A worked-through example — when a Custom AI App is the right choice and how the building blocks fit together
The overview describes when a Custom AI App is a better fit than an assistant or a workflow. This page works through one concrete example so you can transfer the pattern to your own cases.
A purchasing team regularly sends queries about rescheduled delivery dates to many suppliers and has to keep track of the replies. Done by hand, that means: maintaining spreadsheets, writing emails one by one, reading replies, and updating status manually.
As a Custom AI App, this turns into one continuous flow:
An export file from the ERP system (changed delivery dates) is uploaded.
The app automatically writes the matching follow-up email per supplier.
Incoming replies are read by the AI and translated into a clear status (date confirmed, new date proposed, question raised).
The team sees at any time which queries are open, confirmed, or answered.
The overview names three signals — all three apply here:
Structured inputs — the ERP export is a table, not free chat input.
Results to review and confirm — buyers want to see and approve every supplier reply, not adopt it blindly.
A multi-step process with state — a query runs through several steps over days; that state has to persist.
On top of that, two properties that a chat assistant or workflow doesn't capture well:
Its own data store — the app keeps its own database of queries, suppliers, and status.
Background jobs — emails run on a fixed schedule and a daily digest is sent automatically, even when nobody has the app open.
Note
Rule of thumb: as soon as a task needs its own state over time and a guided interface, a Custom App is usually the better choice. For one-off question-and-answer tasks, the chat assistant stays simpler.
The app is a standalone web application with its own frontend and backend, embedded into meinGPT via an iframe. meinGPT handles sign-in and securely forwards the user context to the app.
Building block
Role
Interface (iframe)
Table view of queries, detail dialogs, approval of replies
Backend
Own database, import logic, email sending, AI calls
Auth proxy
meinGPT as the identity source; the backend verifies every request against the meinGPT JWT
Email integration
Sending and receiving via a mail API (e.g. Microsoft Graph), credentials stay server-side
AI evaluation
Incoming replies are passed server-side to a language model and translated into a status
The security-critical part — how your app only accepts requests coming from meinGPT and verifies the user context — is described separately under Security & Authentication Proxy.
Import — The buyer uploads the ERP export. The app creates a query per row and matches it to the right supplier.
Outbound — For each open query, the app generates a follow-up email and sends it via the configured mail API. Sending runs on a fixed schedule in the background.
AI evaluation — When a reply arrives, the language model reads it and proposes a status: date confirmed, new date, or question needed.
Review & approve — The buyer sees the reply and the AI's proposal side by side and confirms or corrects the status. Nothing is marked final without human approval.
Overview — A daily summary goes out automatically; the overview shows at any time what is open, confirmed, or answered.