At on-demand work service Mercari Hallo, our concept of work is simple: “quick and easy for everyone.” With this in mind, our service connects partner companies urgently looking for workers with crew (workers) urgently looking for work. Just over a year since our launch, over 11 million people have registered for our on-demand work service, making us number-one* in terms of new registrations in the industry.

*Based on the number of users registered for on-demand work services operated by member companies of the Japan Spot Work Association between February 2024 and February 2025. (only available in Japanese)

At Mercari Hallo, we carefully review all job listings to ensure that our partners post jobs that comply with regulations and offer a safe, secure working environment for our crew. In addition to manual checks, we use AI/LLM technology to filter listings to verify their safety. Our integrated approach of leveraging both human expertise and technology enables us to accurately predict latent risks in job listings.

In this article, we talk to Maiko Hiyama (@michael), lead of Mercari Hallo’s job listing quality project, engineering manager Shosaburo Minamoto (@gen), and Tomoyuki Habu (@t-habu)* and Yuki Yada (@arrow), who work on projects related to generative AI, about operations and Mercari Hallo’s new feature for predicting latent risks in job listings.

*@t-habu could not join the photo shoot.

The key to more reliable job listing screening lies in combining human expertise and technology

——Could you start by telling us what job listing screening entails?

@t-habu：Job listing screening refers to checking job listings to make sure that they meet our standards for suitable content and don’t contain any inappropriate wording. We screen listings based on a wide range of criteria. As we are the platform on which they are posted and responsible for the content, we screen all the job listings posted on Mercari Hallo. Only listings that pass our screening are posted.
Each month we process tens of thousands of job listing requests, making sure not to post any listing that pose a risk to us or our partners and crew. To do this, we employ a combination of manual checks and the latest technology.

@michael：Previously, we only screened listings manually. However, as the owners of the platform and with the recent rise of AI, we made it our priority to find a way to screen listings more thoroughly. At Mercari Hallo, we aimed to introduce a primary filtering process using generative AI and similar technologies to supplement our traditional human checks and improve screening accuracy. This system is now operational.

——Has using AI/LLMs changed the screening process or improved the accuracy of risk prediction?

@michael：First, the AI/LLM engine screens each job listing against specific categories, classifying them as either “high-risk” or “low-risk.” Then, a human checks job listings that have been flagged as misrepresenting the actual work involved or likely to violate laws and regulations.

@t-habu：Yes, a human performs the final check. By automating the detailed initial screening using AI/LLMs, we can redirect human expertise to the most critical stages of the review process.

@gen：We are confident using LLMs as they have great contextual understanding and a large vocabulary, which has been backed up by research. Using both humans and technology at different stages of the screening process has enabled us to identify all potential risks.

Also, LLMs allow us to fine-tune and modify evaluation criteria to suit various needs, making it a reliable option for long-term operation. We can simply alter the prompt to change how it operates, which means that we can quash any new tricks to bypass the screening process as they appear.

@t-habu：From an operational standpoint, it’s important to progress from the design stage to the verification stage quickly. It only took around one month from planning to implementation, which was a great accomplishment. As platform owners committed to providing safe and secure on-demand work, I believe this represents our fastest new feature implementation to date.

Establishing a framework to continually ensure a safe and secure service for our partners and crew

——What effect do you think improving screening accuracy and risk prediction has had on Mercari Hallo’s UX?

@t-habu：Well, firstly, I think these improvements have made work safer and more secure for our crew. The crew are the ones who are in danger and at a disadvantage if they work a job that’s potentially illegal. Protecting the crew is our main goal, and also an advantage our service offers.

Also, for partners who make the job listings, we minimize the risk of their jobs unintentionally violating laws and regulations. For example, some job listings state that only people of a certain gender can apply, but such a restriction may be illegal depending on the role’s specific requirements and the precise wording of the listing. Posting a listing that violates laws and regulations, even unintentionally, can damage a partner’s reputation. We perform a thorough check of all listings before they’re posted to protect our partners from such legal risks.

@michael：These advancements have made the service safer for both partners and crew, and have also improved the screening process we use as platform owners—now, we are able to screen listings efficiently and accurately.  This has allowed us to reallocate our human resources more effectively, ensuring that no potentially harmful job listings are posted.

Yielding both speed and quality through each member taking ownership of the problem

——You mentioned that you were able to use AI/LLMs for risk prediction in just one month. Why do you think you were able to achieve this so quickly?

@michael：I think one key factor was that there was great teamwork between the teams involved, that is, the team who identified the problem and the team who had the knowledge of how to solve the problem. Without such effective collaboration, it would have taken longer to figure out what options we had to solve the problem. @arrow, who is knowledgeable about LLMs, was everywhere when we needed him and was a great help!

@arrow：I suppose that’s true. I was already aware of the problem at hand, and I think our progress was facilitated by the ease of communication I experienced with you, @michael, and @t-habu. We also collaborated with talented backend engineers, such as @gen, who led the system side of the project for us.

@gen：You made things easy for us engineers because you were very proactive in suggesting ideas. Even those outside the development team actively made suggestions and gave opinions regarding checklist items, for example.

@t-habu：For this project, we formed a small team focused on producing results quickly and set up daily standups to facilitate progress.

With the holidays fast approaching, everyone had the same goal of completing their work by the end of 2024, spurring us on to work hard and produce results in a short period of time.

Unlimited creativity—the anticipated evolution of Mercari Hallo and LLM usage

——Lastly, could you tell us about your future outlook on Mercari Hallo and AI/LLM usage?

@t-habu：In terms of job listing screening, we plan to expand the number of categories of listings that can be reviewed by LLMs. Also, currently there are some types of listings that are only checked by LLMs, and we can make these checks more accurate by adding a rule engine to the LLM.

@arrow：We want to make Mercari Hallo’s user interface more user-friendly—in the future, we want to expand services to smaller business partners, but some owners of small businesses struggle to use technology.

@michael：We’d also like to use AI/LLMs for sales enablement*. We want our sales staff to focus on sales work and negotiations, so using AI/LLMs to draft proposals and analyze phone calls and meetings would be a great help. This would also help them present more accurate sales proposals to prospective partners. We are also looking forward to AI/LLM technology automatically integrating with CRM tools.

*Sales enablement refers to initiatives aimed at strengthening and improving the sales organization.

@gen：I think AI/LLMs could also help with refining sales strategies and researching target companies, for instance. In this new technological era, even sales work can be made more efficient!

Text : Yuri Kato 　Photo : Tomohiro Takeshita