AI Didn't Kill Offshoring After All (At Least Not Yet)

One common prediction about AI was that as it got more advanced, it would eliminate a lot of outsourcing: why outsource jobs to India, or other lower-wage nations, if you can automate them? One answer is that it's still cheaper to outsource some jobs, thanks it part to currency arbitrage. 

Before chatgpt, 1 USD = 75 INR

After chatgpt, 1 USD = 100 INR

You can't kill offshoring faster than money depreciation

— J Sam🌐 (@JaicSam) August 4, 2026

Today's guest post, by Filip Pesek, an entrepreneur focusing on executive assistants, offers more detail on why AI hasn't eliminated those jobs yet. Before we get to it, two brief programming notes: 

1) We've got three short-term trades teed up for later today: a post-earnings recovery trade on a chip company we expect to quickly bounce back from its post-earnings drop, and two pre-earnings trades on companies reporting this week.

As usual, we're making those trades available for free subscribers to our trading Substack to unlock. You can become a free subscriber, and get an email as soon as our trade alert goes out, by clicking on the widget below. 

2) If you're worried the market has rallied back too far and too fast, and you want to hedge, our hedging app is back on the App Store for you.  

No on to Filip Pesek's guest post. 

Authored by Filip Pesek at DonnaPro

The White-Collar Repricing Nobody Is Hedging

The consensus about artificial intelligence and white-collar work is that the machines are coming for the desk jobs. Each quarter brings another forecast of administrative roles automated out of existence, another executive survey planning headcount reductions, another essay on the hollowing of the corporate middle.

The labor data does not support it.

Employment of secretaries and administrative assistants is projected by the Bureau of Labor Statistics to show little or no change from 2024 to 2034. The occupation holds roughly 3.5 million jobs, with about 358,300 openings projected annually over the decade. Office and administrative support in aggregate ran to 18.5 million workers at the last full survey, around 12 percent of national employment.

That is not a collapsing function. Anyone positioned around a thesis of administrative headcount destruction will spend years watching a flat line and wondering why the trade will not work.

It will not work because the displacement is real but the mechanism has been misread. The function is not being eliminated. It is being repriced, relocated, and moved off the domestic payroll line entirely. The headcount number stays flat while it happens, so the standard indicators report a false negative for years while the margin transfer occurs somewhere the data does not look.

The Number That Actually Matters

BLS Occupational Requirements Survey data shows telework applied to only 21.3 percent of secretaries and administrative assistants in 2024.

Four out of five administrative support workers in the United States are still physically co-located with the people they support. In 2026. After the remote-work experiment, after asynchronous tooling matured, after AI became competent at the structured layer of coordination work.

The technology to distribute this function has existed for years. The function has mostly not been distributed. That gap is the entire story: an enormous pool of labor cost structured as fixed, local, and full-time, sitting in a market where the cost of restructuring it as variable and distributed has fallen sharply and keeps falling. Every point of that 79 percent that converts is a fixed-to-variable conversion on the corporate expense base — and unlike a layoff, it does not show up as one.

Where the Work Is Going

India’s global capability center ecosystem now comprises more than 2,100 centers employing over 2.3 million professionals and generating close to $100 billion in annual revenue, per the Nasscom-Zinnov 2026 report. The sector has compounded at roughly 19 percent a year since 2022 and is projected to hold that pace toward 2030.

The qualitative shift matters more than the growth rate. The EY India GCC Pulse Survey 2025 documents a systematic move from labor-cost arbitrage toward capability arbitrage. These are no longer back-office processing floors. They are absorbing judgment-bearing functions once considered non-transferable because they required context and real-time coordination.

AI made the transfer viable, not by replacing the humans but by collapsing the coordination overhead. Context that once demanded physical proximity now moves through documented systems, AI-assisted briefing, and asynchronous protocols. The human still does the judgment. They no longer do it from the next desk.

Compound growth near 19 percent in offshore capability against flat domestic administrative headcount, in the same functional category, is not coincidence. It is a transfer.

What AI Does to the Seat

Wage differential alone understates it, because the same systems act unevenly on the seat itself.

The most rigorous field evidence available is a study of 5,172 customer-support agents at a Fortune 500 firm, published in the Quarterly Journal of Economics. A generative AI assistant raised productivity by roughly 15 percent on average, measured as issues resolved per hour. The average is the least interesting number in the paper. Gains ran to 34 percent for novice and low-skilled workers and to approximately zero for the most experienced and highest-skilled, who registered small speed gains alongside small quality declines.

Read that distribution against the composition of a newly built distributed support function. The gains land almost entirely on the inexperienced seat. The expensive, long-tenured, co-located incumbent captures nearly none of them.

The same study found an agent with two months of tenure using the tool performed at the level of an agent with six months of tenure without it, and that AI assistance measurably improved English fluency, with the effect concentrated among international agents.

Tenure compression and language fluency were two of the last genuine frictions holding executive-adjacent support work in place geographically. Both have now been measured in a controlled setting, moving in the direction that favors the distributed seat. The arbitrage is not simply lower cost for equivalent output. It is lower cost against a quality gap the technology is closing from the bottom.

Why the Indicators Will Miss It

A domestic role eliminated and replaced by an offshore seat shows up as one job lost. A domestic role never created because the function was distributed from the outset shows up as nothing. The second category is almost certainly larger and is entirely invisible in employment statistics.

A firm shifting from a full-time in-house hire to a retained external arrangement records lower payroll and higher operating expense. Headcount falls by one. Nothing about it registers as automation, offshoring, or displacement in any standard dataset. At the smaller end of the market this is already the default: boutique funds, family offices, and founder-led firms have been restructuring support toward distributed arrangements for years, selling a fixed-to-variable conversion with an AI-assisted delivery layer behind it. No employment statistic captures the transaction. Multiply it across the long tail and the aggregate is large, and unmeasured.

The Second-Order Effects

The operating leverage of small firms improves as a fixed cost base turns partly variable, changing the survival math for the long tail of asset managers and advisory practices in a drawdown.

Commercial real estate in administrative-heavy markets faces a slower, less visible erosion than the return-to-office debate frames it. The question is not how many days a week employees attend. It is how many seats are being structurally removed from the domestic footprint.

And the policy response to AI-driven displacement is being calibrated against data that will not register the actual mechanism. Retraining and displacement frameworks built around headcount destruction will address the wrong phenomenon. These workers are not being replaced by software. They are being competed with by a distributed labor pool software made accessible.

The Base Rate

The median annual wage for secretaries and administrative assistants was $47,460 as of May 2024. Across office and administrative support, mean wages sat near $48,000 against a national average above $65,000.

That is a wage base of several hundred billion dollars a year in the United States alone, attached to a function that is stable in headcount, 79 percent co-located, and increasingly deliverable from anywhere at a fraction of the cost.

The narrative says AI is coming for these jobs. What is happening is that AI removed the last practical constraint on moving them, and the market is working through a repricing that will take a decade and barely register in the statistics everyone watches.

The displacement is real. The mechanism is not the one being priced.