Inbound mail building block: classification and suggested reply

One foundation for everything that arrives by email or scan.

WhatsApp
Agent sheet
Category
Back office, finance and administration
Model
open, installed on your premises
Data
never leaves
Subscription
€0 per month
01Overview

Inbound mail is the raw material of half the agents we install. This building block reads it, understands it, classifies it and proposes a reply; the other agents (HR, management, sales admin) rely on it.

The problem

Each department sorts its mail its own way, scanned attachments are not read, the history of an exchange is lost from one message to the next and replies start from scratch.

Our answer

OCR and vision reading of all formats (PDF, images, scans, Office), classification by type and department following your plan, conversion to normalised PDF, suggested reply that takes previous exchanges into account through a long memory stored on your side.

02How it works

Live in four steps.

  1. 01
    Classification plan

    We formalise your categories and their routing rules.

  2. 02
    Reading

    OCR and vision on your real formats, tested on a sample of past mail.

  3. 03
    Memory and replies

    Exchange base (PostgreSQL and pgvector) and reply templates per category.

  4. 04
    Wiring

    Mailboxes, network folders or SharePoint as input; business agents as output.

03What you gain

Results you can measure.

One foundation

Every new agent reuses the same reading and classification.

Contextual replies

The suggestion knows the file, not just the last message.

Everything stays with you

Mail, memory and model on your server.

04Frequently asked

What we get asked before signing.

Are handwritten signatures detected?
Yes, the vision model checks for a signature and a date when the category requires it.
Does the agent need an AI vendor subscription?
No. It runs on an open-source model, picked from the right family, installed on your server or VM. No API key, no token bill, no data sent outside.
What hardware is required?
From 8 GB of RAM on CPU for this kind of specialised task. We measure real throughput on your machine during the audit and write it into the proposal.
How long does setup take?
Two to six weeks depending on the connections to build. You validate an acceptance run on your real cases before going live.

Let's talk about your case for thirty minutes.

A conversation to understand your context, volumes and tools. You leave with a scope and an order of magnitude, no commitment.

WhatsApp
+261 38 66 261 00team@wikolabs.com