A global financial institution spent 2.76 million pounds a year on engineering and realized only a twentieth of the feature value that investment should have produced. New research from ClearRoute ties that outcome to a single structural cause, and the pattern holds across the organizations the consultancy has assessed over the past four years.
ClearRoute CEO James Jarvis argues the gap comes from how organizations operate, not from a shortage of AI tools. Engineering teams can produce code faster than ever, yet that speed rarely translates into features reaching customers on a comparable timeline. In his view, the slower stages are the ones surrounding the code itself. Test suites that break under real load, approval steps that still depend on a person reviewing and signing off, governance rules written for an earlier and slower way of working, and inconsistent environments from one team to the next all continue to move at their own pace, independent of how fast the underlying code gets written.
CTO Sarndeep Nijjar takes the point further, describing the next stage of AI adoption as a question of oversight rather than capability. As engineering organizations expand from single user coding tools to AI agents that can act across a delivery pipeline on their own, the value of that shift depends on the boundaries placed around it beforehand. Nijjar's view is that enterprises which deploy these agents without firm technical limits in place end up carrying more risk than they gain in return, while those that define the limits first put themselves in a position to benefit.
ClearRoute's assessment data supports that view with figures drawn directly from client engagements. One enterprise needed close to nine months, on average, to carry a single business critical feature from completed code to live production, a delay of 266 days. A separate defect classified as critical in production sat unresolved for 158 days before the same organization closed it out. Onboarding shows a version of the same pattern: a different company reported that incoming engineers typically waited three to six months before any of their work made it into a live release.
That unevenness extends to how organizations perform once software does reach production. Elite teams, running mostly automated testing and release processes, keep their change failure rate under 5 percent while shipping several releases a day. Organizations still leaning on manual testing and manual approvals see failure rates run considerably higher, in the range of 10 to 20 percent. AI does not close that distance on its own. It magnifies whatever process discipline already exists inside an organization, so capable organizations extend their lead while less mature ones fall further behind, even when both groups use comparable coding tools.
ClearRoute's research also flags a shift already underway across enterprise engineering. Rather than only producing code for a person to check, AI agents are starting to act on their own within production systems, take steps inside live environments, and move work directly through delivery pipelines. Enterprises that have not established clear rules for what these agents can access and what actions they can take on their own carry a form of risk that increases as this kind of deployment becomes more common.
Deciding in advance who and what can act inside these systems, and how that activity gets reviewed, is what separates organizations that gain from this shift from those that inherit its downside.
Download the full State of the Route to Live 2026 report: https://clearroute.io
Read the original coverage from Compare the Cloud: https://www.comparethecloud.net/news/ai-has-accelerated-the-easy-part-of-software-delivery-the-other-80-is-still-broken
A global financial institution spent 2.76 million pounds a year on engineering and realized only a twentieth of the feature value that investment should have produced. New research from ClearRoute ties that outcome to a single structural cause, and the pattern holds across the organizations the consultancy has assessed over the past four years.
ClearRoute CEO James Jarvis argues the gap comes from how organizations operate, not from a shortage of AI tools. Engineering teams can produce code faster than ever, yet that speed rarely translates into features reaching customers on a comparable timeline. In his view, the slower stages are the ones surrounding the code itself. Test suites that break under real load, approval steps that still depend on a person reviewing and signing off, governance rules written for an earlier and slower way of working, and inconsistent environments from one team to the next all continue to move at their own pace, independent of how fast the underlying code gets written.
CTO Sarndeep Nijjar takes the point further, describing the next stage of AI adoption as a question of oversight rather than capability. As engineering organizations expand from single user coding tools to AI agents that can act across a delivery pipeline on their own, the value of that shift depends on the boundaries placed around it beforehand. Nijjar's view is that enterprises which deploy these agents without firm technical limits in place end up carrying more risk than they gain in return, while those that define the limits first put themselves in a position to benefit.
ClearRoute's assessment data supports that view with figures drawn directly from client engagements. One enterprise needed close to nine months, on average, to carry a single business critical feature from completed code to live production, a delay of 266 days. A separate defect classified as critical in production sat unresolved for 158 days before the same organization closed it out. Onboarding shows a version of the same pattern: a different company reported that incoming engineers typically waited three to six months before any of their work made it into a live release.
That unevenness extends to how organizations perform once software does reach production. Elite teams, running mostly automated testing and release processes, keep their change failure rate under 5 percent while shipping several releases a day. Organizations still leaning on manual testing and manual approvals see failure rates run considerably higher, in the range of 10 to 20 percent. AI does not close that distance on its own. It magnifies whatever process discipline already exists inside an organization, so capable organizations extend their lead while less mature ones fall further behind, even when both groups use comparable coding tools.
ClearRoute's research also flags a shift already underway across enterprise engineering. Rather than only producing code for a person to check, AI agents are starting to act on their own within production systems, take steps inside live environments, and move work directly through delivery pipelines. Enterprises that have not established clear rules for what these agents can access and what actions they can take on their own carry a form of risk that increases as this kind of deployment becomes more common.
Deciding in advance who and what can act inside these systems, and how that activity gets reviewed, is what separates organizations that gain from this shift from those that inherit its downside.
Download the full State of the Route to Live 2026 report: https://clearroute.io
Read the original coverage from Compare the Cloud: https://www.comparethecloud.net/news/ai-has-accelerated-the-easy-part-of-software-delivery-the-other-80-is-still-broken