Is coding becoming the new back office?
Every technological revolution so far has created new high-value work activities while pushing other roles down the value chain as they become cheaper to deliver at scale.
In the last such revolution, the Internet helped turn modular business processes like customer service, data entry, and technical support into tradable services that could be offshored. Between 2002 and 2003 alone, UNCTAD estimates that low- and middle-income countries increased their share of offshoring-related FDI projects from 37% to 51% as firms took advantage of rising connectivity to harness cheaper labour markets. But while ‘call centre work’ generated formal jobs and export earnings for these countries, it delivered little in the way of wider productivity gains or growth if countries did not use it to deliberately upgrade to higher-value specialisms.
The latest AI coding agents are pushing programming toward automation, raising the fear that developer roles - historically high-paid and high-value - could be commoditised into the next call centre work. This calls into question whether growing developer and software hubs in Latin America and Asia are actually the next step up on the value chain for these markets or a treadmill to get stuck on, particularly as analysis of AI chatbot use from Anthropic finds that AI is used disproportionately for development in emerging economies.
To explore this more, we analysed GitHub data to track how programming is changing in the age of generative AI. Widely seen as a reliable hard skill, with high levels of AI adoption and strong, domain-specific measures of AI capabilities, programming is a good test case for how global value chains are set to change, and where intentionality from developing markets can help them leapfrog.
We find that there are real risks: since the launch of ChatGPT in November 2022, the early productivity gains have mostly shown up in high- and middle-income markets, while emerging economies have shifted faster toward the more automatable ends of programming. But 2026 could be a turning point. Developers in lower-income countries are learning fast from the bottom up and starting to close the gap in how they use coding agents, opening the door for policy and ecosystem support that means AI widens opportunities for innovation, rather than hardening coding into the next low-value offshore service.
Developers have become more productive - but the gains from AI are not yet transformative, and have not yet spread to lower income countries
For context, over half of all developers contributing to GitHub originate from ten markets. Within this, the highest growth in developer populations comes from large, fast-growing low and middle income countries - namely, India, Brazil and Indonesia.
Since 2020, the number of developers and volume of programming activity on GitHub have trended upwards at a steady pace. At the aggregate level, the launch of ChatGPT in November 2022 and increasingly capable models in the intervening years has not spurred dramatic changes to developer populations or activity.
In fact, for countries of all income levels, the general trend has been a slight slowdown in the rate of growth for developer numbers and repositories in the years following the launch of ChatGPT in November 2022 relative to the years just before.
However, developers in higher income and middle income countries appear to have become more productive since the advent of generative AI tools; growth in git pushes has accelerated in these markets since 2022. This means that, all else being equal, developers in middle and high income countries are producing more work than before.
With the exception of India - where average annual growth in git pushes has increased by 20 percentage points since 2023 despite growth in developer numbers slowing over the same period - lower income countries have not partaken in this initial rise in productivity. This could be driven by lower AI adoption levels, less effective harnessing of AI tools, or the use of cheaper, outdated models. This is in line with recent evidence that gains from AI-assisted programming accrue to more senior developers and frontier markets using AI tools selectively, with no measurable impact on more inexperienced developers.
Digging deeper, we also find initial evidence that globally, developers are becoming more responsive to changes in AI capabilities. While levels of activity on GitHub were initially only mildly impacted by the release of new models, in the two weeks following the release of Anthropic's Opus 4.5, the number of pull requests made on Github increased by a third.
It remains to be seen whether this AI-assisted increase in activity will persist through 2026 as models become more capable and increasingly agentic. What’s already clear is that meaningful productivity gains are likely to accrue to developers and firms that know how to use these tools well and embed them into workflows, rather than to tool capability alone. That raises a near-term risk for low-income countries: if adoption lags, they may lose competitiveness and substitutability for higher-cost labour in advanced markets and fall further behind in the global value chain instead of using programming to climb it.
Developing countries are increasingly taking on more automatable programming work
As GitHub activity shows, programming has shifted over time towards higher-level ‘applications’ languages like Python, TypeScript, and Java that make it faster to build user-facing software like web and mobile apps, create data tools, and run analytics. Meanwhile, lower-level ‘systems’ languages like C and C++, used for performance-critical foundations such as operating systems, databases, and core infrastructure, have taken a smaller share of new work as more of that plumbing is handled by cloud services and ready-made open-source components.
The advent of generative AI tools has accelerated this trend; GitHub activity in applications languages rose from under two thirds of all activity in 2020 to three quarters in 2025. This suggests that increasingly, programming is being used in a wider range of contexts for product development, tool-building, and research - modulisable, discrete uses to which AI assistance is particularly suited - rather than systems upkeep in narrower software development roles.
In particular, the rapid rise of programming in TypeScript is especially telling because it tends to grow when developers are building lots of product-facing code - web interfaces, APIs, and integrations - and because it rewards the kind of work AI tools are particularly good at: generating UI components, wiring data flows, following framework patterns, and producing large amounts of consistent, type-safe boilerplate.
Our analysis suggests that developers in lower-income markets are shifting toward application-focused languages - and the work they involve - faster than those in advanced economies; a trend further accelerated by generative AI. In India, for example, TypeScript rose from the 11th most-used language in 2020 to the 5th in 2025, and its average annual growth rate after ChatGPT’s launch was almost 50 percentage points higher than before.
More broadly, most lower- and middle-income markets are increasing their use of application languages even though they started from a higher base. That higher baseline likely reflects a newer developer population and less legacy systems work. The continued rise may also be a sign of how work is increasingly being split internationally: web development and tooling are easier to offshore to lower-cost markets, while higher-stakes, complex systems work - where mistakes are costlier - tends to remain concentrated in advanced economies.
This shift isn’t necessarily negative - much of today’s value is created in applications - but developing countries will still need the capability to build and maintain their own core systems and infrastructure. Application-layer work is also easier to automate and to standardise, which can make it more vulnerable to being pushed into lower-value, commoditised tasks.
We analysed the topic tags developers attach to their GitHub repositories across countries. Drawing on literature on the classification of software tasks and GitHub activity1, we group projects into three core buckets: infrastructure and tooling (routine, upstream work like testing, monitoring, and developer tools), platform and framework development (non-routine work building reusable foundations like core libraries, algorithms, and architectural components), and domain application (non-routine, downstream problem-solving for specific business and social needs).
We find that infrastructure and tooling activity has risen broadly everywhere, but the increase is markedly faster in lower- and middle-income markets. That pattern raises a risk that these ecosystems become concentrated in more routine, outsourced, support-style work, with AI accelerating this transition to coding-as-call-centre-work (while simultaneously depleting demand for human labour to do these activities) instead of helping to build the deeper capabilities needed to move into higher-value engineering and develop home-grown local solutions.
(NB: We limit our topics analysis to a smaller set of countries because GitHub topic tagging is less widely used than other indicators and is itself a marker of more established development practices. As a result, countries with sufficient topic data tend to be relatively large or mature developer hubs, meaning that this is likely a conservative estimate of growth in infrastructure and tooling work for lower-income markets.
Due to the limited range of topics data, we test the robustness of our analysis by applying an unsupervised NLP clustering approach to commit messages on GitHub Archive data from 2022 to 2025, allowing natural groupings in the data to surface.This data is much more widely available but requires location data to be mapped on. The data is also messier and dependent on the quality and veracity of commit messages inputted by individual users. Nonetheless, a similar pattern emerges: while AI does not dramatically change the composition of commits in developed markets, developing markets see a notable increase in generic and low-quality commit messages as well as those typically mapped to infrastructure and tooling work between 2022 and 2025)
But bottom-up adoption of the latest generation of AI coding agents could reverse this trend in 2026
Between 2022 and 2025, developers in lower- and middle-income markets appeared to lag in AI adoption while shifting faster towards front-end and application-layer work, increasing their risk of being locked into lower-value roles in the global value chain.
However, more recent evidence suggests an opportunity for a turning point. Anthropic data from late 2025 - representative of a small part of the market as it is - suggests that as coding tools become more agent-like, users in lower- and middle-income markets are moving from using them mainly for simple automation to using them more collaboratively; iterating, giving feedback, and applying them to harder problems.
India - where more than half of Claude prompts relate to coding - shows how quickly bottom-up learning can shift usage. Over three months, Indian users moved from some of the highest rates of directive automation prompts to using Claude to augment their work at levels comparable to much richer countries. By November 2025, the average prompt from India reflected task complexity similar to coding prompts in high-income markets according to Anthropic’s measure of the years of human education required to complete the task.
If the latest coding agents keep leaping forward at current rates, and policymakers in emerging markets supplement this bottom-up learning by developers with top-down support embedding the latest tools in workflows, then coding need not become the next call centre industry. Instead, it could democratise access to tools capable of creating entire products rather than just speeding up routine workflows, enabling more local innovation and raising productivity across the wider economy.
Emerging-market developer hubs like India - an ecosystem with scale, depth, and strong bottom-up adoption - are well placed to design the blueprint for success here. And doing so will mean that software development does not follow earlier global value chains, with high-value design and systems integration staying in advanced economies. The barriers to building an app, launching a dashboard, or prototyping a new digital product have already fallen drastically. If used well, as agent capabilities grow, coding can start to look less like a call centre or manufacturing plant, and more like the next mobile phone: an everyday tool by which emerging markets democratise, harness, and scale innovation.
1. Our taxonomy draws on established classifications of task routineness and software value creation: Autor et al. (2003) on routine versus non-routine cognitive work; Messerschmitt and Szyperski (2003) on ‘upstream’ infrastructure and tooling versus ‘downstream’ domain-facing development; and Baldwin and Woodard (2009) on platforms (reusable components with ecosystem control) versus applications (domain problem-solving built on platforms). We also draw on an analysis of what GitHub tags represent in practice (Kalliamvakou et al., 2014), including distinctions between production, experimental, and educational work, as well as language and community tags.
