Skip to main content
AI Agents & Automation

Email ClassificationAutomation

Turn a crowded shared inbox into labelled, routed work: quotes, complaints, invoices and spam each go where they belong without a person reading first.

Email Classification Automation: what the work involves

A single info@ address often serves as sales desk, complaints line, accounts payable and supplier hotline at once. Staff open each message to find out what it is, forward it, and hope the recipient notices. Important enquiries wait under newsletters, invoices go unfiled, and nobody can say how many messages of each kind came in last month.

DevKey builds a classifier that connects to your mailbox through Gmail, Microsoft 365 or IMAP. Each message is read, assigned an intent such as quote request, complaint, invoice, job application or vendor offer, and its key details, like order number, amount or sender company, are extracted. Rules then apply labels, forward to the correct team, create CRM or ticket records, or file attachments. Ambiguous emails stay in a review folder, and reviewers' corrections update the examples.

What we build

Core features

01

Intent classification

Messages are labelled according to categories defined with your team, with a separate class for unknown or unusual mail.

02

Detail extraction

Order numbers, deadlines, amounts and contact details are pulled from bodies and attachments and attached to the label.

03

Rule-based actions

Each category triggers actions, such as forward to accounts, create a CRM lead, save the PDF to a folder or send an acknowledgement.

04

Thread and language awareness

Replies stay with their original thread, and Urdu, Roman Urdu or English messages are handled by the same pipeline.

05

Review folder with suggestions

Low-confidence emails show the model's top guess and reasons, so a person can confirm in one click.

06

Volume and response reports

Counts by category, response times and unanswered messages help you plan staffing and spot recurring issues.

Planned for

What we get right before launch

Mailbox access and privacy

Mailboxes contain personal and commercial secrets. We use scoped OAuth permissions, process only the folders agreed, avoid retaining message content longer than needed, and document who can see what.

Never lose an email

Automation that mislabels or hides a message is dangerous. Original mail remains in place, actions are reversible, and a daily digest lists everything the system moved or skipped.

Prompt injection in incoming mail

Emails can contain text that tries to instruct the model. We treat message content as data, restrict the actions the system can take, and keep high-impact steps behind approval.

Stack

Tools and technology

  • OpenAI GPT
  • Anthropic Claude
  • Gmail API
  • Microsoft Graph API
  • Python
  • FastAPI
  • n8n
  • PostgreSQL
  • scikit-learn
Email Classification Automation FAQ

Common questions, answered

Can it reply to emails automatically?

It can send simple acknowledgements, and draft replies for a person to approve. Fully automatic substantive replies are possible for narrow categories, but we recommend starting with review until accuracy is proven.

Does it work with Gmail and Outlook?

Yes, through Google and Microsoft APIs, as well as standard IMAP mailboxes. We request only the permissions needed and explain them to your administrator before connecting.

How many categories can it handle?

More than people expect, but accuracy falls when categories overlap. We usually begin with eight to fifteen distinct ones, review the confusion between them, and merge or split based on real mail.

What about attachments?

PDFs, images and office files can be read, with OCR for scans. Complex or password-protected attachments may need manual handling, and the system flags files it could not open.

Ready to start your Email Classification Automation project?

Tell us what you need and we will come back with a clear scope, timeline and the questions worth answering before any build starts.