Narrow a shortlist of Japanese government suppliers with AI

What this procedure is for

Market researchers often need to know which companies have delivered to a particular Japanese ministry, whether to find a partner or to see what competitors have won. Japan's central government publishes its award results on the Procurement Portal (調達ポータル), but the volume is large, and the rows only become a list of candidates once someone filters them by ministry and subject and regroups them by company.

Hashike is a bilingual research service that organizes US federal contract opportunities performed in Japan and Japanese central-government award results. It ties each Japanese award result to its supplier by corporate number (法人番号), the thirteen-digit number Japan's National Tax Agency assigns to each corporation. The number identifies one company regardless of how its name is spelled or whether another firm shares the name.

The steps below hand Hashike's records to an AI assistant and have it build a candidate list keyed by corporate number. At the end you have a table for one ministry listing the suppliers, each one's award count and value, and the contract methods used, and every number in it traces back to a value Hashike returned.

How many award results the procedure works over

Hashike counts central-government award results published on Japan's Procurement Portal (調達ポータル) over the most recent sixty months. In that window there are 167,444 award results, 17,515 distinct suppliers counted by corporate number, and 45 ministries with at least one award result.

Over the same sixty-month window, the ministry with the largest total award value is 国土交通省 (the Ministry of Land, Infrastructure, Transport and Tourism), with 41,311 award results and a total award value of ¥1,105,189,106,402. A single ministry can carry that many rows, which is why the procedure filters by procurement name and fiscal year before grouping.

All amounts are award prices (落札価格, the winning bid) in yen as published. They are not converted to dollars and cannot be added to or compared with US federal contract values.

The tools and how to connect

MCP (the Model Context Protocol) is a standard way for AI assistants to call outside tools. Hashike's MCP server is a remote HTTP endpoint at https://hashike.io/api/mcp and needs no package or local process. It authenticates with a Hashike API key (prefix hsk_live_) or a Supabase OAuth access token, passed in an Authorization: Bearer <token> header.

This procedure uses three tools. Each has a public API endpoint that returns the same content over GET, so a script can call them without an AI assistant.

• ·buyer_profile returns one ministry's award count, total value in yen, number of distinct suppliers, contract-method mix and recent awards. On the public API, the buyers endpoint takes the ministry code in its path.
• ·find_comparable_awards searches award results by ministry, Japanese fiscal year and procurement name. It requires at least one filter and never returns an unfiltered listing. On the public API, the awards endpoint does the same search.
• ·supplier_profile returns, for one corporate number, the exact award count and value over a rolling sixty-month window, the ministries that bought from the company, its contract methods and recent awards. On the public API, the companies endpoint takes the corporate number in its path.

Hashike does not offer a CSV or Excel export. Records come out through MCP, the public API or the Markdown export.

Building the candidate list

These steps assume an AI assistant connected to Hashike's MCP server. Steps one to three work equally well as requests to the assistant or as direct calls to the public API.

1. Open the ministry on Hashike's buyer index at /buyers. The two characters at the end of the page address are its ministry code (府省コード), the code published with the award results, and the later calls take it as ministryCode.
2. Call buyer_profile with ministryCode=<ministry code> and note the ministry's award count, total value and number of distinct suppliers. Those three numbers are the baseline the candidate list is checked against later.
3. Call find_comparable_awards with ministryCode=<ministry code> and a word from the procurement name as query=<keyword>, adding fiscalYear=<fiscal year> if needed. The result holds the total number of matches (totalMatching) and a bounded page of award results, newest first, each with the procurement name, award date, award price, contract method, company name and corporate number.
4. Have the assistant group the returned rows by corporate number. The prompt below is one way to ask. Rows without a corporate number cannot be tied to a supplier, so the prompt keeps them in a separate group.
5. Call supplier_profile with corporateNumber=<corporate number> for each candidate. It returns the exact sixty-month award count and value for each company and the full list of ministries that bought from it.
6. Finish with the checks in the next section. The table that passes them is the candidate list.

Use the Hashike find_comparable_awards tool with ministryCode <ministry code> and query <keyword>. Group the returned award results by corporate number. For each corporate number, list the company name as published, the number of rows, the sum of the award prices in yen, and the contract methods that appear. Put rows with no corporate number in a separate group and do not assign them to any company. Report totalMatching and the number of rows you actually received. Do not convert yen to any other currency, do not estimate values for rows you did not receive, and do not rank or recommend companies. Then call supplier_profile for each corporate number and add its exact sixty-month award count and value as separate columns, labelled as coming from supplier_profile.

Checking the assistant's answer against Hashike

A language model can miscount rows, merge companies with similar names or produce a plausible total that no tool returned. Before using the table, compare the model's output with the values Hashike returned on the following points.

• ·Row count: the rows per corporate number plus the rows without one add up to the number of rows step three actually returned.
• ·Corporate numbers: every number in the table matches the tool output digit for digit, and no rows were merged on company name alone.
• ·Ceiling: for every company, the supplier_profile count and value are at least as large as the figures counted from the step-three rows dated inside the sixty-month window. If one is smaller, the model assigned rows to the wrong company.
• ·Ministry total: the candidates' combined count and value, counted from the step-three rows, do not exceed the ministry-wide count and total value from buyer_profile. The supplier_profile figures include awards from other ministries, so they do not belong in this comparison.
• ·Currency and period: amounts stay in yen, and the period is described as a rolling sixty-month window, not as a fiscal year.

What the list does not cover

The candidate list inherits the scope of the award results behind it, including the limits below.

• ·It is built entirely on public records. Hashike has no access to non-public government information and is not affiliated with or endorsed by any government.
• ·It covers central-government award results only. Prefectural and municipal procurement is not included.
• ·About one point two percent of award rows carry no corporate number (法人番号) and cannot be tied to a supplier. A row missing from a supplier's record means it could not be matched, not that the supplier was inactive.
• ·Multiple-award methods (複数落札) let several suppliers win one procurement, so a row count is a count of award results, not of procurements.
• ·The absence of a record under a set of filters is not evidence that the underlying activity did not occur.
• ·The window is a rolling one over award date, not a fiscal year. Japan's fiscal year runs April to March and the US federal year October to September, so a figure spanning both would silently reconcile two calendars. The fiscalYear filter on find_comparable_awards is the Japanese fiscal year.