Studio S.O

One AI agent runs a real one-person business — consulting, travel, and culinary — end to end: hearing, proposal, Stripe payment, delivery, 14-day AI concierge.

Comment

Studio S.O is a working business in Japan — AI consulting, travel design, and culinary development — operated by one person together with a Gemini agent running on Google Cloud.

The agent owns the customer relationship end to end. It holds the consultation in the customer's own words, writes up the requirements for approval, produces the free draft and then the paid deliverable, stays on as a fourteen-day concierge after payment, and sends a follow-up days later. For the travel and culinary lines, no human is ever in the loop — not before the sale, not during delivery, not after.

The storefront is live and takes real payments through Stripe, including PayPay. Every request is metered: the agent logs its own decisions and its own cost, and is held to a hard ceiling per request.

My job is not to serve customers. My job is to correct the agent when it is wrong.

Business Viability

Studio S.O sells eleven products across three lines, priced from ¥500 to ¥500,000, at sutekioojisan-so.com. Payments run on Stripe in live mode, with both credit cards and PayPay enabled. The first paid order came through on 10 August 2026 and was delivered end to end without a human touching it.

Revenue is early. The operation is not. What is already in place is the part that usually takes a small company years to assemble: a public storefront, statutory disclosures under Japan's Specified Commercial Transactions Act, a privacy policy, live payment processing, automated delivery, post-sale support, and a follow-up loop — all of it running without staff.

The unit economics are measured rather than estimated. Every request writes a cost line naming the model, the token counts, and the inferred cost, so the margin on each sale is a number I can read, not a projection. Marginal cost per delivery is a few yen of inference. That is why a ¥500 product and a ¥500,000 product can sit in the same catalogue without a different cost structure behind either one.

The risk in a one-person business is the person. That risk is what the agent removes: the work that used to stop when I stopped now continues without me.

AI-Native Operations

The agent is not a chat widget bolted onto a business. It is the business's operating layer.

It runs the consultation through a speaking avatar, in Japanese or English at the same quality. It infers which of the three lines a request belongs to and narrows the catalogue accordingly. It writes the requirements sheet the customer approves on screen. It generates a free draft, and then — the moment payment clears — the full deliverable, rendered to PDF, mailed to the customer, and archived for thirty days.

For travel and culinary work it grounds its output with Google Search and verifies the grounded links before they ship, so the places, routes and prices it names are real ones. After payment it continues as a concierge for fourteen days, showing the customer the days and turns remaining. Days later it drafts and sends the follow-up itself, with a one-click satisfaction answer written straight back to the order record.

It also watches itself. A regression harness scores each product's output with an LLM judge against the consultation it was given, so quality drift is caught by a machine rather than by a complaint. Every Gemini call passes through a meter with hard call and token ceilings, so a malfunction costs yen rather than a month of revenue.

The human in this business writes no customer-facing sentence. The human reads the logs and fixes the agent.

Category Impact

The customers are individuals and very small businesses. A builder losing his evenings to sorting site photos. A salon owner losing bookings to a phone that rings mid-treatment. A family trying to plan three days away. People who are not helped by being handed a form and told to fill it in.

So the agent is built to listen first and ask second, never to use a technical word, and to reflect the customer's own phrasing back before it proposes anything. The requirements sheet exists for exactly one reason: so the customer can see they were heard correctly, in their own words, before any money is discussed.

The same flow runs in English at identical quality, which means a business this small can answer an enquiry from overseas on the day it arrives.

What this demonstrates for the category is narrow, and I think useful. A single person with real domain expertise and no staff can now offer the responsiveness and the finish that used to require a firm. The expertise was never the bottleneck. Everything around the expertise was the bottleneck — and that is the part that an agent can now carry.

Google Cloud products in production use

Vertex AI (Gemini) — the consultation, the requirements sheet, and every deliverable. Google Search grounding is enabled for the travel and culinary lines, and grounded links are verified before delivery. Gemini image generation illustrates the deliverables. Cloud Run — the backend service (FastAPI), where the agent runs. Firebase Hosting — the storefront, in Japanese and English. Cloud Scheduler — the daily follow-up batch. Cloud Logging — the agent's own decision log and a per-request cost line. Cloud Build — contgeminiainer build and deploy. Secret Manager — Stripe keys and the webhook secret. Google Sheets API — the order record: consultation, payment state, concierge history, follow-up state. Gmail API — delivery of the finished plan as a PDF, and the follow-up mail. Google Drive API — a thirty-day archive of every deliverable.

The entry requirement is one Google Cloud product. This business runs on eleven.

Built With

Share this project:

Updates