AI coding assistants can help developers complete more coding tasks, but faster code generation does not automatically mean faster, safer software delivery. The central leadership question is how to redesign the work around the tools: what to delegate, how to verify outputs, which skills to protect, and what outcomes to measure.
Does AI make software teams more productive?
The evidence points to real gains in some settings, but not a universal productivity multiplier. A developer may finish a task sooner while the team still waits on review, reworks a change, or encounters a release bottleneck. Task output, end-to-end delivery, quality, and engineers’ experience are related but distinct outcomes.
As an Amazon Associate I earn from qualifying purchases.
| Evidence | What it found | What the finding can—and cannot—tell leaders |
|---|---|---|
| Microsoft Research, three randomized field experiments (2025) | A pooled estimate found an increase of 26.08% in completed tasks for developers offered an AI coding assistant (standard error: 10.3%). The experiments involved 4,867 developers. | This is the strongest causal evidence in the sources for a change in task completion under the studied conditions. The authors describe the estimates as noisy. It is not a guarantee of the same gain in another organization, nor a measure of software quality or company-wide delivery speed. |
| Microsoft workplace study (2025) | In a mixed-methods study at one large multinational software company, 84% of participants reported positive changes in daily work practices and 66% reported shifts in their feelings about work. The study included a randomized trial and a three-week diary study. | It captures workers’ reported experience as well as work practices in one company; it does not establish that every team will experience the same changes. With sustained tool use, perceived usefulness and enjoyment rose, while trust in AI-generated code did not change. |
| DORA / Google, 2025 report | DORA says its research drew on more than 100 hours of qualitative data and survey responses from nearly 5,000 technology professionals worldwide. | DORA frames AI as an amplifier of organizational strengths and weaknesses, with returns depending on the underlying system. These are the report’s described research inputs, not a census of all technology workers or proof that every organization will see a particular result. |
| Anthropic, internal study (data collected August 2025; published December 2, 2025) | Anthropic’s interviews and internal data describe engineers using AI across a broader range of tasks, alongside concerns about maintaining expertise, supervising outputs, mentorship, and collaboration. | The findings show possible benefits and trade-offs worth checking, not a settled workforce-wide effect. Anthropic cautions that its employees had early access to frontier models and may not represent other organizations. |
| GitLab / The Harris Poll, vendor-sponsored survey (released June 23, 2026) | In a survey of 1,528 developers and technology buyers across six countries, 80% said their organization adopted AI tools faster than it developed governing policies, and 92% reported governance challenges with AI-generated code. | These are respondents’ reports published by an interested vendor, not universal estimates. The announcement also describes difficulties distinguishing AI-written from human-written code, fragmented toolchains, and missing origin tracking. |
| McKinsey with Sonar, company case study | The case reports up to 0.2x higher pull-request throughput (up to 20%), up to 0.4x lower pull-request cycle time (up to 40%), and self-reported productivity gains ranging from 0% to 80%. | These are case-specific figures, not controlled evidence that another company should expect similar results. The productivity range is self-reported; the results should be read in the context of the workflow described in the case. |
Together, the studies support a bounded conclusion: coding assistance can change what people complete and how work feels, but outcomes depend on the setting and on what happens after code is generated. DORA’s 2025 report puts the organizational point directly, describing AI as an “amplifier” of existing strengths and weaknesses. A team with unclear ownership or slow verification may simply move its queue downstream faster.
Free tools Windows power users keep installed
One-click scans. No signup required.
Why doesn’t faster coding automatically mean faster delivery?
Software delivery is a chain, not a typing contest. Generated code still has to fit the system, pass checks, be understood by reviewers, and move safely through release. If the constraint is review capacity, test reliability, release coordination, or unresolved technical debt, increasing the volume of proposed changes may shift pressure to that constraint rather than remove it.
#1 Best Overall
This is why code volume, accepted suggestions, or time spent in an editor are weak stand-alone measures of organizational improvement. A useful evaluation follows work across the path from task selection to a production outcome, and checks whether speed came with more rework, defects, security exposure, or maintenance burden. Those are practical measurement dimensions, not a validated universal scorecard.
What should an organization redesign around AI?
“Redesign everything” is too broad to be an operating plan. Start with a defined workflow and make explicit decisions about responsibility, verification, and outcomes. The following dimensions are concrete places to examine.
1. Workflow and operating model
Specify which tasks an assistant may draft or carry out, what context it receives, and where a person must direct, review, or approve the work. Define ownership for changes that span components, and make sure the path from a proposed change to release has a clear human accountable for the result.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →McKinsey’s Sonar case describes an AI-oriented cycle that provides agents with context, generates code, checks quality and security, and uses feedback to resolve issues. Its useful lesson is the workflow design, not a promise that the same implementation or outcome will transfer elsewhere. McKinsey senior partner Martin Harrysson characterized the effort as focusing on the operating model, not just the tools.
2. Quality, security, and provenance
Make verification part of the workflow rather than an afterthought. Decide what evidence a change needs—such as tests, security checks, review, and a traceable record of its origin—and who is responsible for evaluating it. Teams should be able to understand what was generated or materially assisted, what was changed by a person, and how the final decision was made.
GitLab’s survey points to reported friction around identifying code origin and coordinating fragmented toolchains. Those are reasons to establish workable controls, but the survey figures should not be treated as a measure of every company’s risk. Controls need to fit the organization’s actual systems and obligations.
Rank #4
3. Engineering foundations
Generated changes are easier to assess when code is structured, architecture is understandable, tests are dependable, and technical debt is visible. If those foundations are weak, teams may struggle to determine whether a plausible-looking change is safe to maintain. Tariq Shaukat, Sonar’s CEO, argues in the McKinsey/Sonar case that strong foundations matter for agentic development; that is a vendor perspective, but it aligns with the practical need to make outputs checkable.
4. Skills, mentorship, and collaboration
AI can help engineers work in areas beyond their usual specialties, while also making it easier to accept an answer without fully understanding it. Set expectations that engineers remain able to explain, debug, and maintain the systems they change. For junior engineers, consider where foundational practice and feedback come from if an assistant becomes the first stop for questions.
Best Value
Anthropic’s internal study raised concerns about technical expertise, supervision, mentorship, and colleague interaction. Treat these as issues to observe locally, not proof that AI inevitably weakens skills or collaboration. Track whether teams are learning and sharing context, not only whether individual tasks move faster.
5. Measurement and learning
Measure several parts of the delivery system together. A practical local view can include completed work, pull-request throughput, cycle time, rework, defects or incidents, security findings, and team experience. Choose measures that match the workflow being changed; do not assume that any one metric is a universal KPI or that a rise in code output is a gain by itself.
Set a baseline before changing the workflow, compare like with like, and review results with the people doing the work. If throughput improves while review queues, rework, or incidents worsen, the intervention may have moved the bottleneck rather than improved delivery. If experience changes, investigate what changed in the work instead of treating sentiment as a substitute for delivery outcomes.
Quick Recap
How can leaders introduce AI without mistaking adoption for transformation?
- Choose a bounded workflow. Select a repeatable type of work with a clear owner and observable completion criteria. Avoid starting with a broad mandate to use AI everywhere.
- Map the existing path. Record where work waits, how it is reviewed and tested, what causes rework, and who approves release. This establishes the real constraint before a tool is introduced.
- Define delegation and accountability. State what the assistant may propose or execute, what context it may use, and which checks require human judgment. Keep a named person responsible for the change that reaches the next stage.
- Build verification into the trial. Ensure the team can evaluate generated work using its existing quality and security expectations, and identify any missing provenance or audit information that matters to the workflow.
- Compare outcomes, not activity. Review task completion and delivery flow alongside rework, quality, security, and worker experience. Separate observed results from workers’ perceptions, and account for changes in task mix or process.
- Adapt before scaling. If a check becomes a bottleneck, redesign that stage rather than simply pushing more code into it. Expand only when the workflow is understandable, maintainable, and producing results the organization values.
This sequence is a practical way to test a change, not a claim that every team needs the same rollout. The evidence spans randomized experiments, an internal study, a vendor-sponsored survey, and a company case; each answers a different question. Treat results as local until they are demonstrated in the work your organization actually does.
What the evidence does—and does not—justify
- It supports testing AI-assisted workflows. Randomized field experiments found a task-completion effect in their participating companies, while other studies report changes in work practices, experience, and delivery measures.
- It supports redesigning the surrounding system, not just distributing tools. Workflow ownership, review, testing, provenance, architecture, and skills affect whether faster generation can become dependable delivery.
- It does not establish a universal gain. The estimates and reports vary by study design and setting; the task experiment was noisy, the diary study concerned one company, Anthropic studied its own employees, the GitLab poll was vendor-sponsored, and the McKinsey figures are from a case study.
- It does not prove that generated code is inherently lower quality or that agentic development guarantees dramatic productivity gains. Those conclusions go beyond the evidence summarized here.
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.
