Hisayuki Idekoba, the CEO of Recruit Holdings, was on a recent panel along with Vladimir Lukic of BCG to discuss the future of work in Japan and the US. He talked about workforce trends in both countries, how Recruit is utilizing AI in its business to improve the productivity of its staff and some of the challenges in terms of adoption, mistakes etc. Some interesting insights for investors focusing on the “3D” (demographics, digitalization and domestic capex) themes in Japan.
Some notes are below the embedded video (not a full transcript, and any mistakes are my own).
Demographics: Japan’s elderly population (65+) will continue to grow until ~2044, while the average age of Japanese workers is going up. In 2040, they predict there will be a labor shortage of ~11m people in Japan. Japan needs to be all in on AI. If you travel to Japan, you will already see that many restaurants, for example, are struggling to hire servers. While the demographic story is well known in Japan, many people don’t realize that the US is starting to look like the new Japan.
What he means by that is the demographic situation in the US is like Japan ~20 years ago. The working age population is starting to level off. In Jan 2020, right before the pandemic, there were 163m workers (135m native-born, 28m foreign-born). As of Jan 2024, there were 166m workers (135m native-born, 31m foreign-born). Almost all the growth has come from foreign-born labor.
Realistically, both countries will need a certain amount of immigrants. In the US, 1 out of 4 healthcare workers are immigrants, 1 out of 5 nurses are immigrants. But the growing challenge is we don’t have a good pipeline of skilled labor. All developed countries have to think about what is a good balance between labor supply and demand and what is a realistic timeline to improve the productivity of the workforce. The other thing people forget in the debate between AI vs. immigrants is that only immigrants can also be a consumer in the economy.
On AI & the labor market: When they looked at data on the share of US job postings using AI-related keywords, there was actually a big dip after Covid-19, but it is now starting to pick up a little bit. Gen AI jobs are still rare, but they are growing quickly. In their analysis of job postings on Indeed.com, they found that less than 20% of jobs currently face a high potential exposure to Gen AI (which means that Gen AI can already perform at least 80% of the required skills at a “good” or “excellent” level). In Japan, Gen AI usage is still very low, and tends to be higher among men and younger people. In a conversation he had with the President of MIT recently, he framed it as “before Gen AI, humans do the work, and machines do the QA. With Gen AI, it will be more of the opposite.”
At Recruit, for example, they have delivered a generative AI model to help employers write better job descriptions. So far, more than 800,000 employers on their platform have used it. The employers who use it get ~16% more applications. They are also using AI to help recommend jobs to job searchers.
When they first approached it a few years ago, they were trying to automate a recruiter’s job. But over time, they started to realize it is very difficult to automate. They broke it down to describe the job process end to end, analyze where AI might make a difference and then worked on deploying it. First, they needed to separate out each part or process of a recruiter’s job and they realized they have to do a lot of things. But perhaps you can automate a certain step or part. So out of 10 processes, for example, perhaps 1 is relatively easy to automate. At the same time, for the human beings doing these jobs, it can be very tiresome to do the same set of processes over and over again.
On providing recommendations for job seekers, what they saw was that if a recruiter has been doing the job for 10 years, they are very good at recommending jobs. But for the recruiter who just started, there is a big difference in the quality of their recommendations. By supporting them with the AI recommended model, they saw a big jump in productivity for this part of the process especially for people who started in the job recently. Recommendations are also a good use case of AI because it is not a binary answer, there is no definitely right or wrong answer, so it was relatively easy to utilize that new technology.
On AI making mistakes: you need to contextualize the process, look at what is the improvement and if there is an improvement, then use it and start evolving and get the learnings back. For example, when they got push back about AI making mistakes when drafting job posts, the first question is what is the approval rate of a first draft when a human writes it? 30%. Ok, what is the approval rate when the machine does it? 70%. Yes, it might sometimes “hallucinate” but that is also a lot easier to detect. So it is much more improved and easy to detect when it is wrong. So why don’t you use it? They need to pile up failures when using AI technology because that is the best way to train their models.
Adoption and attitudes to AI in Japan: He is actually surprised how Japanese people are welcoming robots and AI, perhaps much more than in many other developed countries. They don’t seem to view it as much as a threat. There are many opportunities to help improve productivity in factories, the manufacturing sector and even in the service industry. But is also requires a lot of time to build trust. What they see from their own data is that when they put a human being as an interface for customers, the success rate is way better (even if the content delivered through human or a screen is the same). If anything, we might be trusting human beings too much. But with a younger generations having AI support from day one, they might have a totally different sense of trust.
Can AI be used to assess candidates? They have been using predictive AI to build assessment tests. But assessments in the US are a very regulated space and he thinks that is a good thing generally. If a bunch of startups are just assessing people with AI, nobody knows which is right or wrong. There are also questions about the use and protection of people’s data. He thinks there should be some standards, like the Fair Credit Reporting Act when somebody is applying for a financial loan. Job applications and a person’s data should be treated similarly. There are some Chinese startups already using AI to assess and rate people, but he is not so sure yet whether that is a good thing or not.