What AI Actually Did to Offshore Support Work
The most cited study of AI at work followed 5,172 agents who were mostly offshore. It found the opposite of displacement, and nobody mentions where they were.
The single most cited piece of evidence about artificial intelligence in the workplace is a study of customer support agents. It is usually described as a study of American workers. It was not.
Of the 5,172 agents in the sample, 89% were outside the United States, mostly in the Philippines, working for a Fortune 500 American software company both directly and through third-party outsourcing firms. That makes it the best available evidence about what these tools do to offshored knowledge work, which is a live question in Nairobi.
What the study found
Access to an AI assistant increased issues resolved per hour by 15% on average. The distribution matters more than the average. Less skilled and less experienced agents improved across every measure, with a 30% increase in issues resolved per hour. Agents with two months of tenure performed as well as untreated agents with more than six months.
For the most skilled agents, the effect was close to nothing, and the quality of their conversations declined slightly.
The study also found something easy to overlook: AI assistance improved English fluency, particularly among international agents, and customers were more polite and less likely to escalate to a manager.
Why this cuts against the displacement story
This is evidence of complementarity in exactly the occupation that dominates business process outsourcing. The tools made offshore agents faster, made their customers less hostile, and compressed the experience curve so that new agents became useful sooner.
That is not a small thing for a country building a digital services sector. It is also not a guarantee: a company that needs fewer agent-hours to handle the same volume can employ fewer agents. The study measures productivity, not headcount.
A number worth getting right
Almost every secondary source quotes this study as finding 14% average improvement and 34% for novices. Those are the figures from the 2023 working paper. The peer-reviewed version published in the Quarterly Journal of Economics in 2025 says 15% and 30%.
It is a small difference and it does not change the conclusion, but a publication that cites the superseded numbers is telling you it did not open the paper. We have written separately about why the "40% of African outsourcing jobs" figure is also wrong, about the skills Botswana says it needs, and about what the developer productivity studies actually show.
Frequently asked
Will AI replace outsourcing and support jobs in Botswana?
The best available evidence points to complementarity rather than replacement. A peer-reviewed study of 5,172 support agents, 89% of them offshore, found AI assistance raised productivity 15% on average and 30% for the least experienced. It measured productivity, not headcount, so it does not settle whether firms will employ fewer people.
Who benefits most from AI at work?
Consistently, the least experienced. In the support-agent study the least skilled improved 30% while the most skilled gained almost nothing and saw a small quality decline. The same pattern appears in studies of consultants, writers and software developers.
What is the correct figure from the Brynjolfsson, Li and Raymond study?
The peer-reviewed Quarterly Journal of Economics version reports 15% average productivity improvement and 30% for less skilled and less experienced workers. The commonly quoted 14% and 34% are from the earlier 2023 NBER working paper.