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.