A May 2026 National Bureau of Economic Research study examined how productivity effects play out as organizations adopt successive generations of AI tools.
The takeaway: the sharp increase in code generated by AI-assisted development does not correlate smoothly with productivity. Production is becoming easy, but curation and distribution remain stubborn obstacles, and therefore opportunities, for those willing to put their AI and human resources in the right places.
The illusion of speed in the boardroom
For the modern CEO, the promise of generative AI in software engineering was supposed to be a “10x” miracle. On paper, the metrics are intoxicating: commit volumes are exploding, and the sheer velocity of keystrokes suggests a revolution. But a frustrating paradox has emerged for founders and board members: despite the massive infusion of AI tools, product release cycles remain stubbornly stagnant.
Our verdict for the board is simple: your engineering team is moving faster than ever, but they are not shipping faster. This analysis examines the hidden costs of “upstream automation” and why your organization is likely producing more noise but not more product.
1. The “17.3x” illusion: diversification without execution
There is a massive disparity between what happens at the keyboard and what reaches the customer. Data on the latest generation of autonomous coding agents shows a staggering 17.3x (1,700%+) increase in lines of code. But this explosion of activity attenuates sharply as it moves toward the finish line.
The data reveals a critical missing layer between activity and output: “repositories touched.” While AI-assisted developers are touching 1.5x more projects, they are failing to close them. For the “sync agent” generation (real-time assistance), actual software releases increased by a mere 20.3%.
The strong productivity gains from AI are attenuated by human bottlenecks in the production chain. Task-level productivity effects attenuate substantially as we move from more granular to higher-level outcomes.
The strategic takeaway: writing code is no longer your bottleneck. Shipping code is. Your team is spreading itself across more projects, but the friction of the final launch remains unchanged.
2. The “O-ring” hypothesis and the 0.25 elasticity trap
To understand why shipping hasn’t accelerated, apply the “O-ring” theory of economics. Just as a single failed O-ring caused the Challenger shuttle to break apart, a single human-performed task, such as code review or security integration, can negate the output of 17,000 AI agents.
The data reveals an elasticity of substitution of 0.25 between AI and human effort. In economic terms, AI and humans are strong complements, not substitutes. Because they are near-perfect complements, doubling AI output without doubling human review capacity leads to a mathematical certainty: a backlog.
Even unbounded automation of the upstream layer yields only finite productivity gains.
Strategic directive: AI has shifted the cost from creation to curation. If you keep “hiring more AI” without evolving your human review processes, you are simply paying for more congestion at the end of the pipeline.
3. The three generations: the bottleneck is moving up
The impact of AI depends on which generation of tools your team uses. The bottleneck is moving “up” the production chain as these tools evolve:
- Generation 1, Autocomplete (+36% commits): operates at the line-of-code level. It helps with syntax and speed but leaves the developer in full control of the unit of work.
- Generation 2, Sync Agents (+109% commits): operate in real time alongside the developer (for example, Claude Code). These agents can edit across multiple files but require constant human supervision.
- Generation 3, Async Agents (+180% commits): operate at the pull request level. These tools (for example, OpenAI Codex) autonomously handle entire tasks.
Autocomplete helped with syntax; async agents now attempt to handle entire pull requests. That moves the bottleneck from the IDE to the “merge button.”
4. The marketplace flood: the supply-side expansion trap
Founders should beware of building in an oversaturated market. Evidence from the iOS, Android, and Chrome stores shows a broad increase in new application releases since mid-2025. Yet total app usage has remained flat.
The ease of building an app does not create more consumer attention. In fact, the data shows that the share of new applications failing to reach even a modest audience has risen. AI is making it easier to build invisible apps.
As a contrast, SourceForge, a platform with an older developer base less inclined to use AI, has seen no such spike. That suggests the “AI flood” is a choice, not an industry-wide rising tide. Do not spend capital building functional products that will simply drown in a market with no extra attention to give.
5. AI as a leveler: reassess your hiring matrix
One of the most significant strategic findings is that AI serves as a “leveler.” Productivity effects are consistently larger for less active or less experienced developers.
For your “top 1%” engineers, the gains are substantial but hit a cognitive ceiling. For junior talent, AI provides a bridge to significantly higher output. The JLytics verdict for the board:
- Reassess your hiring: your “10x developer” is now primarily a reviewer and architect.
- Elevate the floor: your “1x developer” is now a 5x producer. AI is best used to lift junior talent into mid-level tasks, freeing your seniors for high-level architecture and the human-in-the-loop judgment AI cannot replicate.
Conclusion: moving the bottleneck
The cost of software development has fundamentally shifted from creation to curation and distribution.
The explosion of AI-generated code has moved the bottleneck from the keyboard to the merge button. As a leader, you have to be direct with your engineering leadership: are they prepared to manage a review-heavy workflow instead of a write-heavy one?
Yesterday’s managers were evaluated on their team’s ability to write. Tomorrow’s managers must be masters of judgment. In an automated production chain, human judgment is the most valuable and scarcest resource you have. If the merge button is the new site of technical debt, your strategy has to change accordingly.
Ready to work with an executive advisor ally who will build you a personal data environment to maximize performance while reclaiming your free time?
