AI coding tools speed up individuals, but bigger pull requests, weak measurement and legacy processes keep engineering productivity and confidence stalled. The industry is witnessing a paradox where the individual efficiency gains provided by artificial intelligence are being negated by the surrounding systems, which slow everything right back down. This lack of result is significantly amplified by company size and pull request size. To the point that, while AI investment has increased 28 times for most companies, and especially those with more than 99 engineers, velocity measures are stagnant and even down. Such is the finding of the recently released State of AI Impact in Engineering from DX, which measures engineering organizations across speed, effectiveness, quality, and impact.
Justin Reock, deputy CTO of DX, tells The New Stack, “It is concerning because, when the cost has gone up 28x — which, literally, the only exponential metric is cost — and velocity is not exponential, we’re not shipping exponentially more.” While AI spend continues to skyrocket, the innovation ratio — the allocation of engineering effort spent on new feature work versus maintenance, toil, and operational overhead — remains flat. This means AI is not freeing up engineers’ time to spend on interesting business solutions, according to the report. How is the industry spending so much more money on agentic and AI developer tools, while also doing layoffs, to no avail?
It is possible that the sector is still in an inflection point where saved time is being absorbed by tech debt and backlog items not necessarily tagged as new features, Reock suggests, still hopeful that the gap between AI cost and benefit is just growing pains. However, the tension between code maintainability and change confidence is a major concern. These two drivers make up the Developer Experience Index (DXI) benchmark: Code maintainability, defined as feeling comfortable making changes to the code and understanding it, and Change confidence, the feeling that releasing code into production will not break things. Traditionally, these have positively correlated, as the former makes engineers more comfortable with the latter.
“AI is making it easier to understand what’s in front of you, and even to make changes to it. But change confidence is now in the negatives,” Reock says. “We’re more afraid to release the code. So we can understand the code, maintain the code, look at the code, and make changes to it more easily. But we trust less what we’re releasing.” This lack of trust carries a significant financial penalty. For every point of improvement in DXI, you return ten hours a year to each engineer. This is the first time that DX has witnessed a drop in this measurement, by two points industrywide, which equates to a loss of 20 hours per engineer annually.
Smaller organizations spend more on AI and get more out of it, DX finds, while legacy software organizations are struggling to see any return on investment. Martin Davidson, CTO of micro-consultancy a2bic.ai, takes this to the extreme: He and his co-founder, with a combined experience of about 80 years, can wrangle teams of AI agents to do the work of 100 junior to mid-level engineers. “Small orgs don’t have to pay the non-linear coordination and communication taxes that get worse with the size of the org. Remember the Mythical Man-Month: communication channels grow as n(n−1)/2, so a team of 10 has 45 channels of overhead,” Davidson explains. “Three of those people are effectively there just for alignment. But once you only have one or two people, the comms cost evaporates.”
Related: Coding agents can be evaluated by assessing their work
Medium-to-large organizations are trying to fit AI into existing systems across people, processes, and technology. AI represents a fundamental technological and operational paradigm shift, leading to what Davidson calls the renovation problem, where some structures are no longer fit for purpose. “You can’t retrofit this. We’ve got processes and structures which were designed when writing code was the expensive thing. That’s no longer true — code-writing is now essentially free, but we still cling to the old structures. And they aren’t cheap,” Davidson continues. “Company structures are like buildings — at some point you realize they are no longer fit for purpose and they need to be demolished and rebuilt from the ground up.” This reasoning mirrors the historical struggle enterprises faced moving to the cloud.
Davidson posits that the winners may not be the companies that successfully transform, but rather those that start fresh, unencumbered. While legacy organizations face this structural dilemma, AI is proving very useful at pattern recognition, helping to unravel the mystery of legacy systems, migrate them to the cloud, and rewrite them with AI in mind. Yet, many teams are ignoring universally accepted success patterns at the speed of AI, specifically how smaller batch sizes contribute to more stable releases. The DX report finds that in July 2025, the median PR size was 42 lines of code, while a year later it’s at 72 lines of code.
Of all the DXI indicators measured over the last quarter, incremental delivery — engineers reporting that they get to work on small, incremental changes — took the biggest hit. “Which carries all kinds of forward benefit — revert, less review, more understandable documentation, better unit test cases, like all this stuff,” Reock says. “That, in correlation with PR size, I think, is also concerning around quality.” LinearB’s AI engineering productivity gap report supports this, ranking 253 organizations into four buckets. This research finds that smaller pull request size directly ties to more successful AI adoption. “Elite organizations” in the top 10% had average pull requests of less than 100 lines of code, while those at the bottom — labeled as “Needs focus” — have pull requests of more than 228 lines of code.
Engineering is a science, meaning you cannot improve what you do not measure. According to research from LeadDev’s AI Impact Report 2026, only 31% of teams interviewed are measuring AI’s impact at all, defining these measurements as the retention of core engineering skills. Among organizations that have actually started adopting AI-powered developer tools, 70% describe themselves as having adopted “widely or completely,” yet only 26% report that AI has boosted engineering productivity by more than 25%. Michael Hill, managing editor at LeadDev, clarifies that the 26% includes respondents going on instinct, not just confirmed data. “The productivity optimism — 26% seeing big gains — and the measurement gap — only 31% actually tracking it — are two separate findings from two different questions,” Hill says, meaning “most of the people reporting gains aren’t the same people who can prove it.”
Even for those actively measuring, the results are worrying. Reock points to the DXI score going down for the first time last quarter. “We should really be paying attention to that because our customer base trends up, right? The data is heavily biased because they [DX customers] are investing in developer experience, like actively. They bought a product, and so that number tends to trend upward for our cohort of customers. So to see this actually go down is very concerning.”
