The productivity paradox every CTO is living through

Since GitHub Copilot went mainstream in 2022, every vendor in the software development space has published a version of the same claim: AI tools make developers dramatically faster. The numbers cited range from 35% to 90% productivity improvement. CTOs read these figures, watch their teams adopt AI tools enthusiastically, and then look at their engineering budget — which has not gone down.

The data is more nuanced than the headline numbers suggest. And understanding the nuance is where the actual cost strategy lives.

What the research actually shows

Three large-scale studies from 2024–2025 give the clearest picture. They point in different directions — and that divergence is itself the signal.

GitHub’s own study: tasks complete 55% faster

A GitHub-commissioned study involving 4,800 developers found that participants using Copilot completed defined coding tasks 55% faster than those without it. Code approval rates in peer review rose by 5%. Code readability improved by 3.6%. This is the most-cited figure in the AI productivity literature — and it is the most favourable to AI tools.

The key qualifier: the study used controlled, well-specified tasks, most of which involved writing new code from a clear brief. Completing a ticket is not the same as completing a sprint.

METR’s field study: experienced developers took 19% longer

A 2025 randomized controlled trial by METR tracked 16 experienced developers completing 246 real-world tasks in mature open-source codebases — the kind of work most production engineering teams actually do. With AI tools including Cursor and Claude, those developers took 19% longer to complete their work than without AI assistance.

The same developers perceived they were 20% faster. The gap between perception and measurement is one of the most significant findings in recent AI productivity research. It suggests that AI tools create a feeling of momentum that doesn’t always translate to output.

McKinsey’s analysis of 300 companies: top quintile gains 16–30%

McKinsey analyzed nearly 300 publicly traded companies for its November 2025 report on AI in software development. The top quintile of performers achieved 16–30% improvements in productivity and time to market, plus 31–45% gains in software quality. The other 80% saw materially lower returns.

The distinguishing factor was not which AI tools a company used. It was whether they had redesigned their workflows around AI, rather than simply adding AI tools to existing workflows. Companies that gave developers Copilot without changing how work was structured saw little to no measurable improvement.

Sources: GitHub Blog (2024), METR arXiv:2507.09089 (2025), McKinsey “The AI revolution in software development” (Nov 2025), Stack Overflow Developer Survey 2025, Harness SEI case study (2025).

Where AI saves real time — and where it does not

The productivity findings become consistent when broken down by task type. AI tools deliver clear, measurable gains on specific categories of work. They provide little benefit — and sometimes negative returns — on others.

Sources: emorphis.com blog (Nov 2025), CodeRabbit December 2025 report (1.7× more issues in AI-assisted PRs), Stack Overflow 2025 (45% of developers cite debugging AI code as time-consuming).

The cost structure has not changed — only what you get for the money

Here is the part the AI productivity headlines consistently miss: the cost of hiring a developer in Israel is determined by the labour market, not by their tool stack. A mid-level developer with Copilot still costs $10,500–$13,000 per month all-in. The salary, Bituach Leumi, pension, Pitzuim, and Keren Hishtalmut payments are unchanged.

What AI tools change is the output per developer-hour — for specific categories of work. If your team’s bottleneck is boilerplate and test generation, AI tooling is a genuine force multiplier. If your bottleneck is architectural clarity, product spec quality, or experienced judgment on complex systems, it is not.

*CodeRabbit (Dec 2025): AI-assisted PRs contain on average 1.7× more issues requiring review, creating new review overhead that partially offsets generation speed gains.

The real arithmetic: where does the saving actually come from?

The honest calculation of AI’s impact on an engineering budget looks like this for a typical Israeli mid-stage startup:

This is genuinely valuable. But it does not reduce the budget. It improves the output for the same spend. The only way to reduce the budget is to either reduce headcount or reduce the per-developer cost. AI tools, by themselves, do neither.

How the highest-performing teams are actually restructuring costs

McKinsey’s 2025 data found that only 5.5% of organizations are seeing real financial returns from AI investment — but the top performers share a pattern. They do not treat AI as a tool layered on top of existing workflows. They redesign the workflow around AI, then make deliberate decisions about which roles require local senior expertise and which do not.

The practical implication: teams that combine AI-native workflows with distributed engineering are achieving the multiplicative effect that AI tools alone cannot deliver.

The mechanism is straightforward. A senior developer in Vietnam, working with AI tools and a well-structured workflow, costs $6,500/month all-in. A senior developer in Israel, on the same tools and workflow, costs $13,000–$16,500/month. The AI productivity gain is roughly the same for both. The cost difference is not.

Calculations use: Israeli mid-level $11,750/mo all-in; Israeli senior/lead $15,000/mo; Vietnamese mid-level $6,500/mo (israviet rate). AI productivity premium of +15% applied uniformly. Output units are approximate and task-dependent.

What this means for team composition decisions

The insight that falls out of this analysis is not that AI makes local hiring unnecessary. It is that AI sharpens the distinction between what requires senior, local expertise and what does not.

Work that benefits most from local senior developers: product strategy, architecture decisions, stakeholder communication, specification writing, and anything requiring deep context in a legacy codebase. These are the tasks where AI tools provide the least leverage, and where experience, language fluency, and proximity matter most.

Work where AI tools are highly effective and where geography is irrelevant: new feature implementation, test coverage, documentation, boilerplate, code review first passes, API integration, and well-specified bug fixes. These are the tasks where a skilled, AI-native developer working remotely performs at or near parity with a local equivalent — at 48–52% of the cost.

The teams seeing the strongest results in 2025 have recognised this distinction and acted on it: a lean Israeli core (technical leadership, architecture, product interface) supported by a larger, AI-native offshore team handling implementation. The Israeli headcount focuses on the decisions that require their seniority. The offshore team delivers the volume that AI tools have made structurally more accessible.

The numbers in one place

What to take away

AI coding tools are not a cost reduction strategy on their own. They are an output amplifier — and the degree of amplification depends heavily on task type, developer seniority, codebase complexity, and whether the workflow has been redesigned to take advantage of them.

For Israeli tech companies, the cost pressure is structural and unaffected by tooling adoption. The budget does not get smaller because developers use Copilot. What changes is how much you can accomplish with a given headcount — and by extension, how you should think about the composition of that headcount.

The teams that are compounding both advantages — AI-native workflow design and a distributed team structure with a lower per-developer cost base — are capturing the double leverage that neither strategy achieves alone. That is where the real engineering budget story of 2025 is being written.