Almost a 3rd of developer time is now consumed by invisible work, akin to reviewing AI-generated code, fixing bugs, and context-switching between instruments.
[Image source: ChetanJha/MITSMR Middle East]
Synthetic intelligence is reshaping software program engineering quicker than many organizations can measure it. Whereas AI coding assistants are serving to builders ship code extra rapidly, a lot of the promised productiveness is offset by work that exhibits up in no metrics.
Harness has launched The State of Engineering Excellence 2026 report, exhibiting that AI coding instruments have remodeled the day-to-day work of software program builders quicker than the trade’s measurement frameworks can sustain.
Throughout the US, the UK, India, France, and Germany, 700 engineering practitioners and managers have been surveyed for the report. Almost 89% of engineering leaders stated developer productiveness improved after adopting AI coding instruments, whereas 88% reported improved developer satisfaction.
But the report additionally highlights a widening disconnect between how organizations measure productiveness and the way engineering work is definitely altering.
Based on the findings, 81% of engineering leaders stated builders now spend extra time reviewing AI-generated code. Greater than 1 / 4 reported that code-review workloads had elevated by over 30%. Almost a 3rd of developer time is now consumed by “invisible work” akin to reviewing AI-generated code, fixing bugs, and context-switching between instruments.
“It isn’t the work organizations are attempting to speed up; it’s the overhead connected to the work,” the report famous.
The findings add to rising considerations that enterprises could also be overestimating the effectivity beneficial properties from generative AI by counting on legacy software program metrics targeted on output quantity and supply cycles. Whereas AI-generated code can shorten improvement timelines and enhance throughput, engineering groups are more and more being requested to tackle new obligations in governance, high quality assurance, and danger administration.
“AI coding is the primary expertise shift in fashionable software program that has modified not simply what builders construct, however how they spend their hours,” stated Trevor Stuart, SVP and Normal Supervisor at Harness. “Cloud and the web have been infrastructure revolutions layered beneath the developer. AI is reshaping the developer’s job completely, and the measurement frameworks that the trade has relied on for the previous decade weren’t constructed for this new unit of labor.”
The report argues that many organizations proceed to guage engineering efficiency utilizing techniques designed earlier than generative AI entered workflows. Metrics akin to DORA benchmarks, cycle time, and velocity stay broadly used, however they typically fail to seize elements like validation time, technical debt, burnout, and cognitive load.
That contradiction is mirrored within the survey information. Whereas 89% of leaders stated their present metrics precisely mirror AI’s impression, 94% concurrently acknowledged that important elements — together with developer fatigue, code high quality, and technical debt — are lacking from these frameworks. Solely 6% stated current techniques have been totally able to addressing the hole.
The most important problem organizations recognized was not AI adoption itself, however visibility into its actual impression. Respondents cited measuring true productiveness beneficial properties, sustaining code high quality, and proving return on funding to management as the highest considerations.
The stress on builders can be intensifying. Harness’ findings counsel builders are more and more uneasy about how AI-driven productiveness information could also be used. Greater than half of respondents feared particular person efficiency evaluations primarily based on AI metrics, whereas 46% reported stress to work quicker than was sustainable. One other 46% raised considerations round surveillance and privateness.
The notion hole between management and practitioners was notably stark. Managers have been almost 4 occasions extra probably than frontline builders to say that they had no considerations about how AI productiveness information could be used.
To adapt, the report recommends that engineering organizations start measuring what it calls the “new unit of labor” created by AI-assisted improvement. That features monitoring validation time, AI-agent accuracy, code acceptance charges, burnout indicators, and cognitive load alongside conventional supply metrics.
It additionally urges corporations to separate improvement-focused analytics from worker efficiency evaluations and contain builders straight in defining how productiveness needs to be measured in AI-enabled workplaces.
“Engineering leaders are being requested to make multi-year AI funding selections utilizing dashboards constructed for a special period of software program improvement,” Stuart stated.