mercan

Mercari’s Bold Move to Become AI-Native—The Story Behind Creating an AI Task Force of 100+ Members

2025-9-1

Mercari’s Bold Move to Become AI-Native—The Story Behind Creating an AI Task Force of 100+ Members

Share

  • X
  • Facebook
  • LinkedIn

In May 2025, Mercari announced its transformation into an AI-Native organization.
The announcement was made as a Slack post by Group CEO @suadd (Slack name of Shintaro Yamada).

This does not only apply to specific companies/divisions or engineering roles; it’s a change that involves every single one of us.

We will not be a company that merely implements AI; we will redesign our entire organization premised on AI.

Just a few weeks after this message was posted, the AI Task Force, a cross-functional organization comprising over 100 Mercari members, was formed. Engineers who are well-versed in AI technology account for 40 members of the Task Force, and have been reassigned to the Task Force as their main focus. This core team of over 100 will drive the company-wide effort to audit existing operations and redesign workflows around AI (as of July 1).

We are already in awe at the speed and scale at which they work. In this Mercan article, we will look at the kickoff event for the AI Task Force.

Mercari is not the only company or organization that needs to become “AI-Native.” Many companies are now realizing that they cannot shy away from AI. We hope that this article will provide valuable insight for companies starting their journey with AI.

The real meaning behind @mark’s statement, “We can now bet on AI”

The first words spoken at the AI Task Force kickoff were from @mark, who oversees Mercari’s Japan business. @mark’s address was not a passionate speech filled with buzzwords, but rather a simple statement that reflected the effort and struggle behind all the preparation up to that point.
“We will make a serious bet on AI. The technology has now advanced to a point where we can say that. In fact, it took three to four years to get to this point.”

For a long time, Mercari has focused on becoming more of a tech company and leveraging AI. However, for these ideas to evolve beyond mere concepts and into a serious, company-wide redesign, it was necessary to build a technical foundation, foster the right culture, and most importantly, build a reserve of willpower.

@mark emphasized that, “We are trying to change not the tools or features we use but the fundamental way we work.”
“We need to face the experience of working in the age of AI head on. Let’s focus on redesigning our work experience (WX), the same way we do for products and UX.”

We are not going to be a “company that uses AI.” We will be a company that works together with AI. We now understand that this structural transformation is the very heart of being AI-Native.

“Create an organization that can operate autonomously in six months”—@kimuras’ call to action

Mercari CTO @kimuras took to the stage next. His tone was calm, yet there was a strong sense of concern and expectation as he spoke.

“The goal of this Task Force is not to finish auditing operations and create proofs of concept (PoCs) within the next six months. We want to go further and create a state where all teams across Mercari can operate autonomously.” 

As @kimuras explained, we want to go further than simply using AI to assist with daily tasks, and establish an organizational culture in which operations will continue to improve with AI over the next six months. That is the true nature of this transformation, and what makes it difficult. After stating the overall objective of the Task Force, @kimuras presented three specific policies:

  • Project Managers will dive deep into business structures and AI technology, adopting an engineer’s (Enabler’s) perspective
  • Enablers (engineers) will form hypotheses and prototype solutions, taking the lead rather than waiting for direction
  • Both these roles will collaborate cross-functionally to involve others and spread knowledge

“Over the next six months, I want to change the way we work while leveraging AI myself.” 

This strong declaration from a leader of the company sparked a quiet realization across all members: their daily work is the true stage where AI will shine.

Insights from en. of Nulogic: “Reconstruction,” not “AI implementation”

In the second half of the kickoff, we welcomed an external guest, Nulogic representative en., to give a talk about practical tips and key insights for implementing generative AI.

en. specializes in connecting AI technology with UX and has supported AI implementation in numerous startups and businesses. Their expertise extends beyond comparing implementation tools and prompt techniques to include insight into why things work and how things should be

Introduction: Implementing AI means redesigning thought and structure

To start off, en. stated that there is an important question to ask now that AI utilization is gaining traction, and posed the following to the audience:

“Is generative AI really changing the way we work?”

This question was not about the specific advantages or disadvantages of certain models or prompts. It was asking what it means to implement generative AI into workflows. en. began by framing the issue not as a matter of choosing the right tools or designing PoCs, but as a question of how to change the very structure of our thinking.

One particular phrase in their presentation stood out:

“Implementing AI is not the act of adding something to make a task easier. It’s the act of restructuring your way of thinking.” 

en. then unpacked the meaning of these words, walking through several key principles and sharing over 10 specific examples.

The first golden rule: Analyze work using “input processing output” 

When attempting to implement generative AI into workflows, many people incorrectly assume that they just need to decide on the right tools and find out what the AI can do for them.

The core of the issue is whether you can first break down your work into a simple process: “input → processing → output.”

Starting a PoC without this process might lead to a situation where you input something into the AI and it gives you an output, but you’re unsure what you wanted to do to begin with.

en. wanted to convey that, to avoid this risk, structuring the workflow should be the first step in the process.

For example, for a specific idea like “summarize a year’s worth of chat history from a messaging app as ‘Memories,’ the work can be categorized as follows:

  • Input: Natural language log on LINE (non-standard, miscellaneous text) 
  • Process: Teach Gemini your thought process using the “Chain of Thought” format (a prompt design method that explicitly describes steps for reasoning or thought when processing complex questions or tasks) 
  • Output: Text sorted into categories such as “memory,” “event,” and “important saying” and summarized 

The most important thing here is to create your own definition of a memory before inputting anything into the AI.

“To leverage AI, you must first try to clarify any ambiguities in your thinking.” 

The second golden rule: The prompt is the design outlining your thought process 

Next comes prompt design, which is also very important.

en. explained that writing a prompt is like describing your work to someone else. That is, instead of giving the AI orders without any context, you should provide it with context and use the “Chain of Thought” method to convey your thoughts.

“Imagine a new employee asks you how they should approach a certain task. You would explain that you make certain decisions at certain stages of the process, and do things in a certain order. That’s exactly what a prompt should be like.” 

Going back to our example of classifying text in a chat history on LINE, the prompt would look something like the following:

  • “Focus on the context of honorifics to estimate positive or negative tones”
  • “Instruct the model to not arbitrarily convert other people’s name into kanji”
  • “After extracting the text, output a summary for each classification label”

 As you can see, instead of simply writing “extract memories,” providing instructions on what criteria define a memory, rules for conversion, and your desired output format will ultimately lead to reproducible output.

The third golden rule: Don’t aim for “perfect automation” right off the bat 

Then, en. mentioned the importance of “Human-in-the-Loop.”

“AI is inconsistent and prone to mistakes.  So, failing to set up the AI in a way that requires human verification will ultimately create more operational risks.” 

Take screening resumes as an example. As a task that involves discretion, a process like the following would be best when implementing AI:

  • Use Vision API to perform OCR on the resume PDF and convert it to text (pre-processing step)
  • Use a model like GPT-4o to score how well the text matches the job listing criteria
  • Output the scoring results in a list, and have a human check and make the final judgment

As you can see, starting small and implementing AI to improve specific tasks is an important part of AI culture.

“Human-in-the-Loop is a type of workflow design for the generative AI era.” 

The fourth golden rule: Leveraging general tools and templates = standing on the shoulders of giants

Some people tend to think that they need to create in-house tools in order to implement generative AI, but the frontend of AI tools is always evolving. The UI and APIs are also constantly changing. Despite this, many companies want to make their own tools.

As an alternative approach, en. suggested implementing AI based on the following order of priority:

  1. Leverage general-purpose tools like ChatGPT, NotebookLM, and Gemini in a way that best suits your needs
  2. Add minimal AI-compatible extensions to in-house tools
  3. Use SaaS integration and automation (e.g., n8n)
  4. Develop an in-house tool only when absolutely needed

en. suggested that following this order and allocating more time to the heart of the problem—the redesign of workflows—is more productive.

After explaining these four principles, en. moved into the latter half of the presentation, providing more specific examples and introducing various cases in which they have implemented generative AI.

en. explained the technology used in each case, and also delved into the thought process behind implementation, such as why the AI was designed that way and what failures or discoveries occurred, providing practical hints for participants to relate to their own work.
en.’s examples were based on the “input → processing → output” process for structuring work presented earlier, as well as the principles of “trial speed” and “reusability.” These were all deeply insightful cases for the AI Task Force as well.

Conclusion: Implementation in the generative AI era begins with verbalizing workflows

The words en. reiterated to conclude the session were calm, yet deeply impactful.

“Implementing generative AI means being able to explain your work to others.

It involves transforming ambiguous tasks, individual-dependent processes, and things conveyed through intuition … into structures and prompts.” 

Any work being carried out in a way that only you can understand cannot harness the power of AI. In a way, en.’s statement made us realize that we can start to effectively leverage AI’s power once we articulate and structure our work.

It was interesting to see both the Enablers (engineers) and Project Managers of the AI Task Force nod along in agreement.

“First, try to reframe your work using structure” was the quiet resolve shared among all participants.

What we will create together

This project is not tied to a fancy product launch. It is about analyzing each and every one of our numerous tasks, understanding the structure of our work, and redesigning workflows. While these tasks may seem dull, they will form the backbone of our transition to an AI-Native company.

Mercari is not trying to create any “special magic” with AI. We are simply trying to change by reviewing our teams and tasks, one by one. Over the next six months, we will strive to embody that.

Mercari’s transformation into an AI-Native company has just begun.

As we are still learning, we want to keep publishing honest information about both our successes and failures. Some of you may want to use AI but don’t know where to start, or maybe you created a PoC but it doesn’t work as expected. 

We hope that this article will provide you with some hints for leveraging AI.

Share

  • X
  • Facebook
  • LinkedIn

Unleash the
potential
in all people

We’re Hiring!

Join us