Offshore didn't kill the IT career: the delivery model did
Offshore outsourcing was never as cheap as the rate card suggested; fully loaded, the offshore team often costs as much as the onshore team it replaced. AI is now dismantling the specific labor-arbitrage mechanism that made offshore attractive for new development: the cheap translation of business intent into code. The article frames the industry as three eras, shows where the evidence supports the shift and where it does not, and argues that competition is moving from cost to judgment and context.
Run the fully loaded cost of your offshore program and you will find something uncomfortable: once rework, oversight, and the timezone tax are counted, the offshore team often costs as much as the onshore team it was supposed to replace. I have watched this arithmetic come out the same way on SAP programs for a decade. The offshore model was never really about cheap labor; it was about scale, and scale is exactly what AI is about to make obsolete. Think of the three eras as my framework for reading the industry, not a chapter of its official history: information technology has moved through three phases, and the second one, the offshore phase, is coming to an end not because companies suddenly love quality, but because the economic reason for shipping work across the world is dissolving.
Why offshore won, and what it actually cost
The first era of IT delivery was simple: you hired people, you put them in a building near the business, and they built and ran the systems. High quality, high cost, and no way to scale fast. Then the dot-com boom arrived and the demand for software outstripped the supply of local engineers. Companies could not hire fast enough in Boston, London, or Munich, so they followed the only lever that moved quickly: labor arbitrage. A developer in Bangalore or Manila cost a fraction of one in the West, and the internet made handing work across the world possible. Offshore became the answer to a scale problem, not a quality problem. That distinction matters, because it means the model was built on a specific economic mechanism, and mechanisms can be replaced. What the cost comparison usually left out was everything after the invoice. The specification had to be written in enough detail that a team twelve time zones away could act on it without asking questions, so a whole layer of analysts and coordinators appeared to translate business intent into tickets. The code came back, and local seniors reviewed it, fixed it, and often rewrote it. Every handoff was a tax. Add the rework, the oversight layer, the status calls at 6 AM, and the fully loaded offshore rate stopped looking like a bargain. On one program I worked, we ran the numbers properly for the first time and found the offshore delivery was within a few points of the onshore alternative. Nobody had checked before, because the headline rate was so seductive. None of this means the people offshore were bad at their jobs. They were doing exactly what the model asked: turning well-specified tickets into working code, at volume. The model's failure was structural, not personal, and that is the honest way to read the history.
Where the model started to crack
The cracks appeared long before AI, in the way software delivery changed. Agile methods demand tight collaboration, daily feedback loops, and the ability to change direction mid-sprint. All of that is hard across a twelve-hour timezone gap. The handoff became the dominant activity: write the spec at the end of your day, receive questions at the start of yours, and lose a day every time the answer was not in the ticket. The work I know best, SAP transformation, shows why this matters. A cutover weekend is not a series of tickets; it is a sequence of judgment calls about a system that has been running a real business for years. When a transport does not apply cleanly at 2 AM, the person fixing it needs context: which plants are affected, which process was mid-flight, what the business can tolerate. That context does not travel in a ticket; it lives in people who have sat in the room. SAP delivery was always an awkward fit for pure offshore, which is why the delivery market in my corner of the industry is already pulling back toward smaller, AI-augmented local and nearshore teams for the judgment-heavy work. This is the fulcrum of the whole argument: there is commodity work and there is judgment work. Offshore staffed the commodity layer, the translation of specs into code. The economics of the model depended on that layer being large and cheap, and that layer is exactly what AI is dismantling.
What AI actually changes
AI does not just write code faster; it removes the middleman layer that offshore economics relied on. The junior developer translating a spec into code is exactly the task that code models now do directly, and the evidence on where the productivity gains land is consistent. In the GitHub Copilot randomized controlled trial, developers completed a bounded coding task 55.8 percent faster with the tool. In Brynjolfsson, Li, and Raymond's study of 5,000 customer support agents, generative AI raised average productivity by 14 percent, with a 34 percent lift for the least experienced workers and almost nothing for the most experienced. Read that second finding twice, because it is the one that matters: AI's biggest measured gain goes to people doing well-defined, bounded work, which is precisely the profile of the offshore commodity layer. The model is not a threat to judgment work; it is a threat to the work that was outsourced because it did not require judgment. There is a second finding that cuts the other way, and intellectual honesty requires it be put next to the first. METR's 2025 study of experienced open-source developers found they were 19 percent slower with AI on complex, familiar codebases, while perceiving a 20 percent speedup. The perception gap is a warning: for senior people doing hard, context-heavy work, the tool can slow you down while feeling like a speedup. Taken together, the evidence points one direction. AI's gains concentrate in bounded, well-specified work, and the human judgment that used to be the expensive part of a project is now the scarce part. That inversion is what kills the offshore model, not the raw speed of any single tool.
The honest counter-argument
The strongest objection deserves a straight answer: if AI makes a $30 per hour offshore developer and a $150 per hour onshore developer both faster by a similar multiple, the cost ratio does not change, and the cheaper labor stays cheaper. Worse, the NBER result suggests AI helps the less experienced worker the most, which would seem to favor offshore juniors, not onshore seniors. And the scale problem is real: a handful of AI-augmented seniors cannot staff a 400-person program or a global support desk. Add the fact that the big offshore providers, TCS, Infosys, and Accenture, are retooling their own delivery models around generative AI, and the clean story of three eras with offshore as the casualty starts to look naive. Here is the more defensible version, and I think it is the true one. AI does not make onshore labor cheaper per hour, and offshore is not dying overnight. What AI does is shift the competition from cost to judgment and context. The work that used to be routinized, the spec-to-code translation layer, is now nearly free wherever it is done, which removes the pricing advantage of doing it in a low-wage country. The work that remains expensive, the judgment work, was never offshore's strength anyway. The run, maintain, and support workload that is most of offshore's actual revenue base will persist, and the providers that adopt AI fastest may keep their position. But the specific economic mechanism that made offshore attractive for new development, cheap translation of business intent into code, is being dismantled in front of us. It is a directional shift, not a completed transition, and pretending otherwise would be as dishonest as the old headline-rate math.
What era three looks like
The third era is not the return of the first one. In-house IT of the 1990s was expensive because people did everything by hand. The AI-augmented local team of the next decade is different: a small core of senior people who understand the business, using AI for the boilerplate, the tests, the documentation, and the first-pass implementation of clearly specified features. Fewer people, higher judgment density, and a cost profile that no longer depends on a wage gap. The practical question for every IT leader is which side of the commodity line each piece of work falls on. Boilerplate integration code, standard reports, data migration scripts: these go to AI, wherever the team sits. Cutover strategy, architecture decisions, the conversation with the business about what the system should do: these stay with people who have context. The mistake to avoid is assuming the line maps to geography. It does not. A local junior using AI on well-defined tasks will outperform a remote team doing the same, at a similar cost, with none of the handoff tax. That is the whole trade in one sentence.
The bridge is behind you
I will leave you with the image I keep coming back to. It is the end of a transformation program, run by a team a third the size of the offshore-heavy one that would have staffed it a decade ago. The cutover weekend runs with people in the room who know the business, the AI handled the migration scripts overnight, and the program closes on time because there was no twelve-hour gap between the question and the answer. That is what the third era looks like from the inside: not cheaper per hour, but radically fewer hours, and the hours that remain spent where they matter. Offshore outsourcing was a bridge that got the industry across the river during the scale crisis of the internet era. We have crossed the bridge, and the maintenance bill for keeping it standing is now the thing draining IT departments of their judgment work. Your offshore program will end eventually. The only open question is who ends it: you, deliberately, while the money and the momentum are on your side, or the market, one renegotiated contract at a time. The good news is that you are allowed to choose. Start by running the fully loaded cost. The number will surprise you, and it will tell you which era you are actually living in.
Sources
Frequently asked
Is offshore outsourcing dead?
No. Run, support and scale work still justify offshore delivery. What AI erodes is the labor-arbitrage advantage on judgment-heavy, net-new development work, and even that is a directional shift, not a completed transition.
Does AI help offshore developers more than onshore ones?
The evidence points that way for routine work: Brynjolfsson, Li and Raymond found the biggest productivity lift went to the least experienced workers. That is why the defensible claim is not that AI makes onshore cheaper per hour, but that it shifts competition from cost to judgment and context.
How should an IT leader read the three eras?
As the article's own framework, not an established historical taxonomy. The practical takeaway is to run the fully loaded cost of the offshore program and decide which side of the commodity line each piece of work falls on.
Need this in your organisation?
I work with a small number of clients each quarter on ERP strategy and IT-department automation. If the questions raised above are live in your team, get in touch.
Start a conversation →