The common answer to “what does AI help with” covers writing, research, image generation, and coding. These answers conflate AI’s capability ceiling with AI’s current real-world effect. This article asks only the latter: what has the 2025–2026 LLM state actually helped with?

This question is worth answering because the “AI saves money” narrative assumes AI works for all knowledge work. That assumption decides who should invest in AI and who should worry about being replaced. Pinning down the scope of the benefit shows who actually gains and who actually loses.

The main answer: AI’s current state really solves one thing — the labor required per unit of software. The magnitude depends on task complexity: routine tasks save a lot, complex tasks save little. The reason is that software is a low-verification-cost domain: AI’s effect is measurable and checkable in software, but in law and medicine verification is expensive and hard to measure.

1. Question: the scope of AI’s current help

“What can AI help with” asks about the capability ceiling. “What has AI’s current state helped with” asks about the real-world effect of 2025–2026 commercial LLMs in actual work. This article asks only the latter.

“Current state” means: the performance of currently commercial LLMs (GPT, Claude, Gemini) in real work. It is not the laboratory ceiling, nor a prediction of the future.

Why limit to the “current state”? Because “what AI will help with in the future” is a prediction that cannot be verified. “What AI’s current state has helped with” has data to check. This article uses only claims with data support, not predictions.

2. Evaluation framework: verification cost determines measurability

AI generates content faster than humans. Generating fast does not mean generating correctly.

Whether AI’s effect in a domain can be measured depends on verification cost. In a low-verification-cost domain, AI’s output can be checked quickly and objectively. In a high-verification-cost domain, AI’s output requires expert judgment to check, which is slow and subjective.

Software has low verification cost. Whether code compiles is objective. Whether tests pass is objective. Whether a bug is fixed is objectively checkable. AI writes a version of code, runs the tests, keeps it if it passes, and sends it back to rewrite if it fails. Verification is automated, and a single test run takes seconds.

A low verification cost also means the effect can be measured with a randomized controlled trial. METR’s trial does exactly this: the same developers, some using AI and some not, compared on completion time. This design is not possible in law or medicine, because those domains have no checker that says “run it once and know if it is right.”

Law, medicine, strategy, and writing have high verification cost. When AI writes a legal opinion, a lawyer’s review takes hours, because there is no automatic checker for whether a legal opinion is correct. Whether a diagnosis is accurate or a strategy is good all require expert judgment, and the judgment itself is contested. For AI’s output in these domains, the saved generation time is offset by verification time.

Conclusion: software is the domain where AI’s current effect is most measurable. This does not mean AI is useless in other domains; it means that in those domains, AI’s effect is hard to measure and easily overestimated.

3. Comparison: AI’s magnitude in software varies by task complexity

AI’s effect in software is real and measurable, but the magnitude varies by task complexity. The table lists the subjects and results of five studies.

StudySubjectResult
GitHub Copilot (Peng et al. 2023) [2]routine programming tasks55% faster completion
McKinsey (via summary) [5]routine vs complex tasks46% saved on routine, <10% on complex
METR 2025 (randomized) [1]experienced OSS developers, their own complex repos19% slower with AI
Microsoft 2026 (via summary) [4]tens of thousands of engineers, large-scale24% more merged PRs
METR 2026 (update) [3]same design as METR 2025from -19% to +18%

The five studies’ results differ because they measure different task complexity, developer experience, and tooling maturity. GitHub Copilot measured routine tasks, so the magnitude is large (55%). METR 2025 measured experienced developers in their own complex repos, so the magnitude is negative (-19%). Microsoft measured the overall effect of a large-scale rollout, so the magnitude is moderate (+24%).

Three observations. First, routine tasks (boilerplate, CRUD, simple features) save a lot, 46–55%. Second, complex tasks (experienced developers in their own familiar repos) save little or even slow down; METR 2025 measured 19% slower[1]. Third, after tooling improved, the effect flipped positive; METR 2026 went from -19% to +18%[3].

So the claim “AI dramatically reduces software labor” holds for routine tasks and does not hold for complex tasks.

4. Judgment: who gains, who loses

The side that gains is the company that uses AI for routine software work. For the same routine software output, it uses fewer engineer-hours. What it saves is “the labor cost of building software,” not “the price of the software product.” What a software product sells for depends on pricing power, not on how many hours it took to build. A team that uses AI for routine features produces the same output with fewer hours; the saved hours become company profit, or the capacity to take on more projects.

The side that loses is the software engineer, especially the newcomer. The unit value of routine work (boilerplate, CRUD, fixing small bugs) is squeezed — these are exactly the parts AI now does well. Senior engineers are temporarily safe on complex work, because complex work has high verification cost and a small AI effect. Newcomers’ learning path (writing boilerplate, fixing small bugs) is taken by AI, reducing their growth opportunities. This is the most hidden impact: it does not fire existing people; it cuts off the next generation’s learning path. For the company, this is a cost reduction. For the industry, this is a narrowing of the talent pipeline.

Applicable conditions: software (low-verification-cost domain), routine tasks.

Limitations:

  • Whether total employment falls depends on software demand (Jevons paradox). If AI makes software easier to build and demand grows with it, total employment may not fall, but “unit value is squeezed” holds.
  • “Saving labor cost” is not “saving money in your pocket.” The labor cost a company saves goes into company profit, not into the engineer’s raise.
  • Other domains (law, medicine) have high verification cost; AI’s effect is hard to measure, and this article does not judge them.

Ending

Back to the question: what has AI’s current state actually helped with? The answer: it helps “companies that use AI for routine software work” save labor cost, at the cost of “the unit value of software engineers’ routine work being squeezed.”

Applicable conditions: software (low-verification-cost domain), routine tasks. Limitations: complex tasks save little, other domains are hard to measure, total employment depends on demand.

Next article: where does the cost a company saves go — to your phone (13% more expensive in 2026).