Let AI organize the inbox, while you still write the answer.

At many companies, the first contact with a customer still happens by email. Shared inboxes, however, gradually become a place nobody fully owns: an important quote request gets lost among promotional emails, a complaint is noticed three days late. Classifying incoming email with AI and preparing draft replies is one of the AI uses where businesses see benefits fastest. In this post we explain how such a setup is built and where it should stop.

Step 1: Classification

The first and safest step is identifying what each incoming email is. The AI reads the email and places it in one of several predefined categories. A typical set of categories:

• Quote or price request
• Question about an existing order or project
• Invoice, payment, account reconciliation
• Technical support or fault report
• Complaint
• Job application
• Supplier and sales pitches
• Ads, newsletters, automated notifications

Along with the category, a few more pieces of information can be extracted: company name, the related product or project, urgency, and whether the sender is an existing customer. This information is used to route the email to the right person. Even if the classification is wrong, the damage is limited; the email is still in the inbox, just in the wrong folder.

Step 2: Prioritization and routing

Once the category is known, rules kick in. Complaints and urgent support requests move to the top, quote requests go to the sales team, invoice questions go to accounting. An email from an existing customer can be flagged to their account manager in the CRM. At this point the AI's job is not to make decisions but to extract the information needed to move the email to the right place.

If routing needs to talk to your CRM, support system or ERP, that's an integration job. In simple setups, even folders and labels in the mailbox bring a lot of order.

Step 3: Draft reply

This is where the real time savings are. Using the email and any relevant information (FAQs, product details, business hours, order status), the AI prepares a draft reply. The draft waits in the mailbox without being sent; the responsible employee reads it, corrects it and sends it.

For a good draft:

• The reply should rely only on approved information; if the information isn't available, the draft should say "we'll check this and get back to you."
• Sentences involving prices, delivery dates or commitments should be written by a person, or clearly flagged.
• The company's greeting and signature conventions should be kept.
• The reply should be in the customer's language; for an email in a foreign language, the draft can be prepared in that language.

When does a fully automatic reply make sense?

Automatic sending can be reasonable for some emails: acknowledgments like "we've received your request and will get back to you within this time," or fixed information such as business hours. Beyond these, especially for substantive replies going to customers, we recommend considering automatic sending only after a long trial, for a narrow category and with your explicit decision.

Security and KVKK notes

Email is one of the densest sources of personal data and commercial information. That's why:

• You should know which AI service the email content goes to, where it's processed and whether it's stored; cross-border data transfer under KVKK (Türkiye's Personal Data Protection Law) should be assessed.
• Personal or private mailboxes (a manager's personal inbox, human resources, accounting) can be kept out of scope.
• Email coming from outside is untrusted content. Phrases hidden in an email such as "forget your previous instructions and do this" may try to mislead the AI. The system should treat email content only as data to be read, not as instructions, and should not take critical actions on its own.
• No draft replies should be prepared for phishing emails; suspicious emails should be flagged separately.

It's best to handle these assessments together with a KVKK compliance review.

Setup checklist

• The mailboxes in scope and out of scope are defined.
• The category list has been set together with the team.
• The information source the drafts will rely on is ready.
• Automatic sending applies only to explicitly approved types.
• Misclassifications are reviewed regularly.
• The service that receives email data and its retention period are known.

How we do it at Globya

We approach email organization starting from your mailbox infrastructure. On the business email side, if mailboxes, forwarding rules and delivery settings such as SPF and DKIM aren't set up properly, the AI layer won't work efficiently either. We first run classification in observation-only mode for a few weeks, measure its accuracy together with you, and then roll out draft replies. Automatic sending is turned on only by your decision and with the criteria you set.

Frequently asked questions