Nippon Life Grades 8,830 Meeting Records

Summary

• Nippon Life re-graded 8,830 meeting records from ten years of engagement with AI.
• It filtered out the AI’s bias toward reading the company’s own remarks generously, and checked more than 10,000 requests against disclosures and financial changes.
• It shows that past qualitative records can be measured with a consistent yardstick and used as a basis for future strategy.

Were ten years of meeting records actually useful

When an institutional investor meets the management of a company it invests in and asks for better governance, a dividend policy, or greener management, the activity is called stewardship. Nippon Life, one of the largest life insurers in Japan, has kept up this kind of dialogue every year with about 700 investee companies. Each officer alone covers close to 100 companies a year.

Japan introduced a stewardship code for institutional investors in 2014 and a corporate governance code for companies in 2015, which sped up governance reform, and Nippon Life has treated stewardship as a core duty for more than 10 years. In the interview video, Tomochika Ishii, general manager of the equity investment department, explained that Nippon Life announced its acceptance of the stewardship code in 2014, then strengthened its setup and increased both the amount and the quality of dialogue. He said that about 700 companies a year is probably among the largest scales in Japan.

But the bigger the activity grew, the more a basic question followed. Had the conversations really influenced the investee companies’ behavior, or had they stayed ceremonial talk? The topics of dialogue had also widened, from the early focus on governance and financial metrics to environmental and social areas. Records piled up without a break for ten years, but it was close to physically impossible for people to reread so many qualitative records one by one and measure their effect by a consistent standard. They could only guess that the work had been effective, relying on each person’s memory and individual impressions.

According to a client case report and an interview video published by the global consulting firm Oliver Wyman, Nippon Life decided to use generative AI not just as a tool for automating work, but as a tool for measuring past records. It would re-evaluate the 8,830 dialogue records built up over the past ten years with one yardstick. And it would use AI to check whether the more than 10,000 governance requests it had passed to companies had led to real changes in the companies’ disclosures or finances.

Setting the grading standard before asking the AI anything

The Nippon Life and Oliver Wyman project focused on confirming three questions. Yasuyuki Sekioka, a partner in Oliver Wyman’s Tokyo office, said in the interview video that the aim was to verify objectively whether Nippon Life’s dialogue had really been of high quality, whether the investee companies had received it well, and whether, as a result, a change in company behavior that can be seen from outside had actually appeared. Nippon Life’s reason for choosing Oliver Wyman as its partner was clear. They had worked together on stewardship projects for the previous two years, so Oliver Wyman understood the company’s way of working and its philosophy deeply. The two already shared goals and standards, so Nippon Life believed Oliver Wyman could produce practical solutions even for a difficult task.

The first step of the analysis was to make the AI understand Nippon Life’s investment philosophy. Then the team defined clearly which conversations count as good ones. By Nippon Life’s standard, a good dialogue was one that is “constructive, specific, and capable of influencing the other side’s behavior.” Once the criteria were set, the team picked 300 meeting records for a pilot test and refined the prompt, the instruction text, step by step.

In this process, a practical obstacle that no one had expected came to light. Generative AI is very good at grasping context. But because of this strength, it tended to read the remarks of Nippon Life’s staff in the records more positively and more generously than the real context warranted. It was as if a grader were guessing the intent behind a short answer from a student on their own team and adding points. The project team filtered out such cases in the sample-check stage. The final prompt, finished after several rounds of revision, came to about ten A4 pages. Tomochika Ishii, general manager of Nippon Life’s equity investment department, explained that this step of analyzing dialogue quality precisely took the most time in the whole project.

In practice, this ten-page prompt was probably not a simple piece of writing but something close to a detailed grading manual. It likely stated the investment philosophy at the start. In the middle it likely held the definitions of grades such as High Quality, High, Good, and Effective, together with rules for judging borderline cases, and at the end it likely set the output format. The team probably did not let the AI grade a whole set of minutes at once. It likely sorted the topics within each record first, and then forced the AI to quote supporting sentences directly from the minutes each time it gave a score. In the 300-record pilot, the team likely compared the model’s scores with scores that staff had given by hand. Each time the model had scored too loosely, they likely added a prohibition rule and a counterexample, such as “do not treat a mere statement of principle as a commitment,” and that is how the prompt grew to ten pages.

With the criteria refined, the AI analyzed all 8,830 dialogue records covering ten years in a consistent way. The original English article calls them “dialogue minutes,” which can read as minutes of conversation time. But as Tomochika Ishii said directly in the video, the figure is the number of meeting records. The team did not rely on the AI’s judgment alone. It ran human-led verification and governance checks in parallel. The human verification probably did not mean going over all 8,830 records again. It was more likely done by drawing samples by year and by topic and checking that the supporting sentences the model quoted matched its grading results. Through this process, consistent patterns emerged from the unstructured meeting records that had been scattered across years, topics, and companies.

Loading disclosure reports and financial data into the system beforehand

Measuring the quality of dialogue was only half the result. The key was to find out whether companies actually moved after the dialogue. In step two, the team sorted the more than 10,000 governance requests that Nippon Life had passed to companies in meetings into quantitative requests and qualitative requests. Quantitative requests were number-based demands such as improving financial metrics. Qualitative requests were institutional demands such as reorganizing governance or expanding disclosure.

The project team loaded the relevant data into the analysis environment in advance. For the quantitative evaluation it put in financial data, and for the qualitative evaluation it put in the securities reports and integrated reports that companies disclose every year. This was probably a device to keep the AI from searching the internet at random and citing wrong information, and to make it look for evidence only inside the documents of a set company and year, so that results could be repeated consistently. The AI cross-checked the content of past requests against later disclosure documents and changes in financial metrics, and judged whether the company’s change had moved in the “same direction” as the request.

Even Tomochika Ishii, who had proposed this analysis himself, was of two minds about the result at first. In an interview for the article, he looked back on his honest feelings at the time.

“We were half-skeptical whether generative AI could really make such sound judgments, but Oliver Wyman managed to build an admirable system environment with AI, which truly surprised us.”

Nippon Life’s healthy skepticism came together with system design and human-led sample verification, and the result was a data analysis framework that is consistent, repeatable, and trustworthy.

The path of stewardship, confirmed as trends and not as numbers

The analysis results were summarized as overall trends and patterns, not as specific percentage figures. First, as the years went by, the share of dialogues rated High Quality grew steadily. The data confirmed that Nippon Life’s dialogue grew more refined and more effective each year. Tomochika Ishii mentioned that the share receiving positive ratings, in both dialogue quality and company response, rose every year.

Gaps between topics also showed up clearly. Dialogues on finance or governance more often received positive ratings such as High, Good, and Effective. Environmental and social topics, which expanded in earnest from around 2018, had more mixed results. Tomochika Ishii said that in practice this gap was something they had somewhat expected. This is probably because finance and governance have standards built up over a long time, while environmental and social issues were relatively new challenges for both companies and investors.

A link between the quality of dialogue and changes in company behavior was also confirmed. When the dialogue was effective and the company’s response was positive, a behavior change that can be observed in disclosures or financial data was more likely to appear.

However, Nippon Life did not rashly claim a causal link, that is, that companies changed thanks to its dialogue. Tomochika Ishii explained that it is true that the share of companies with confirmed behavior change rose every year, but this cannot be credited to Nippon Life alone. Government policy changes, demands from other institutional investors, and the companies’ own efforts at renewal all worked together, so the point was that the results should be read as meaning that its requests and the direction of company change were broadly aligned.

A way to look again at qualitative work that could not be measured

Stewardship is not the kind of work that produces visible results in a short time. It usually takes three to five years, or longer, before a company accepts an investor’s demand and changes its governance or discloses a new metric. So a short-term, one-year focus on results makes it hard to measure the true value of dialogue.

With AI, Nippon Life can now systematically see, over the long span of ten years, where progress has been made, where it should put in more effort, and how the quality of dialogue connects with real company change. With each officer covering close to 100 companies, Tomochika Ishii pointed out that this analysis gave them an opportunity to consider how to improve the way they work so as to raise the share of effective dialogue.

The meaning of this case does not stop at one financial firm’s success story. Inside companies, countless qualitative records are piled up and left alone, valuable but hard to quantify: meeting records, minutes, customer service logs. When many organizations adopt AI, they focus on writing or summarizing new documents automatically. But Nippon Life used AI as a yardstick to evaluate and measure the organization’s own past footprints.

When you want to evaluate an organization’s qualitative activity with AI, the first thing to do is not to choose a tool but to put into writing and rules your internal philosophy and standard of “what is good activity.” You must also correct strictly, through sample checks, a language model’s habit of reading context too positively. And when you define the result carefully as a match in objective direction, instead of dressing it up as an exaggerated causal link, AI becomes a trustworthy measuring tool that turns an area once run on vague instinct into a solid basis for decisions.

Sources

1. Oliver Wyman — How generative AI helped Nippon Life elevate stewardship
   https://www.oliverwyman.com/our-expertise/insights/2026/may/generative-ai-elevate-nippon-life-stewardship.html
   Published: 2026-05-28 (last modified 2026-07-31, author Yasuyuki Sekioka)
2. Embedded interview video on the same page — Elevating Nippon Life’s corporate governance and stewardship with AI (13:26, Japanese)
   https://vimeo.com/1123436648
   Yasuyuki Sekioka (Partner, Oliver Wyman Tokyo) interviews Tomochika Ishii (General Manager of Equity Investment Department, Nippon Life). Video references in the text are based on Japanese auto-generated captions (accessed: 2026-09-24)