AI speeds up coding but creates new risks & software delays
Recent research from Harness has found that while artificial intelligence is helping developers produce code at a faster rate, bottlenecks elsewhere in the software development process are negating these initial productivity gains and increasing overall risk and cost for organisations.
The findings, published in Harness's State of AI in Software Engineering 2025 study, are based on a survey of 900 engineers, platform leaders, and technical managers across the United States, United Kingdom, France, and Germany. The research highlights several challenges that companies are facing as they adopt AI solutions into their development workflows.
Productivity shift
According to the study, 63% of surveyed organisations reported that code creation has accelerated since introducing AI-powered tools for their developers. However, this increased speed in generating code does not always translate to swifter or safer deployment. Some 45% of deployments involving AI-generated code have led to problems, and nearly three-quarters (72%) of organisations said they have already experienced at least one production incident attributed to code written by AI assistants.
The transition of bottlenecks from the coding stage to downstream processes such as testing, deployment, and security is described in the study as the "AI Velocity Paradox." Trevor Stuart, Senior Vice President and General Manager at Harness, commented:
"The AI Velocity Paradox is real. Teams are writing code faster, but shipping it slower and with greater risk. Without automation in place, productivity gains at the front end are erased by downstream bottlenecks - more bugs, higher cloud costs, and greater security exposure. To truly benefit from AI, enterprises need to extend it beyond code creation into testing, quality, and deployment. That's how you turn speed into lasting advantage, delivering software that is faster, safer, and more resilient."
Risks increase downstream
Security concerns are prominent, with 48% of leaders expecting AI-driven developments to increase software vulnerabilities. Despite the perceived risk, only 41% of organisations are fully confident in their governance protocols to catch issues before they reach production. Additionally, 70% worry that inefficient AI-generated code could fuel uncontrolled cloud spending, putting further strain on technology budgets.
Tool sprawl is also emerging as a concern. Development and engineering teams are now using, on average, between eight and ten distinct AI tools, while more than a third (36%) of respondents indicated they are managing even more tools. This tool fragmentation is having implications for team productivity, as 71% say that context switching between them hampers their efficiency.
Automation, a critical area to address these challenges, remains underdeveloped. Just 6% of organisations described their continuous delivery process as fully automated. According to the report, those with even moderate automation were more than twice as likely to realise velocity gains, with the rate rising from 26% to 57% when moving from low to moderate automation.
The concern around so-called "vibe coding" - where developers rely on intuition and AI assistance rather than structured review and process - is also growing. Sixty-three percent of respondents believe that "vibe coding is a disaster waiting to happen", fearing that it will overwhelm skilled engineers with downstream rework.
Downstream danger
The Harness study also found that 73% of respondents warn that unmanaged AI assistants in the software development lifecycle could significantly widen the impact of failed releases, and 74% believe that companies failing to integrate AI safely and securely could become obsolete.
Priorities for improvement
The research points to several priorities for organisations that want to realise the benefits of AI without incurring greater risk. These include pairing AI for coding with automated tools for testing, deployment, and security checks; focusing investment on downstream automation; and consolidating fragmented toolchains to reduce complexity.
The report also highlights that the most significant performance gap exists not in the creation of code, but in software delivery. Investment in continuous delivery and governance automation is cited as key to converting the speed of AI-driven development into tangible business gains.
Fragmented tool environments and less experienced developers using AI assistants present further governance challenges, potentially leading to more incidents and hidden costs. The study suggests a shift to unified platforms and the introduction of AI-powered guardrails to keep teams focused on their objectives and to limit risk.
Stuart noted:
"Harness is building this foundation today. By applying AI across the entire delivery lifecycle, we help enterprises escape the paradox and free developers to focus on innovation instead of rework."