---
title: "Processing emails with AI: from intake to approval"
description: "The four steps of email processing — classification, extraction, proposal and approval — and what is best kept deterministic at each step."
source_name: "Gen8"
source_url: "https://gen8.app/en/kennis/emails-automatisch-verwerken/"
site_home: "https://gen8.app/"
source_description: "Managed AI service for SMEs. We set up, connect and manage; you stay in control."
locale: "en"
authors:
  - name: "Tijs Berens"
    url: "https://gen8.app/en/kennis/author/tijs-berens/"
copyright: "© 2026 Gen8 VOF · Dutch Chamber of Commerce 92978789 · Delft"
published_at: "2026-08-29T00:00:00.000Z"
modified_at: "2026-10-22T13:48:14.670Z"
---

# Processing emails with AI: from intake to approval

“Automate the inbox” is not an executable assignment. Split it into four steps — classify, extract data, propose a reply, obtain approval — and each step becomes a separate decision.

By Tijs Berens · August 29, 2026 · Email

## Why the inbox as a whole is unmanageable

A shared inbox contains requests, complaints, invoices, newsletters, internal cc messages and one email that fits no category and is the one that really matters. A workflow trying to handle all that in a single movement cannot be tested or trusted. Four separate steps can.

## Step 1 — Classify

Determine what a message is about and where it goes next. This step offers the highest return and the lowest risk: nothing changes outside the inbox, and an incorrect classification is visible and reversible.

Keep the number of categories small — five to eight — and add one you need that is often missing: **unclear**. A model forced to put everything in a box gives no signal when it loses its way. An explicit uncertainty category is the cheapest quality tool you have.

## Step 2 — Extract the data

For each category, retrieve the fields you need: order number, client name, requested date, amount. Only do this for categories where you actually use those fields; extraction with no recipient is work with no return.

Important: let the model *produce* the fields, but let rules *validate* them. Does the order number exist, is the client known, does the date fall in a working week? Those checks belong in code, not in a model, and catch most errors before anyone sees them.

## Step 3 — Propose a reply or action

Only now does writing begin. A proposal is a draft: reply text, a task, a CRM update or a combination. Two design rules make the difference:

- **Include the supporting evidence.** What is the reply based on — which earlier email, knowledge article or order details? Without that line, the approver has to redo the work.
- **Let the model decline.** "I have insufficient information to answer" is a valid proposal and should be a normal outcome, not an error message.

## Step 4 — Approve and send

Human approval is the starting point for everything sent externally. Approval belongs in the inbox itself: a draft waiting in the email environment the employee already has open, with the option to edit before sending. A separate approval screen stops being opened after two weeks.

Automatic sending is defensible for one type of message: a straightforward acknowledgment of receipt with no substantive judgment. Once a message contains a commitment, amount or deadline, it is no longer an acknowledgment.

## What to measure at each step

Do not measure "AI quality"; measure something concrete at each step: how many messages end up in *unclear*, how many validated fields are corrected, and what share of draft replies are sent unchanged. Those three numbers tell you exactly which step needs attention — and they also provide the argument for later giving a step more freedom.

## Where this goes wrong

The most common pattern: starting at step 3 because draft replies are the most impressive. Without reliable classification and validation, that produces fluently phrased answers to the wrong question — the most expensive type of error because it looks good.

Also: an inbox that receives confidential messages too \(job applications, sickness notifications, legal correspondence\) should be split first. Filtering within the workflow is a weaker guarantee than a separate mailbox.

## What to do with this

Take a hundred messages from the past week and manually sort them into your intended categories. The share you cannot classify without hesitation yourself is the lower bound for what the workflow will mark as unclear — and that number determines whether step 1 is worthwhile. How those four steps look as a working workflow is described under [inbox triage with human approval](<https://gen8.app/en/mogelijkheden/inbox-triage/>).

## Frequently asked questions

Can an AI reply go directly to the client?

Only for a message without substantive judgment, such as an acknowledgment of receipt. Once it contains a commitment, amount or deadline, a person should be involved — which takes seconds.

How many categories should we distinguish?

Five to eight plus an uncertainty category. More categories make classification more fragile while the onward routes usually converge anyway.

What happens to attachments?

Treat them as a separate step with its own decision: which file types are inspected, what happens to the content and how long it is retained. An attachment is often more sensitive than the surrounding email.

## Want to apply this to your own process?

Bring one process to a conversation. We assess suitability, risk and the cost of the first step.

[Book an introduction](<https://gen8.app/en/contact/>)

- Reply within one business day
- No obligation