This Time, AI Doesn't Have an Alibi

A few weeks ago, I argued that AI did not kill the entry-level job, and the timeline backed me up: the decline started years before the models were good enough to matter. This piece is the other half of that story. An alibi that holds through 2025 is not an alibi for 2026 and everything after. The honest question is no longer whether AI will remove work. It is which work, how fast, and what a leader should do about it.

Start with a fact that looks like a coincidence and isn't. The jobs most exposed to AI today are, to a striking degree, the jobs we shipped overseas in the 2000s. Call centers and first-line support. Claims processing, data entry, bookkeeping, and first-line IT. Plus, a large share of software engineering: QA and entry-level development moved to India, China, and Eastern Europe, while the architects and senior engineers stayed on staff. We have relocated this exact category of work before, and we called it offshoring. It was an arbitrage on geography, the identical work bought in a cheaper labor market. What is happening now is a second arbitrage on the same work, except the destination is not another country. It is a model. AI agents like Claude, ChatGPT, Hermes, or OpenClaw (for the brave ones) now do the tier we once offshored, and then some.

The reason the two waves rhyme is buried in how offshoring worked. To hand a process to a team 10,000 km away, you first had to write it down. Script it, break it into tickets, test plans, and service-level agreements, and strip out the tacit judgment so a stranger could run it to standard. Offshoring was, without anyone intending it, a twenty-year codification process. It turned tacit work into explicit process maps, and those maps are precisely the structure a machine needs to run it. Offshorability, it turns out, was an early read on automatability, because both measure one property: how much of a job is written down, and how little of it is judgment.

That gives a leader something more useful than dread. The test that decided what could be offshored in 2006 is the same test that decides what AI runs well in 2026. A process that is clean enough to send abroad, clearly defined, repeatable, and measurable, is a process clean enough to automate. And that clarity is also what keeps a model reliable: a well-specified process acts as a guardrail, starving the ambiguity a model would otherwise fill with confident nonsense. So one question does three jobs at once. How clearly is this process defined? is the offshorability question, the automatability question, and the hallucination-risk question, one property across three eras. The cleaner and more measurable the work, the closer AI already is to doing it.

This is no longer a forecast. It shows up first where the work was most codified. India's IT-services firms, the companies that built the offshore delivery model, have cut fresher hiring by roughly 80% from their early-2020s peak as they pivot to AI-first delivery. The tooling is mainstream rather than fringe: by late 2025, around 90% of developers reported using AI coding tools, and over 80% said those tools made them more productive. Boilerplate, test coverage, routine tickets, the scriptable tier is exactly what those tools do best. Support and the back office are on that path too, for the reason above.

The tool, though, is only an amplifier. The 2025 DORA research found that AI magnifies whatever system it lands in: strong engineering practices compound, and weak ones only get faster and less stable. The gains come from the workflow around the model, not the model itself. The bottleneck was never the typing. It was the handoffs, the reviews, the queues, the shape of the work. Offshoring taught this lesson the expensive way: lifting a broken process to a cheaper location did not fix it; it exported the mess. Automating a broken process in place only repeats it faster. A machine bolted onto a human-shaped workflow inherits every bottleneck it has.

When the machine is genuinely better, redesign is the reason. My Oura ring developed a failing battery within a year, and the entire warranty replacement was handled by an AI agent, start to finish, in two or three minutes, with no queue and no need to repeat myself to a chain of people. It was faster and cleaner than the human version had ever been. What made it work was not the bot. It was that they automated the friction, the queue, the intake, the status chase, and then trusted the machine with a real decision that cost the company money: issue the replacement. Contrast that with the reflex most firms reach for, dropping a chatbot in front of the old script. In support, AI deflects far more than it resolves; deflection rates run high while verified-resolution rates sit far lower, often in the low 40s. A deflection that never resolves is a customer giving up, the old process with a shinier bottleneck. The principle is small, and it is the whole game: automate the friction, redesign the flow, and point the human at judgment.

That redesign question matters most for the places that took this work first. India's technology and business-process sector employs more than 5 million people and stands among the largest pillars of its economy; the Philippines' equivalent employs nearly 1.9 million workers and accounts for more than 8% of GDP. These are national systems built on codified work, now meeting the technology that runs codified work best, and the hiring cuts are the leading edge of it arriving.

The reflex is to read that as a slow-motion disaster. I read it as the opposite. The very fluency that exposes them, running codified processes at enormous scale, is exactly the capability the AI era rewards, provided it climbs a level: from executing the process to designing, supervising, and auditing the AI that executes it. The largest concentration of people who understand how these processes actually work is an asset for building the layer above them. And markets with no legacy to defend tend to skip a generation. Much of sub-Saharan Africa never built out landlines and went straight to mobile; Kenya's M-Pesa put mobile money in most adults' hands while much of the West was still standing in line at a branch. An offshore delivery firm has no on-premises workflow to protect and every reason to rebuild delivery AI-first, which is a real path to leapfrogging incumbents who are busy bolting AI onto processes they refuse to change.

Pilots have a phrase for the discipline that all of this asks for: "stay ahead of the plane." Configure now for the state that is coming, not the one that already arrived, because the pilot who falls behind the plane spends the flight reacting to things that already happened. Staying ahead here means designing for the human-and-machine combination that is arriving, rather than defending the human-shaped process that is leaving. The economics reward it. Making execution cheaper has never shrunk the amount of work worth doing; it enlarges it (economists call it the Jevons paradox), and the roles it opens, redesigning workflows, supervising models, owning the exceptions and trust, are the ones worth moving toward. The winning unit was never the machine alone or the human alone. It is the workflow deliberately built around both.

So the Monday-morning move is a small one. Pick the most codified process you own and resist the urge to pave it exactly as it is. Ask what it would look like if you designed it today from scratch for a person and a machine working together, then build one step toward that, rather than a bot in front of the old script. The work already left once. Whether it leaves you behind this time comes down to whether you are willing to redesign it, or only to re-staff it. Stay ahead of the plane.