Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →AI can help produce code and complete some software tasks faster, but neither a code change nor a higher task count proves that the engineering work is finished. The result still has to solve the right problem, fit the system around it, and be checked for correctness and maintainability. Current studies show that productivity effects vary by setting and by what researchers measure—not that engineering has become unnecessary.
Does AI actually make software engineers more productive?
Sometimes, on some measures. The evidence does not support one universal productivity number: studies examine different developers, tools, work environments, and outcomes. In particular, time per task is not the same measure as the number of tasks completed, and neither automatically measures all the effort required to deliver and maintain software.
| Evidence | Setting and participants | What was measured | Finding and qualification |
|---|---|---|---|
| DORA, 2025 | Global research synthesis drawing on more than 100 hours of qualitative research and nearly 5,000 technology-professional survey responses | How AI fits into software-development organizations | DORA describes AI as an amplifier of existing organizational strengths and weaknesses. This is a synthesis, not a universal causal estimate of productivity. |
| METR, early 2025 | Randomized trial involving experienced contributors working in established open-source repositories | Time to complete assigned tasks with AI assistance | Tasks took 19% longer with AI use in this study; the confidence interval ranged from 2% to 39% longer. The result applies to these participants, tools, tasks, and period. |
| Cui, Demirer, Jaffe, Musolff, Peng, and Salz, Management Science, 2026 | Three randomized workplace experiments at Microsoft, Accenture, and an anonymous Fortune 100 company; 4,867 developers combined | Completed tasks after access to coding assistants | The combined analysis found a 26.08% increase in completed tasks, with a standard error of 10.3%. Less experienced developers had higher adoption rates and greater gains. This is not an estimate of time saved on every engineering activity. |
| Microsoft Research, 2025 | Randomized workplace study of generative AI coding tools | Participants’ reported experience using coding tools | 84% of participants reported positive changes in daily work practices. Sustained use increased perceived usefulness and enjoyment, while views about the trustworthiness of generated code did not change. |
These findings should not be averaged into a single “AI productivity” figure. METR measured time per task in a specific open-source setting; the workplace experiments measured completed tasks in company deployments. The populations and outcomes differ, so a slower result in one context does not invalidate a task-count gain in another.
If AI writes the code, what work is left for the engineer?
Writing code is one part of engineering, not a substitute for deciding what should be built or establishing that a change works in its intended context. An assistant may generate a patch quickly, but that output alone does not show whether the change meets the need, respects system constraints, or can be maintained.
#1 Best Overall
- Frame the problem: clarify the user need, requirements, and what success means before asking for an implementation.
- Supply context: identify relevant code, interfaces, conventions, dependencies, and constraints so proposed changes fit the existing system.
- Evaluate the change: determine whether the implementation is correct for the intended behavior and whether tests or other checks support that judgment.
- Own the result: decide whether the change is suitable to deliver and maintain. Generated code is an input to that decision, not evidence that the decision has already been made.
The cited studies do not measure the full lifecycle of engineering work, nor do they establish a universal amount of time saved on review, integration, or maintenance. It would therefore be too strong to infer from faster code generation or a higher task count that an equivalent share of end-to-end engineering effort has disappeared.
Why do some AI coding studies show slower work while others show gains?
The work and participants were different
METR’s early-2025 trial focused on experienced open-source contributors working in established repositories. The three Management Science experiments took place in workplace deployments with a broader developer population. A tool can affect task speed differently depending on the work, the developer’s experience, and the surrounding codebase; the studies do not isolate one effect that applies to every team.
The outcomes were different
METR’s 19% result concerns time taken to complete tasks in its trial. The workplace experiments’ 26.08% result concerns completed task counts across their combined sample. A change in task count cannot be translated directly into a change in time per task or total delivery effort. The studies also differ in design and setting, so their estimates are not competing measurements of the same quantity.
Organizational conditions matter
DORA’s 2025 report frames AI as an amplifier of an organization’s existing strengths and weaknesses. That is a broad synthesis rather than a controlled causal estimate, but it highlights why the same coding assistant may be useful in one environment and disruptive in another: outcomes depend partly on the practices and conditions around its use.
Later METR results have an important limitation
In its February 2026 update, METR said its subsequent experiment gave an unreliable signal of AI’s current productivity effect. The organization reported that 30% to 50% of surveyed developers said they chose not to submit some tasks because they did not want to do them without AI. It also described difficulty measuring time for some developers using concurrent agents. METR’s researchers believed developers were likely more sped up in early 2026 than the early-2025 estimate suggested, but characterized the evidence for the size of that increase as weak. These selection and measurement issues make the later study unsuitable as a clean replacement estimate.
Can you trust AI-generated code without reviewing it?
The available findings do not justify either blanket distrust or blanket trust. Microsoft Research found that sustained use increased developers’ perceptions of coding tools as useful and enjoyable, while perceptions of generated-code trustworthiness did not change. That is evidence about reported perceptions, not a technical audit proving code safe or unsafe.
Review should answer the questions relevant to the particular change: does it implement the intended behavior, fit the surrounding system, and have appropriate checks? The studies do not establish a universal review procedure, a defect rate, or a long-run maintenance outcome. Treat generated code as a proposed change whose suitability must be established, rather than assuming that fluency or apparent completeness is verification.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What should teams measure when they adopt coding assistants?
Choose measures that reflect the outcome the team actually cares about. Task counts, time per task, reported experience, and end-to-end delivery are different indicators; none should be silently substituted for another. A useful evaluation also records enough context to interpret results, including the kind of work, developer experience, tool use, and changes in workflow.
Best Value
- Define what counts as a completed task and whether completion includes the checks and integration expected for that work.
- Separate code-generation speed from delivery outcomes rather than treating them as interchangeable.
- Track quality and maintainability questions explicitly; the cited studies do not establish defect rates or long-term maintenance effects.
- Interpret results in the team’s own setting. DORA’s synthesis and the contrasting experimental findings offer reasons to avoid assuming that another organization’s result will transfer unchanged.
The defensible conclusion is narrower than either “AI makes developers faster” or “AI makes developers slower”: assistants can change how code and tasks are produced, but the measured effect depends on context and metric. Engineering still includes deciding what to build and establishing that the result is fit for its purpose.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
