most organizations cannot write tickets like that. They write novels, or they write nothing. Models are excellent at novels that sound like tickets. If your process was already fuzzy, AI did not replace the work. It replaced the silence with a diff you now have to argue with.

This is adjacent to the junior queue problem — the pile of small tickets that used to be how people learned the system — and I am leaving that pile for its own argument. Here I only need the craft point: recipe work is cheap to generate and expensive to specify. Teams that invert that (cheap specs, expensive generation) are the ones posting about how AI did not work.

Still untouched: the work that is a decision under uncertainty

Incidents

When production hurts, the job is not “write a patch.” The job is to decide what you are willing to break at 2 a.m., whose story you believe, and when to stop looking. Models will offer a cause. They love the last deploy. They love the log line that matches a blog post. They do not love the cache that is lying or the vendor who is paging you because their other customer melted. I have let an agent draft the timeline after we had a theory. I have not let it pick the theory. The two times we did, we lost an hour.

The “why we won’t”

Architecture that ships is mostly refusal. No, we will not add a second queue. No, we will not “just” put the flag in Redis. No, we will not rewrite this before the audit. A model will generate the yes. Generating the no requires a memory of cost that is not in the repo: the person who left, the budget, the customer who will call, the migration that already failed. You can paste some of that into a prompt. You will not paste all of it, and the model will not tell you which part you forgot.

Novel debugging

If the failure has a stack overflow answer, the model will find a version of it. If the failure is “this only happens for one tenant after a timezone boundary and a retry,” you are back to being a detective. Agents can fetch files. They cannot be curious about the tenant who is not in the fixture. Curiosity is still a human defect, and I mean that as praise.

Taste that has to survive a year

Naming, module cuts, the shape of an API you will hate in twelve months if you get it wrong. Models have median taste. Median taste is how you get a utility package and three slightly different error types. A person who will still be here next year has a reason to care. A model does not live in the codebase. That is not poetry. That is an incentive.

Accountability

Someone still gets the ticket when the money moves twice. Someone still talks to the auditor. Someone still says, in a room, “I merged that.” Until a model can hold that sentence, it has replaced typing, not the role. I am tired of essays that skip this because it sounds like HR. It is the constraint the rest of the stack hangs from.

On-call engineer with a laptop on a kitchen counter at night

The middle that people keep mislabeling

Code review did not get replaced. It got heavier. There is more plausible text to read. If you treat review as optional because “the AI already checked it,” you have replaced a programmer with an unverified generator. That is a staffing choice. It is not a productivity win. I wrote about the review bottleneck after a year of agents; I will not retell the workflow here. The 2026 fact is simpler: review is now part of the job description for everyone, including the people who used to only write.

Pairing did not get replaced. It changed shape. I pair with a model the way I used to pair with a very fast intern who has not been paged. Useful. Not a substitute for the colleague who will tell me my idea is a bad month.

Learning did not get replaced, but the path changed. You can get an explanation of a file in ten seconds. You cannot get the scar that file gave the last person unless they wrote it down. Teams that stopped writing the scar because “the agent will explain the code” are going to pay for that in a year. The code is not the scar.

What I tell people who want a number

There is no honest percentage for “how much of programming is gone.” There is an honest list. Drafting work that was already a pattern: mostly gone as scarce labor. Specifying, reviewing, refusing, debugging the unread case, owning the outcome: not gone. If your week was already 80 percent pattern, your week changed. If your week was already 80 percent refusal and incidents, your keyboard changed and your job did not.

That split is why the headcount stories keep missing. They average those two weeks together and announce a revolution. I have both weeks on my calendar. They do not feel like the same profession. 2026 just made it harder to pretend the pattern week was the whole job. If your planning assumes they are the same, you will cut the people who do the second week and keep a generator for the first. The generator will be fine. The service will not.

If you want a next move, do not “learn AI.” Use the tools on the pattern work and spend the recovered hours on the parts a model still cannot sign for: the incident, the no, the tenant who is not in the fixture, the review you cannot do in a glance. Those hours are the job that is left. They were always the job that mattered. They are just less padded now by typing we can no longer bill as craft.

More articles for you