Collecting orders from each customer's Web-EDI in one list

Every customer sends orders its own way. Customer A has its own ordering system, B a Web-EDI, C a supplier portal. Each morning someone signs in to each one, checks the orders, downloads a CSV, puts the columns in order, and enters them into the sales system. The more customers, the longer the morning.

This column explains how to automate collecting orders from each customer’s site into one list, with a run we recorded.

Key points

• If a customer offers a data feed, such as an API or file delivery, use it before automating the screens
• When you automate the screens, the work is in the sign-ins. A user ID and password run unattended; with a one-time code, a person types only the code
• On three sites, the first run learned the steps with AI, and the second replayed them without AI

First, check for a way that skips the screens

Some customers can send order data as files. In Japanese retail, the standard EDI is 流通BMS (Ryutsu BMS); by its council’s estimate, more than 21,600 wholesalers and manufacturers had adopted it as of June 1, 2025. Other customers offer an API, or deliver CSV files over SFTP or email.

Where those exist, nothing breaks when a screen changes, and large volumes stay steady. Automating the screens is for customers with no data feed, whose sites you can only sign in to and read.

What differs from site to site

Customers’ sites agree on neither look nor steps:

• Sign-in: what the ID is called (login ID, user ID, email), the button’s wording, and whether there is an extra check
• CSV format: column names and order, date formats, and units (some send cases and units per case)
• Terms of use: whether automated use is allowed (check the terms before automating)

How each kind of sign-in is handled

Kind of sign-in	Can it run automatically?	How Kitewell handles it	What you decide
User ID and password	Yes	The password is kept on the PC as a secret and typed by the browser; AI sees only its name	Whether the customer will issue an account for the automation
One-time code (email, SMS, app)	A person types only the code	The run stops and asks for the code; once someone enters it in the run’s page, the same browser carries on	Who receives the code, and when
Client certificate	On a PC that holds the certificate	Add the certificate for the site in Device settings	Which PC holds it, and who manages it
IP address restriction	On a PC on an allowed line	It runs on your own PC, so it uses your office line as is	Whether to ask the customer to add you to the allow list
Image challenges and other bot checks	Don’t automate	If the site refuses the browser, the run stops and records the refusal	Ask the customer about a data feed

A one-time code is recorded with the run, so this is only for codes that expire quickly. Never type a password into a run’s page.

A recorded run

On October 19, 2026, we ran it against three stand-in sites built for this column. The setting is a food wholesaler that takes orders from a supermarket, a food distributor, and a home center. All three sign in with a user ID and password, and one also asks for a one-time code. The sites and Kitewell ran on the same Mac, and the AI model was Claude Sonnet 5.5.

The workflow has four steps: three browser steps that each download one customer’s CSV, and one that merges them into a single order list.

On the site that asks for a one-time code, the run stops and asks for it. Once someone enters the code, the browser continues on the page it left open.

	First run (AI learns the steps)	Second run (replays them)
Supermarket	15 s	8 s
Food distributor (one-time code)	14 s	11 s
Home center	14 s	8 s
Operations that ran without AI	On each site, only opening the page, the check, and the download	Every operation on each site (7 / 7, 9 / 9, 7 / 7)
AI model usage	About 33,000 tokens across the three sites	None

The three sites were collected in parallel, and the times exclude waiting for a person to type the code. In the first run, AI read each page to decide where to type and what to click, and the steps recorded those operations. The second run replayed them as recorded, without asking AI. When a screen changes and a recorded operation can no longer be found, AI looks for that one operation again (why RPA robots stop when screens change).

Putting the CSVs in one format

The three CSVs differed in column names and in how they gave quantities. The merge step maps them to shared columns:

Shared column	Supermarket	Food distributor	Home center
Order number	発注No	order_id	伝票番号
Delivery date	納品予定日	delivery_date	納品日
Ship to	納品先	ship_to	店名
Item code	JANコード	item_code	商品CD
Quantity	発注数	qty	cases × units per case
Unit price	原単価	unit_price	単価

The result was one order list (CSV) of 9 rows from 3 customers. This run stops at the list; it does not enter it into a sales system. For that, use the sales system’s CSV import, or add a step that works its import screen.

When this does not fit

• The customer offers a data feed: 流通BMS, an API, or file delivery is steadier
• Every sign-in needs a one-time code, and you want it to run when no one is around: someone must type the code, so it is never unattended
• The site turns automation away with image challenges or similar: don’t automate it
• Very large daily volumes: a data feed suits that better than the screens

Summary

• For each customer’s ordering site, first check for a data feed, and automate the screens only where there is none
• The kind of sign-in decides how it runs: a user ID and password run unattended; with a one-time code, a person types only the code
• Steps AI learns on the first run are replayed without AI from the second

Running this collection at a set time with no one present is recorded on the same sites in scheduling Power Automate Desktop.

In a PoC, we start with one of your customers’ sites and test on your real screens and data (early access partners). For the related use case, see take in orders from retailers’ Web-EDI.

Sources

• LOGISTICS TODAY: 流通BMS adoption spreads to smaller companies, passing 21,600 (July 8, 2025; the council’s estimate; in Japanese)