Document routing for public administration. One pipeline runs from the first connector onward, is configured while in operation, and collects everything it does not yet recognize until someone assigns it.
Apache 2.0 · Self-hosted · Model choice per task
AssignedBusiness registration is now a type — with its rule.
INTAKE groups unknown documents into stacks on its own. One assignment creates a type and its rule, and the whole stack is processed. Documents and values are illustrative.
* The pipeline is the product
INTAKE opens on the stage, and the stage is the pipeline. Everything else hangs off it: the contents of a pool, a single document, the assignment of a stack, the list of issues.
UNKNOWNdocuments waiting for an assignment
ISSUESwhat fails, what it means — and one action
01 — The problem
In a public administration, documents arrive through many channels: the mailroom's inbox, the SharePoint folders of individual departments, the scanner, the citizen portal. Someone has to look at every document, classify it and forward it — and for invoices, applications and objections, keep an eye on deadlines too.
The usual tools are inbox rules, which only react to sender and subject, or document management systems, which file but do not understand.
INTAKE takes over the sorting and the preparation. The person stays where they are needed: at the approval.
02 — Principles
One connector is enough to start. What the pipeline cannot yet assign, it collects — and processes as soon as someone has created the assignment. INTAKE forms stacks of similar documents on its own; when a case worker assigns a stack, it becomes a new type with its rule. The pipeline grows in operation, not in a project beforehand.
An approval is a normal step in the pipeline, not a special case. Every step an agent takes is logged and can be checked in the document's history: which field was recognized with which confidence, which responsibility was looked up, which reply was drafted. Writing actions to the outside world — informing residents, for example — always need a human confirmation.
Whoever changes a classification prompt, an extraction schema or an agentic task tests it against documents whose results people have already confirmed. The editor shows what is right, what is confused and what falls below the threshold — and what improved or is still wrong compared with the live version. A change can only be applied after a test run against the current state; documents in progress stay on their version.
03 — Roles
Pools, stacks, approvals, issues. Configuration is hidden — the stage shows what needs doing.
The same picture, with handles, plus points and settings in the same place. Changes apply immediately to new documents; what is already waiting can be processed again.
There is no separate admin application where you first have to learn where everything is.
04 — In operation
The Unknown counter opens the stacks INTAKE has formed, each with a suggested type. “Apply to all and create rule” turns a stack into a new type with its own rule; the counter drops and the type appears in the pipeline. Undo is possible.
The document detail shows the scan on the left, the extracted fields with their confidence in the middle, and the history of every agent step on the right. Three actions: approve and deliver, send a query, choose a different type.
Every destination has its delivery routes — SharePoint immediately, a digest email in the afternoon, a notice in the team channel. Routes can be paused and resumed.
An expired token, an unreadable scan, a missing channel, an objection deadline about to run out: every issue says what is failing, what that means, and offers exactly one action. No error code to look up first.
05 — Setup
Seven editors. Each one tests against documents whose results people have confirmed.
06 — Data sovereignty
The choice is a setting, not an architecture decision. A document type with personal data can stay on your own model while non-critical tasks use a larger one.
Queries to specialist systems — the population register, for example — are logged. What an agent looked up is recorded in the document's history.
07 — Example
Fictional: 60,000 inhabitants, a mailroom and six offices. The pipeline “City hall inbox” reads the mailroom's inbox, two SharePoint folders, the scanner and the citizen portal.
08 — Start small
From there, the pipeline grows in operation — with every stack someone assigns. INTAKE runs in your infrastructure; documents, history and audit log stay in-house. We will show you INTAKE on your own inbox.