AX Field Notes

All casesNo. 056

Mizuho Turns Its Clerical Group into Process Designers

5,000 people's worth of clerical work, not 5,000 clerks: process design first, then AI

Case no.
056
Primary source
Company presentation at a Labour Policy Council hearing (MHLW) + press reports
Source published
2026-10-02

1. Why this case

On 2026-02-27, media reports said Mizuho Financial Group planned to use AI to replace about 5,000 clerical jobs over ten years. Bloomberg’s report, which described Mizuho as “the country’s third-largest lender,” carried a company statement: “We are planning to enhance our earnings capabilities by shifting human resources to our focus areas by fully utilising AI,” and “It is not a headcount reduction.”

Seven months later, on 2026-10-02, a deputy general manager from Mizuho FG’s Process Planning Department explained the program at a hearing of the Basic Labour Policy Subcommittee of Japan’s Labour Policy Council, run by the Ministry of Health, Labour and Welfare. The 24-slide deck he submitted is the primary source for this case. It covers what Mizuho changed in its organization, the AI foundations it built for all employees, and four projects where AI now reads documents and makes first-pass judgments, each with hours saved per year.

The case is useful for three reasons. First, the company states its target in units of work, not people, and it published the five places where it says the press got that wrong. Second, the change it describes first is organizational: the group that ran clerical operations was rebuilt as a group that designs processes. AI comes after that in its account, and the deck’s summary states the order as a principle: “foundation, data, and process transformation, then AI.” Third, the audience was a labour policy body, so the deck speaks directly about headcount decline and about what happens to the time AI frees up.


2. Problem

2.1 A workforce that will shrink on its own

The deck’s reason for the change (slide 8) is demographic: “given the organization’s demographic structure, headcount is expected to ‘decline naturally’ over the next 10 years.” From that it draws the requirement: technology must cover the gap between the clerical capacity the bank needs and the people it will actually have. The chart that quantified this gap was marked “not disclosed.”

2.2 Resources tied up in keeping things stable

Slide 7 describes the starting point: “previously, much of our resources went to maintaining ‘stability,’ and we built a solid, reliable clerical setup.” Inside those stability resources, the slide marks a block labeled “room for process kaizen,” with six kinds of waste listed: sideways, back-and-forth, inconsistent, excessive, roundabout, and outdated. The resources spent on “transformation” were the smaller share.

The clerical function’s own identity matched this. Slide 9 places the words associated with clerical work (accuracy, stability, carefulness, backstage support) at the “static” end of a spectrum, and the words associated with process work (flexibility, improvement orientation, standardization) at the “dynamic” end.

2.3 The plan was read as a headcount cut

When the plan became public, several reports described it in terms of people. Slide 5 shows a newspaper article, a Bloomberg Japanese article, and a TV Tokyo news caption, and says the reports drew a large reaction from early morning on 2026-02-27, with follow-up articles spreading beyond major media.

The slide lists five “points of factual misreading,” each with Mizuho’s correction:

  1. “Cut 5,000 clerical staff” → cut 5,000 people’s worth of clerical work
  2. “At subsidiary Mizuho Bank” → across the whole Mizuho group
  3. “Introduce AI at the operations centers” → introduce AI to clerical work company-wide
  4. “Many clerical staff will become unnecessary” → much clerical work will become unnecessary
  5. “Reduce clerical staff by one third” → reduce clerical work

Slide 6 shows Mizuho’s own tally of the reaction on X. Posts matching “Mizuho” and any of AI, cut, clerical, or placement numbered 22,776 in seven days, 15,471 of them on the first day. Mizuho classified 87% as neutral, 12% as negative, and 1% as positive, and it counted posts aimed at anyone other than Mizuho as neutral.

Confirmed facts: headcount expected to decline naturally over 10 years (slide 8); gap between required clerical capacity and headcount must be covered by technology; gap chart not disclosed; resources previously weighted to “stability”; five corrections of 2026-02-27 reports, with Mizuho’s own unit being “5,000 people’s worth of clerical work” across the group; 22,776 X posts in 7 days, 87% neutral by Mizuho’s classification.


3. Solution

3.1 The clerical group becomes the Process Design Group

The central change is organizational. Slide 9 says Mizuho “set up a new Process Design Group, built on the clerical [operations] group,” and “newly set up a Process Transformation Department as a company-wide, cross-cutting unit.”

Slide 10, titled “Conclusion: what did we change in the end,” states the new remit. The group’s job is now explicitly “designing business processes across the company,” process transformation is “recognized as a key management issue,” and the work involves “close discussion with the management layer.” The organization’s stated goal is to optimize “the business processes of all of Mizuho end to end” and become “an organization that designs processes that bring operational efficiency and stability.” The slide’s last line redefines it as “the hub of process transformation.”

3.2 Shifting resources from stability to transformation

Slide 7 lays out the timing. In FY2025, Mizuho “increased transformation resources and accelerated ‘consolidation, review, and digitization.’” Under the heading “process design,” the freed resources come from “review through MPS (Mizuho Process System), AI use, digitization, and similar measures.” For FY2026 onward, the chart draws the transformation share larger again and labels it “the main targets of consolidation, review, and digitization.” Across all three periods, a dotted band states that risk management, regulatory response, new products and services, and business continuity planning continue and are strengthened. The chart carries no numbers, and the slide that explains MPS was not disclosed.

3.3 Foundations for everyone: tool, people, community, in-house builds

Before the work-specific AI projects, the deck describes four “foundations for AI use.”

A tool at the top of the portal. The internal generative AI chat tool sits “at the top of the internal portal as a tool opened every day” (slide 12). This is Wiz Chat: slide 17 describes Wiz Chat as the general-purpose generative AI chat tool, and a slide 16 post refers to each employee’s “WizChat environment (Gemini 3.1 Pro, etc.).” A screenshot of its model list shows seven models, each with a stated use: GPT-4.1 as the main model, Gemini 3 Flash for long documents, Gemini 3.1 Pro and Gemini 2.5 Pro as reasoning models for strategy and planning, GPT-5.2 as the newest high-accuracy general model, and GPT-image-1.5 and Gemini 3 Pro Image for images. “Use cases suited to each model’s traits” are added daily. The visible menu groups them into sales activities, internal work (documents, meetings, research and analysis, risk management and compliance), and everyday tasks (translation, email, contract checks, programming).

Culture and behavior measures (slide 13). The slide’s stated approach is layered measures to build a culture of “taking on challenges and transforming,” accepting that “it is one-way at first” and changing mindset and behavior over time. The counts, with the workshop and knowledge-session figures dated FY2025 on the slide:

  • More than 95% of Process Design Group members hold Stage 1 certification in MPS kaizen.
  • Digital workshops ran 202 times and produced 34 projects.
  • Digital knowledge-sharing sessions ran 42 times, with 2,500 cumulative participants.
  • DX Cafe and Power Platform hands-on sessions ran 5 times in 3 regions.

A slide on personnel measures was not disclosed.

An internal community (slide 16). The AI community “Wiz Buzz ~WORK WITH AI@MIZUHO~” is described as the largest community on the internal social network, with more than 10,000 members (the screenshot shows 10,003). It is used for sharing know-how, questions and requests, and event notices. Screenshots show prompts shared company-wide as custom templates for the chat tool, and a report from a one-day “DX Cafe BUILD” where five teams from operations planning, operations promotion, customer relations, HR strategy, and operations audit departments built AI prototypes by “vibe coding.”

Tools built in-house (slide 17). “Company-wide tools are developed in-house, so know-how accumulates and they are improved daily,” and “many PoCs run in parallel, and interested employees can join easily.” The tools shown include Wiz Chat (a general-purpose chat tool in the intranet environment that can use customer data and internal information), Wiz Search for clerical procedures and for credit and lending procedures (for the problem of not knowing “where to look”), a tool that writes meeting records from audio, RM Studio for corporate relationship managers’ sales materials, Wiz OCR for printed and handwritten text on paper forms with item extraction and missing-entry checks, a slide generator, translation tools, and a sounding-board chatbot. Several are marked as PoCs, and some as temporarily paused.

3.4 AI on real clerical work: reading and first-pass judgment

Slide 19 shows where AI has moved into existing clerical work. Its stated method: “combine in-house AI-OCR and tools to make existing work more efficient,” and because “end-to-end process transformation takes time, start with the visible scope.” It names four projects:

ProjectWhat changedHours saved
Customer attribute determinationDetermining the attributes of corporations and organizations: visual check → AI reading and first-pass judgment3,500 per year
Garnishment noticesExtracting the data the downstream flow needs: visual check → AI reading and data hand-off2,800 per year
Risk management judgmentAI judges where risks lie in each department of the clerical division and offers suggestions2,000 per year
Outsourcing contract checksA custom template that can check outsourcing contracts, created and rolled out1,200 per year

Two of the four state the before-and-after directly: staff used to check documents by eye, and now AI reads them. In attribute determination, AI also makes the first-pass judgment. In garnishment notices, AI passes the extracted data on to the next step. The deck does not say how the later steps are handled.

Three further slides on applied AI were not disclosed: a project to turn tacit knowledge into explicit knowledge (slide 18), building AI agents (slide 20), and a target image for structured clerical procedures (slide 21).

Confirmed facts: Process Design Group created from the clerical group; Process Transformation Department created as a cross-company unit; remit redefined as designing processes end to end; FY2025 shift of resources to consolidation, review, digitization; AI tool at top of internal portal with seven models; culture metrics (95%+ certified; FY2025: 202 workshops/34 projects, 2,500 participants/42 sessions; 5 hands-on sessions in 3 regions); 10,000+ member internal AI community; tools built in-house with many PoCs; four AI projects with 3,500, 2,800, 2,000, and 1,200 hours per year saved; slides 8, 14, 15, 18, 20, 21 not disclosed.


4. Impact

The deck reports results in hours of work removed per year, for each of the four projects on slide 19:

  • Customer attribute determination: 3,500 hours/year
  • Garnishment notice reading: 2,800 hours/year
  • Risk management judgment: 2,000 hours/year
  • Outsourcing contract checks: 1,200 hours/year

The slide does not add them up. Together they come to 9,500 hours a year. The deck gives no start date, measurement method, or before-and-after baseline for these figures.

The other numbers in the deck measure adoption, not output: more than 95% of Process Design Group members certified at MPS Stage 1, 202 workshops yielding 34 projects, 2,500 participants across 42 knowledge-sharing sessions in FY2025, and an internal AI community of more than 10,000 members.

The deck does not give a figure for the overall goal or progress toward it. The 5,000 figure appears only in its corrections of the press reports, as “5,000 people’s worth of clerical work” across the group. The ten-year horizon and the number of clerical staff come from the February reports. Bloomberg reported that Mizuho has “about 15,000 clerical positions at its core banking units and other group companies” and that “affected employees will be transferred to other roles, the bank said.” Kyodo’s own report described the plan as replacing up to about 5,000 people’s worth of clerical work with AI. It said the target work was checking documents customers submit in procedures such as account opening and entering information into systems, and that surplus staff would move to retail sales at branches and corporate sales support, with reskilling support.

On what the freed time is for, the deck’s own answer is a priority, not a number: “‘how to use the time we create’ is the most important issue.”


5. Insight

The unit of the plan is work, not people. Mizuho’s correction list changes one thing in every line: the object of the cut moves from staff to clerical work. That framing is consistent with the slide 8 premise. If headcount will fall naturally over ten years, the problem the bank describes is too much work for the people it will have, not too many people.

In Mizuho’s account, the organization comes before the AI. The first thing the deck says Mizuho changed is the clerical group’s job. It became a Process Design Group with an end-to-end, company-wide remit, beside a new Process Transformation Department, and it was reframed from “static” virtues such as accuracy and carefulness to “dynamic” ones such as improvement and standardization. The AI tools come after that in the deck, and its summary line states the order as a principle: foundation, data, process transformation, then AI. The deck does not date the reorganization against the rollout of the AI tools.

The first AI work is reading and first judgment. All four measured projects sit at the entry of a clerical flow. AI reads a document someone used to read by eye, makes a first-pass call, or checks a contract against a template. The deck says why it started there: end-to-end change takes time, so begin with what is visible. It presents these projects as that starting scope, not as the program’s total.

Broad adoption is treated as a separate workstream. The general chat tool, the use-case menu, workshops, certification, and a 10,000-member community are counted on their own slides with participation numbers. They are presented as the foundation that makes the process projects possible, not as the source of the hours saved.

Limits. This is the company’s own presentation, and the hours saved are not independently verified. The deck withholds the slides on the staffing gap, personnel measures, what MPS is, the tacit-knowledge project, the AI agent build, and the to-be structure of clerical procedures. It gives no headcount change, no redeployment numbers, and no timeline for the 5,000 people’s worth of work. The ten-year horizon, the 15,000 clerical staff, and the moves to sales come from press reports. No minutes of the hearing were found.

One-line takeaway: Mizuho’s answer to a shrinking workforce was to turn its clerical group into the unit that designs processes across the company, and to aim AI first at reading and first-pass checks, while saying openly that the hard question is what people do with the time that frees up.


6. Implementation hypotheses (not in the source)

6.1 Attribute determination as AI first pass with a staff confirmation queue

Basis: Slide 19: customer attribute determination for corporations and organizations moved from “visual check” to “AI reading and first-pass judgment,” using in-house AI-OCR and tools, saving 3,500 hours a year.

The word “first-pass” implies a second pass. The likely flow is that registration documents or forms for a corporate or organizational customer are scanned, the in-house AI-OCR extracts fields such as entity name, type, and registration details, and a model or rule set proposes an attribute category. A staff member would then confirm or correct the proposal instead of reading the whole document. Cases where the OCR confidence is low or the document does not fit a known type would probably go straight to the old visual check. Under that design, most of the 3,500 hours would come from reading and transcription, while the final decision stays with staff. The confirmed category would then be recorded in the customer system, so the AI’s proposal and the staff decision can be compared over time.

6.2 Garnishment notices as structured data handed to the next step

Basis: Slide 19: for garnishment notices, AI extracts “the data the downstream flow needs,” replacing “visual check” with “AI reading and data hand-off,” saving 2,800 hours a year.

A garnishment notice tells the bank to hold a debtor’s funds, and the bank has to find the right customer and accounts quickly. The phrase “data hand-off” suggests the AI output is not a summary for a person to read but a set of fields, such as debtor name, address, issuing authority, case number, claimed amount, and date of service, passed into the next step of the flow. The next step most likely matches those fields against customer records to find the accounts. Because holding the wrong account is costly, a staff member would likely confirm the match before funds are held, with the AI removing the transcription work in front of that check.

6.3 Contract checks as a shared prompt template, not a separate system

Basis: Slide 19: the outsourcing contract project “created and rolled out a custom template that can check outsourcing contracts,” saving 1,200 hours a year. Slide 12’s use-case menu lists contract-check entries marked “New!” for outsourcing of system development and operation (2025) and outsourcing of clerical work (2025). Slide 16 shows prompts shared company-wide as custom templates for the chat tool.

The contract check was probably delivered inside the general chat tool rather than as a new application. A custom template would hold a fixed checklist of the clauses an outsourcing contract must contain, together with output instructions. A staff member would open the template from the menu, attach the contract, and get back a list of items found, missing, or needing attention. Separate templates for system and clerical outsourcing suggest the checklist differs by contract type. The review decision would stay with the contract owner. In this setup the 1,200 hours saved would come from the first read-through, and maintaining the checklist becomes the main ongoing work, done by whoever owns the template.


7. Sources

  1. Mizuho Financial Group — Mizuho no Purosesu Henkaku ni tsuite (On Mizuho’s process transformation), 24 slides in Japanese, quoted here in translation; submitted as Document 1 to the 41st meeting of the Basic Labour Policy Subcommittee, Labour Policy Council
    https://www.mhlw.go.jp/content/12602000/001754475.pdf
    2026-10-02 (accessed 2026-11-17)
  2. Ministry of Health, Labour and Welfare — 41st Basic Labour Policy Subcommittee meeting page and agenda (hearing with Mizuho FG, East Japan Railway, and Hashimoto-gumi)
    https://www.mhlw.go.jp/stf/newpage_76504.html
    https://www.mhlw.go.jp/content/12602000/001755226.pdf
    2026-10-02 (accessed 2026-11-17)
  3. Bloomberg via The Star — Mizuho plans to replace 5,000 clerical jobs with AI in 10 years (Taiga Uranaka, Hideki Suzuki)
    https://www.thestar.com.my/tech/tech-news/2026/03/02/mizuho-plans-to-replace-5000-clerical-jobs-with-ai-in-10-years
    2026-03-02 (Bloomberg original 2026-02-27; accessed 2026-11-17)
  4. Kyodo News via OANDA Lab market news — Mizuho: 5,000 people’s worth of clerical work to AI (Japanese)
    https://www.oanda.jp/lab-education/market_news/kn_2026022701002015/
    2026-02-27 (accessed 2026-11-17)