Free tools Windows power users keep installed
One-click scans. No signup required.
The highest-performing AI-assisted engineering teams do more than give developers a coding assistant. In McKinsey’s 2025 survey, they were more likely to use AI across several stages of software development, train people on real work, redesign responsibilities, measure business and engineering outcomes, and actively manage organizational change. Those practices distinguish the survey’s top performers, but the findings show association—not proof that any one practice causes better results.
What does “top 20%” mean in this comparison?
McKinsey’s November 3, 2025 analysis surveyed nearly 300 senior leaders at publicly traded companies across the Americas, Asia, and Europe. Of those, 100 assessed performance impacts across software quality, time to market, team productivity, and customer experience. McKinsey defined “top performers” as the highest quintile across those four measures and “bottom performers” as the lowest quintile. The reported gap between the groups was 15 percentage points. McKinsey’s analysis spans multiple sectors; it is a comparison of survey-reported outcomes and practices, not a controlled experiment.
As an Amazon Associate I earn from qualifying purchases.
Top performers reported improvements of 16–30% in team productivity, customer experience, and time to market, and 31–45% in software quality. These are ranges reported by the surveyed organizations, not forecast gains or a guarantee for teams adopting AI now. The analysis does not establish that every company measured outcomes identically, or that the practices below caused the reported differences.
Which practices distinguish the stronger teams?
They apply AI across the development lifecycle
Top performers were six to seven times more likely than peers to have scaled four or more AI use cases. Nearly two-thirds of leaders reported four or more use cases at scale, compared with 10% of bottom performers. The applications spanned design, coding, testing, deployment, and tracking adoption—not just generating code. McKinsey’s account therefore points to breadth of workflow integration as a distinguishing pattern, rather than simply having access to a tool.
#1 Best Overall
They adapt roles and team structures
McKinsey describes engineers taking broader responsibility across product decisions, architecture, testing, and AI-assisted workflows, alongside clearer ownership of releases. That kind of redesign can connect AI-assisted work to decisions about what to build and whether it is ready to ship. The article’s Cursor example is based on an interview, so it illustrates one organization’s approach rather than proving that the same structure works everywhere.
They teach through practical work
Hands-on workshops and one-to-one coaching were reported by 57% of top performers, compared with 20% of bottom performers. McKinsey recommends practice tied to actual work, such as code review, sprint planning, and testing. For an engineering organization, that means giving people a chance to learn how AI fits into the tasks they own—not assuming that tool access alone creates consistent skill.
Rank #2
They track outcomes, not just AI activity
Among top performers, 79% tracked quality improvement and 57% tracked speed gains, according to McKinsey. Its suggested outcome measures include cycle time, release quality, and customer satisfaction, considered alongside input measures such as tool adoption. Code volume or the share of code generated by AI does not, by itself, show whether customers receive better software sooner.
They make adoption an organizational responsibility
Nearly eight in ten top performers linked generative-AI goals to both developer and product-manager reviews. Among bottom performers, 10% did so for developers and none for product managers. McKinsey advises recognizing useful behaviors—such as finding automation opportunities or improving quality—rather than rewarding raw tool usage. That distinction helps align incentives with the results the organization wants, rather than with visible activity that may not create value.
Rank #3
What should a team measure to tell whether AI is helping?
Use a small set of measures that covers both engineering performance and the consequences for users. Record a baseline before changing the workflow, then compare like with like after adoption. The survey percentages above are not targets for an individual team; local measurement is needed to determine whether a change helped.
- Delivery: track cycle time or time to market for comparable work, not just how quickly a developer completes an isolated coding task.
- Quality: examine release quality and defects, using the same definition and observation window before and after the change.
- Customer outcomes: include customer satisfaction or another relevant user outcome so faster production is not mistaken for better delivery.
- Productivity: choose a team-level measure tied to completed work and its quality; treat code volume and AI-generated-code share as activity indicators, not outcome measures.
- Adoption: monitor whether and where tools are used to understand implementation, but do not treat adoption alone as evidence of performance improvement.
McKinsey’s reported ranges do not specify that every company used the same measurement method. A team should therefore define its own measures consistently and avoid attributing a change to AI if other workflow or organizational changes could explain it.
Rank #4
How can an organization put these findings into practice?
- Choose a real workflow constraint. Identify a bottleneck in design, coding, testing, deployment, or feedback. Select an AI use case that addresses it rather than starting with a tool and looking for work to assign to it.
- Set a baseline and intended outcome. Decide which delivery, quality, customer, or productivity measure should change, and record its current level before the workflow changes.
- Make expectations and safeguards clear. Explain the organization’s AI plans, address developer concerns, and set usage policies. DORA’s January 2025 guidance identifies clear communication, time to learn, and explicit policies as practical adoption measures. Its analysis used 1,000 developer and developer-adjacent respondents and Bayesian regression on self-reported team AI usage; its estimates are associations with adoption, not universal causal effects or productivity gains. See DORA’s adoption guidance.
- Give people time to practice on real tasks. Use workshops, coaching, and supported learning time to build skill within the workflows the team actually uses.
- Expand only when the use case is useful. If results support it, extend AI use to connected lifecycle stages and clarify who owns workflow changes, quality checks, and release decisions.
- Review outcomes and incentives together. Compare follow-up measures with the baseline, consider other changes that may have affected results, and reward quality and customer outcomes rather than raw usage.
Why tool adoption alone is a weak benchmark
In GitHub’s online survey of 2,000 non-student respondents at companies with at least 1,000 employees—500 each in the United States, Brazil, Germany, and India—more than 97% said they had used AI coding tools at work at some point. The survey ran from February 26 to March 18, 2024, and was updated April 15, 2025; it did not measure frequency of use. Its self-reported result describes those four markets, not all engineering teams, and says little about whether usage improved outcomes. GitHub’s survey is useful adoption context, not evidence of team-level causal gains.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesDORA’s 2024 report drew on responses from more than 39,000 technology professionals worldwide and discusses AI alongside platform engineering, user-centricity, and stable priorities. Its abstract offers broad context for technology-organization performance, but does not establish the effect of any particular AI practice. DORA’s report abstract should not be read as a specific test of the practices described above.
What the comparison can—and cannot—tell engineering leaders
The practical signal is that stronger performers appear to treat AI as a change to the development system: workflows, skills, responsibilities, measures, and incentives all matter. The survey does not provide a guaranteed recipe, a causal ranking of those practices, or a promise that reported improvement ranges will transfer to another company. Treat the findings as hypotheses to test against your own baseline, with measures that reflect both delivery and the quality of what reaches users.
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.
