Our editor wants to open a seaside café. An AI wired into Japan's government data skipped the busiest station — and worked out the rent he can afford

Kaito Ogasawara, DATA WORLD's editor-in-chief, has a quiet dream.

Someday, he wants to open a small café near the sea.

Twenty seats or so. On weekdays, neighbors drop in with their laptops. On weekends, families stop by on the way back from a walk on the beach.

"If I do it, it'll be in Shonan. But which station?"

Shonan is the stretch of coast southwest of Tokyo, about an hour away by train. He had narrowed it down to three neighboring stops on the JR Tokaido Line: Fujisawa, Tsujido and Chigasaki. Ask around the office and you get three answers. "Fujisawa, obviously — that's where the people are." "Chigasaki. Rent's cheaper."

To do this for real, he'd need a loan. Japan Finance Corporation, the government-owned lender that funds many first-time business owners, publishes a startup guide for restaurants. Its checklist asks one blunt question:

Gut feeling and office opinions don't count. You need numbers: how many people use the station, how old the neighbors are and how that will change, how busy it gets on weekends, what land costs, and whether a café is even allowed on the lot. Chain-store site planners layer this kind of data on mapping software to define a "trade area" — the zone your customers come from. For an individual, just collecting the numbers from scattered government websites and files has been enough to stall the plan.

Now an AI can do the collecting. In November 2025 and February 2026, Japan's Ministry of Land, Infrastructure, Transport and Tourism (MLIT) opened gateways that let AI talk directly to its data. You ask a question; the AI goes and fetches the government data and writes up an answer.

So Ogasawara decided to ask an AI connected to those gateways, in plain everyday language. His first message, in Japanese, was just this:

I want to open a café in Shonan. Which station area would be best — Fujisawa, Tsujido or Chigasaki? Compare them with MLIT data and put it in a report with a map.

Just over three minutes later, the AI had a comparison report with a map. On the second message, it reversed its own conclusion. On the third, it built a screen where you can change the numbers and watch sales and affordable rent move. How close did Ogasawara's café dream get to a plan with a real "basis for projected sales"? Here are the answers up front.

The first thing Ogasawara wanted to settle was the station. Within about ten seconds of his message, the AI had looked up which tools the MLIT gateways offered and announced: "First, land prices and passenger numbers for the three stations."

Three minutes and 21 seconds later, the answer came back: "For an independent café, Chigasaki is the top pick."

Its reasoning: Chigasaki has the cheapest land near the station of the three, and the population right by the station holds roughly flat through 2050. Fujisawa has about 3.5 times as many passengers as the other two, but prime land in front of the station runs ¥1.91 million per square meter.

Reading it again, though, Ogasawara noticed a gap. The population figure covered only a single 250-meter grid square sitting on top of each station. Traffic and nearby facilities weren't in the picture at all. So he did what anyone would do and sent a follow-up:

Can you add traffic, the number of nearby daycares, hospitals and care facilities, and future population by age group, and compare again? Update the report too.

The AI pulled population by age group for the nine grid squares around each station (about 750 meters across), counted facilities and added road traffic counts. About ten minutes later, its answer opened like this:

The deciding number was the 20-to-39 age group, which the AI treated as the core café crowd. Tsujido is the only one of the three where that group doesn't shrink by 2050. It is also the only station whose passenger numbers are growing, up 21% since 2011.

But does "the young adults stay" actually translate into customers walking through the door? That was the next question.

A café's trade area is basically walking distance. MLIT's handbook for evaluating cities treats a radius of 800 meters as the standard walking range, and Japan's real estate advertising rules count 80 meters as one minute on foot. So 800 meters is a 10-minute walk.

We re-ran the same population projections the AI used, this time for an 800-meter radius around each station.

Change the boundary, and the result holds: only Tsujido keeps most of its young adults.

Why that matters for a café shows up in Japan's Family Income and Expenditure Survey, run by the government's Statistics Bureau. In 2025, households headed by someone in their 30s spent ¥17,896 at cafés over the year. Households headed by someone 70 or older spent ¥7,268, a gap of about 2.5 times (national average, households of two or more). A neighborhood that keeps its young adults keeps its café spending.

The other thing to check is who passes through. MLIT publishes nationwide people-flow open data: estimates of how many people are present in each 1-kilometer grid square, built from smartphone location data. It's broken down by weekday versus weekend and midday versus late night, as monthly averages.

Run the three stations through it, and their personalities split clearly.

Fujisawa draws about the same midday crowd on weekdays and weekends — a commuter town of offices and schools. Tsujido is a weekend town: on weekend middays it holds 1.32 times as many people as on weekdays. A large shopping mall, Terrace Mall Shonan, opened in November 2011, connected directly to the station's north exit. The station's passenger growth since 2011 tracks that redevelopment.

One small surprise: there's barely any seasonal swing. Around Tsujido, the weekend midday crowd was about 48,000 in August and about 50,000 in January. It's a beach town, but the station area isn't a summer-only place.

Neighbors on weekdays, shoppers and walkers on weekends. The two kinds of customers Ogasawara had pictured line up neatly with how people actually move around Tsujido.

With a station chosen, it was time for the part the loan application cares about: the basis for projected sales. Ogasawara sent a third message in the same conversation.

In English, his request read roughly: "Say I do it in Tsujido. With about 20 seats, how much could I expect to sell? And how much rent could I afford? Add a screen where I can change the numbers — and make it match our site's stylesheet so it's easy to read."

The answer: in the base case, about ¥2.21 million a month in sales. Rent up to ¥220,000 a month is safe, and the ceiling is about ¥270,000. But the AI suggested starting the search with units at ¥150,000 to ¥200,000, a range that stays out of the red even in its cautious case.

It also added an interactive calculator to the report.

Sales are seats × table turnover (customers per seat per day) × average spend, plus takeout. Subtract the cost of goods, labor, utilities and fit-out loan repayments, and what's left is split between rent and the owner's income. The Japan Finance Corporation startup guide uses the same "average spend × seats × turnover" formula for projected sales, so this drops straight into a loan application. The most revealing part was the lookup table.

Half a turn per seat on weekdays moves the affordable rent by more than ¥70,000 a month. That's why the AI kept stressing: "Turnover matters most. Stand in front of the unit you're considering, count the people walking by, and test the turnover assumption."

Still, turnover, average spend and cost of goods are assumptions the AI chose as "typical for an independent café." The AI said so itself: "These are not actual figures." So we checked its numbers against official statistics from two directions.

First, from real cafés. Japan Finance Corporation's small business benchmarks for fiscal 2025, which surveyed 35 cafés it had lent to, put median annual sales per seat at ¥1,298,000. For 20 seats, that's about ¥2.16 million a month. The AI's base case of ¥2.209 million lands right at the median café. Its 30% cost of goods is close to the survey's median of about 28%, too.

Second, from local wallets. According to the household survey, households of two or more in the Kanto region (Tokyo and its surrounding prefectures) spent ¥14,296 at cafés in 2025. Per person, that's about ¥5,000. Across the roughly 24,000 people within a 10-minute walk of Tsujido Station, it adds up to about ¥120 million a year.

That money gets split among every café nearby. With, say, 10 cafés within walking distance, residents alone would bring each one about ¥920,000 a month. To reach ¥2.21 million, the café needs the station's 111,000 daily passengers and the weekend shoppers. The numbers confirm it: being close to the station, and to that weekend crowd, really does matter.

The same survey carries a sobering figure. Of the 35 cafés, only 20 were both profitable and carrying positive equity. Their average break-even ratio was 104.6%, meaning the typical café runs right on the line between profit and loss. The startup guide itself advises projecting "sales on the low side, costs on the high side." The AI's advice to start with ¥150,000-to-¥200,000 units doesn't look overly cautious.

Once the budget is clear, it's time for a property. Imagine a vacant unit near Tsujido Station's north exit that caught his eye. In a separate conversation, the request went like this:

I want to run a café in a vacant shop around here, by Tsujido Station's north exit (35.3422, 139.4470). Before I sign anything, check the zoning, disaster risks and land prices with MLIT data.

The answer came back in 1 minute 6 seconds: the lot is zoned industrial, but a café is allowed. A district plan applies, though, so check its details with Fujisawa City before signing. An industrial zone right by a busy station was a bit of a surprise.

The lot sits outside every flood, storm surge, tsunami and landslide zone. The nearest prefectural land price survey point was ¥820,000 per square meter, up 5.1% from the year before. From that, the AI suggested checking the rent-review clause in the lease, since the landlord may ask for an increase at renewal.

The link the AI provided opens MLIT's official map, where you can confirm the zoning and district plan yourself.

The AI also listed, from general knowledge, the paperwork that follows a decision to lease. The main items:

• A restaurant business license from the public health center
  Cafés need a license under the Food Sanitation Act. In Fujisawa, where Tsujido Station is, the city's own public health center handles it.
• A designated food hygiene supervisor
  Each shop must name one. Most people qualify by taking a short course.
• Notifying the fire department
  Before using the space as a shop, you file a notice that you're starting to use the building.
• A building permit review for change of use over 200 m²
  Converting a unit larger than 200 square meters into a restaurant requires a building confirmation application. A 2019 amendment raised the threshold from 100 to 200 square meters.

Here's how far government data and official statistics carried the plan.

The left column used to mean hopping between websites, files and mapping software. In three rounds of conversation, it came together as one comparison report and a working calculator. The right column is the same work it always was: asking real estate agents, counting people in front of the unit, confirming at city hall.

For Ogasawara's loan application, the groundwork for "why Tsujido" and "the basis for projected sales" is now in place, thanks to government data and AI. What's left is a weekend job: standing in front of a candidate unit and counting passersby on a weekday and on a weekend.

We also found three things to watch when you hand this kind of work to an AI.

• Follow-up questions make the answer deeper
  With a one-line request, the AI answered quickly using four kinds of data. When asked to add more, it combined eight kinds and changed its conclusion.
• It can misread an empty result
  The first answer said "the tsunami inundation API returned no results for any of the three stations." In fact, all three stations sit outside the projected inundation zone, so there was nothing to return. In the site check, too, it reported a fire prevention zone as "not designated" when the data had simply come back empty.
• Check displayed numbers against the source
  Every number in its tables matched the source data, but the chart headings in its report mixed up a ratio with a percentage change.

You need four things to get started.

• Two API keys
  Apply for the Real Estate Information Library key on its API application page. For the MLIT Data Platform, registering an account on its site unlocks the API. Both are free.
• Python 3.10 or later
  Both gateways are published as programs you run on your own computer.
• An AI app that supports MCP
  MLIT's setup guide uses Claude Desktop as its example. You add each gateway's location and your API keys to a settings file.
• The gateway programs
  Download the geospatial gateway and the data platform gateway from GitHub.

Plain, everyday language works fine. Ask for a quick comparison in one line, add what you care about in a second message, then ask about sales and rent in a third. That order turned out to be the order of a store plan itself. For locations, give a station name or an address — or latitude and longitude, if you have them — so the AI doesn't guess the wrong spot. Note that the gateways and the data cover Japan only.

EDITOR'S TAKE ── How Editor-in-Chief Ogasawara reads it

Government data now fits the playbook for opening a shop

Once the plan for a 20-seat café in Tsujido was done, two things stayed with me.

On passengers alone, Fujisawa wins. On cheap land alone, Chigasaki wins. Tsujido came out on top only because two numbers you'd rarely think to look at — the future count of people aged 20 to 39, and the weekend crowd — were layered in.

That kind of comparison used to belong to companies with site-planning teams and budgets to match. Choose the town, know the customers, estimate sales and rent, check the property: nearly every step of the store-planning playbook can now draw on free government data. And the AI takes on almost all of the work of gathering and layering it.

The third answer's calculator was very convincing on screen. But turnover and average spend were the AI's own assumptions. Only after we held them up against Japan Finance Corporation's café figures and the household survey did we know it was "a median café" estimate.

Ask the AI "What else should I look at?" and "Add that too." Then check its numbers once against official statistics or the source data. Make those two things a habit, and an AI becomes a reliable advisor for big decisions like opening a shop.