AI coding assistants have moved from experiment to everyday tooling at remarkable speed. Google’s DORA research programme reported in 2025 that 90% of technology professionals now use AI at work, and for most organisations the conversation has now moved on to what that investment is returning.
Engineering teams usually answer measures with things like code arrives faster, boilerplate disappears and engineers saying they feel more productive. However, many technology leaders looking at their delivery figures see a less convincing picture – release cadence has barely moved, lead times look much as they did a year ago, and in some teams the number of hotfixes and incidents has crept upwards. Both observations can be accurate at the same time.
This article explores why gains at the engineer level often fail to reach production, what the research says about the gap between perceived and measured productivity, where the slowdown tends to hide, and how technology leaders can measure and rebalance the delivery system so that faster coding becomes faster, more reliable delivery.
Why does faster coding not always mean faster delivery?
Writing code is one stage in a longer journey. Before a change reaches users it typically passes through not just coding, but also refinement, peer review, automated and manual testing, integration, security checks, change approval and release. Each of those stages has its own capacity, and the overall pace of delivery is set by whichever one is slowest.
In many large organisations, coding was rarely the slowest stage to begin with. Changes have long waited on environment availability, specialist reviewers, release windows and change advisory boards, and those constraints remain in place after an AI rollout.
This principle, familiar from Eliyahu Goldratt’s theory of constraints, has an uncomfortable implication for AI adoption. If coding was never the bottleneck, making it faster does little to speed up delivery. The extra output joins a queue at the next constraint, work in progress grows, and changes spend longer waiting between stages.
Engineers pick up new tasks while earlier ones sit in review, so context switching increases and feedback arrives later. A useful comparison is a motorway widened ahead of a junction where lanes merge: more traffic reaches the merge sooner, and the overall journey time stays the same.
DORA’s 2026 report on the return on investment of AI-assisted software development describes this directly, arguing that without strong platforms, clear workflows and aligned teams, AI creates isolated pockets of productivity that are lost further down the delivery process. The report also expects most organisations to experience a dip before gains materialise, driven by the learning curve, the effort of verifying AI-generated code and the need to adapt testing and change approval to cope with higher volumes of change.
What does the research say about perceived productivity?
A second reason the mirage persists is that productivity with AI tools is surprisingly hard to judge from the inside. During early 2025, the research organisation METR ran a randomised controlled trial with 16 experienced open source developers working on 246 real tasks in mature codebases they knew well. At the outset, participants forecast that AI tools would cut their completion time by 24%. Afterwards, they estimated a 20% saving. Measured against tasks completed without AI, they had in fact taken 19% longer.
However, that headline slowdown has dated quickly. METR’s February 2026 update reported that developers are likely to be faster with current tools than they were in early 2025, while cautioning that its newer data offers only weak evidence of the size of the improvement, partly because many developers now decline to work without AI at all. The more durable finding is the gap between how fast people believe they are working and how fast they are measured to be, and METR’s later survey of technical workers flags reasons to be sceptical of the large gains people report.
The gains also vary sharply with the type of work. Research presented by Yegor Denisov-Blanch of Stanford’s Software Engineering Productivity Research Group, drawing on code data from around 100,000 developers across more than 600 companies, found gains of 30 to 40% for low-complexity tasks on new codebases, falling to between 0 and 10% for complex tasks on established ones. That distinction matters for large organisations, where most engineering effort goes into extending and maintaining existing systems.
Developer sentiment points in a similar direction, with Stack Overflow’s 2026 Developer Survey finding that 48% of developers trust AI output when they can easily verify it, while just 6.6% trust it for important work decisions. Trust that depends on verification places a steady load on reviewers and testers, which can be overlooked in an engineer’s sense of how quickly a task went.
Where does the slowdown tend to hide?
For example, consider a hypothetical scenario of a team of eight engineers that adopts AI coding assistants. Within a quarter, the number of pull requests raised each week rises noticeably, and the team reports that development work feels quicker. Lead time from commit to production, meanwhile, stays flat. Closer inspection shows changes waiting longer for review, a growing backlog awaiting test, and a handful of releases rolled back after issues surfaced in production. The examples below are typical of where this hidden cost sits, though the pattern will differ between organisations.
Code review:
Every change still needs a reviewer, usually a senior engineer whose time is already in demand. AI tools tend to increase both the number and size of changes, and larger changes take disproportionately longer to review well. As queues grow, reviewers face pressure to approve quickly, which shifts risk further down the pipeline. DORA describes the effort involved as a verification tax, and it falls on the people organisations can least afford to slow down.
Rework and instability:
The 2025 DORA State of AI-assisted Software Development report, drawn from a survey of nearly 5,000 technology professionals, found that higher AI adoption is now associated with greater delivery throughput and, at the same time, with greater delivery instability. DORA’s 2026 ROI analysis builds that instability into its financial model as a direct cost, recognising that failed changes and unplanned fixes can erode the return on AI investment.
Testing and integration:
Where automated test coverage is thin, manual testing absorbs the additional volume of change, and test environments become contested. Release cadence slows down to accommodate the backlog, or changes pass through with less scrutiny than they deserve. (We explore a related risk, test automation that reports success while missing defects, in a companion piece on self-healing test automation.)
Understanding of the codebase:
When fewer engineers have reasoned through the code they push, shared knowledge of the system thins out. The effect of this can quietly build up, with incident response taking longer because nobody can quickly explain what a component does, and future changes costing more because each one starts with rediscovery. Over time, knowledge concentrates in the few people who review most changes, creating a key-person dependency.
What questions can leaders ask?
A handful of direct questions to engineering leads will often reveal where work is gathering. Leaders might ask, for example, whether the time a change waits for review has increased since AI tools were introduced, whether change failure is moving in the same direction as deployment frequency, and how lead time compares with the growth in code being produced, with the answers pointing to where attention and investment should go next.
Which metrics give a reliable picture?
Activity metrics are the most tempting to report because they are easy to collect and move quickly after an AI rollout. Lines of code, suggestion acceptance rates and pull request counts all tend to rise by design once AI tools are in place, which makes them poor indicators of whether anything valuable has changed. As soon as an activity count becomes a target, teams naturally optimise for the count.
The five software delivery performance metrics published by DORA offer a sturdier foundation. Change lead time, deployment frequency and failed deployment recovery time describe throughput, while change fail rate and deployment rework rate describe instability. Read together, they reveal the signature of unbalanced acceleration: throughput climbing while instability climbs alongside it. Supplementing them with measures of flow, such as the time changes spend waiting in review or test, helps pinpoint where the constraint sits.
It is also worth resisting the urge to compare teams on these figures. Products, architectures and starting points differ, and the value lies in tracking each team’s direction of travel over time.
These measures also translate into terms a board recognises:
- Lead time reflects how quickly the organisation can respond to customers and regulators.
- Change fail rate and rework rate reflect the cost of quality and the exposure of customers to defects.
- Recovery time reflects service resilience.
Establishing a baseline before or early in an AI rollout makes the trend far easier to interpret, and pairing the figures with regular feedback from engineers explains why they are moving. (an approach we cover in Engineer Experience: Why and How You Should Measure It.)
How can organisations rebalance the delivery system?
With the constraint visible, the investment case becomes clearer. Guidance from DORA notes that reducing the batch size of changes is one of the most reliable ways to improve all five measures, because smaller changes are easier to understand, review and release. For AI-assisted teams, this often means agreeing limits on change size and resisting the temptation to let generated code arrive in large blocks.
Review capacity should be considered. Clear review standards, protected time for reviewers, static code analysis and AI-assisted first-pass checks can all reduce the load, leaving human reviewers to concentrate on design, risk and intent. Automated testing and continuous integration need to scale with the new volume of change, and the internal platform and pipeline need to support frequent, low-risk releases. DORA’s ROI research identifies platform quality, clear workflows and team alignment as the conditions under which AI investment produces measurable returns.
Change approval processes often need the same scrutiny. Manual approval gates designed for a monthly release rhythm struggle when the volume of change rises sharply, and many organisations respond by bundling changes together, which raises the risk of each release. Moving approval evidence into the pipeline, through automated tests, security scans and audit trails generated as part of each build, allows governance to keep pace without becoming the next constraint.
Additionally, a focus on ownership should mean that accountability for every change stays with the engineer who delivers it, whoever or whatever wrote the first draft, a principle we set out alongside wider governance and rollout considerations in AI-Assisted Engineering: Building the Foundations for Adoption and Scale. Leaders should also expect the dip that DORA describes and plan for it, so that a temporary slowdown is understood for what it is and funding stays in place until the return arrives.
Conclusion
The productivity gains from AI coding assistants are real at the point where they occur. Working code now arrives faster than before, and that capability will continue to grow. AI adoption and delivery discipline produce the strongest results when they are funded and managed together. Technology leaders who measure delivery across the whole system, look for where work is gathering and invest at that point will capture far more of the value their AI tools promise, and will be able to demonstrate it in terms their boards recognise.




