The cost of a shared inbox like info@, orders@ or support@ is rarely the replies. It is the reading. Someone opens every message to find out whether it is an order, a reschedule, a complaint, a vendor invoice or spam, then copies details into another system. That sorting and retyping is where automated email processing pays back, and it is also where mistakes are cheapest to catch.
This page covers the full chain: classify, parse, respond, route, and create calendar events, CRM records and tickets. It also covers Gmail and Outlook permissions and the real risk of letting software send email for you. It also says when a few inbox rules would do the job.
Classify: sorting messages by intent and urgency
Classification makes everything else possible. Each message gets one category from a short list agreed in advance, such as new order, order change, appointment request, cancellation, billing question, vendor invoice, complaint, sales enquiry and not actionable. It also gets an urgency flag and a confidence score. Low-confidence results go to a person, never to an automatic reply.
Keep the list to roughly eight to fifteen categories, built from a sample of a few hundred real emails from your inbox. Known senders, such as a supplier's invoice address or a booking platform's notifications, can be sorted by a plain rule. The language model handles free text a rule cannot read: a customer who writes three paragraphs and asks to move Thursday's delivery in the last line.
Parse: pulling orders, dates and contact details into your systems
Email parsing turns a message and its attachments into named fields. For an order email that is customer, ship-to address, item codes, quantities, requested date and PO number. For an appointment request it is name, phone, preferred times and reason. Vendor invoices are covered on the invoice processing page.
Each field is checked before it goes anywhere. Item codes are matched against your product list and quantities against sensible limits. A date like "next Friday" is resolved against the send date and the sender's time zone, and shown to a person if ambiguous. Customers are matched by email address before a new record is created.
A PDF purchase order, a spreadsheet of line items and a photo of a handwritten form each take a different extraction path. Fields that could not be read are marked unknown, not guessed.
Respond: the email automated response types you approve
Replies come in three levels, chosen per category. Acknowledgments (we received your order, here is your reference, a person confirms by a stated time) are low risk and can send automatically. Answers from approved content, such as your published return policy or a status pulled from your order system, can send automatically once the wording is reviewed and the data source is reliable.
Everything else is drafted, not sent: pricing, complaints, refunds, exceptions, anything that makes a promise. The system writes the draft in the thread and a person edits and sends it.
Hold rules matter as much. Never auto reply to mailing lists, other auto replies, bounces or no-reply senders, or two systems will answer each other in a loop. Skip threads a person already answered, and cap automatic replies per sender per day.
Route to the right person or queue
Routing moves the message to its owner with the parsed fields and draft attached: a label and assignment in Gmail, a folder or category on an Outlook shared mailbox, or a ticket in your help desk.
Record who claimed each message and when. Set a time limit per category, so a complaint untouched for a set period moves to a manager. The limits come from your business, not a default.
Calendar events, CRM records and tickets created from email
To create an event in Google Calendar from an email, the workflow reads the requested time, checks the right calendar's availability, and books the slot or places a tentative hold a person confirms. Outlook calendars work the same way. The confirmation goes to the sender only after the event exists.
The same pattern creates a CRM contact and deal from a sales enquiry, a ticket with category and urgency set from a support request, or a draft sales order from an order email. The workflow searches for an existing record first, writes draft or tentative records until the error rate is known, and links every record to the original message.
Gmail and Outlook specifics and permissions
Gmail and Google Workspace mailboxes are read through Google's mail access with OAuth, where an account owner approves what the app may do. A Workspace admin can restrict which third-party apps are allowed, so involve the admin early. Reading and labeling is a different permission from sending; ask for the narrowest that works.
Microsoft 365 mailboxes, including Exchange shared mailboxes, are reached through Microsoft Graph. Many organizations require admin consent for mail permissions, and an admin can limit an app to specific mailboxes rather than the whole tenant. Set that limit. Benian creates these connections in your accounts with credentials you hold, so revoking access is your setting.
Safety: untrusted senders, attachments and instructions hidden in email
Anyone on the internet can write to your inbox. An email can contain text aimed at the model, such as an instruction to forward recent invoices to a new address. This is prompt injection, and the defense is design, not a better prompt.
The model that reads email should only classify, extract and draft. Sending outside approved reply types, changing bank details, issuing refunds and forwarding attachments sit behind fixed rules and a person. Payment detail changes are confirmed by phone using a number already on file. Unexpected file types are quarantined, and messages that fail sender authentication checks go to review.
When not to build this, and what to measure if you do
If one person handles a few dozen emails a day comfortably, Gmail filters, Outlook rules and saved templates cover most of the value. If most email is structured notifications from one platform, connect to that platform directly. If your team cannot agree on the answer to a common question, settle that first.
When you build, measure weekly: classification accuracy on a sample, parsed fields a person corrected, time to first human touch per category, automatic replies later corrected, and messages left unclaimed past their limit. These numbers decide when a reply type widens or goes back to draft.