A shared inbox - info@, customerservice@, planning@ - receives mail throughout the day. Someone reads that mail, determines what it is about and forwards it to the right handler or department. This often happens via an email system linked to a ticketing system. The question is not whether AI can read mail - it has been able to do that for a long time - but whether AI can independently handle categorization and forwarding, or whether a human must keep monitoring.
This task scores high on volume: an inbox with dozens or hundreds of emails per day is exactly the kind of repetitive work that automation targets. High volumes make it worthwhile to train a system or set up rules, even if not every email is handled perfectly.
The structuredness of the task, however, is low (2 out of 5). An email is free text, written by a customer, supplier or colleague who does not adhere to a fixed format. One email is a complaint disguised as a question, another is a quote request with three sub-questions in it, a third is spam that happens to sound relevant. Where a form with fixed fields is easy to classify, an AI model must first recognize the intent in free text before it can categorize it. That is possible, but it goes wrong more often than with structured input.
Customer contact also scores low (2 out of 5), and that is a reason for caution: many of these emails come from customers or partners. A wrongly forwarded email from an angry customer who waits two days for a response costs more than the time the triage was supposed to save. The higher the customer contact, the more careful you must be with full automation - a pattern you also see with tasks in the HR department, where personally sensitive communication often requires just a bit more oversight than expected (see which work can AI take over in the HR department).
Suppose an email comes in with the subject "question about invoice 2024-1187". A classification rule or trained model recognizes the pattern (invoice number, keyword "question") and forwards the email to the finance department. In most cases this goes well and is a textbook example of text recognition that AI can handle today.
But suppose the same email starts with "I have already been called three times and still haven't been helped, this is the last time that I..." and only then mentions the invoice. A model that steers purely on keywords or intent categories may miss the urgency and the need for escalation. This is where judgment (3 out of 5) comes into play: assessing how urgent or sensitive something is, is less clear-cut than simply recognizing a subject.
The outcome depends heavily on what comes into the inbox. An inbox that mainly receives automated confirmations, order statuses and standard questions has much higher structuredness than assessed here. Think of an inbox linked to a webshop with a limited number of recurring question types - there, classification can work almost flawlessly, and the outcome is closer to category 1.
Conversely: an inbox of a law firm or a complaints department, where every email can be legally or emotionally sensitive, has a lower customer-contact score and higher error costs than the average assessed here. There the outcome shifts toward category 3, human work with AI as a tool for pre-sorting.
This is exactly why a task-specific scan says more than a general statement about "email triage". The same task name can be largely automatable at one company and remain mostly human work at another, depending on the type of mail, the sensitivity of the senders and the consequences of a mistake.
For an average shared inbox with mixed content, email triage falls into category 2: partly automatable, with human oversight. Based on classification rules or a trained model, AI can make an initial classification and forward most routine messages directly. In case of doubt - an unclear subject, an emotional tone, an unknown sender - the system should escalate to a human who approves or corrects. Full takeover (category 1) is only realistic for inboxes with strongly predictable, repetitive content and low customer contact. Purely human work (category 3) applies to inboxes where almost every email is sensitive, legally charged or critical to the relationship.
Incidentally, this classification does not touch on staffing decisions: whether and how an organization redistributes the freed-up hours falls outside this assessment and has its own legal requirements. Similar trade-offs between volume and sensitivity also play a role in planning tasks, where you can read which work can AI take over in planning (/welk-werk-kan-ai-overnemen-op-de-planning), or in document processes as shown on the page about the document signing process.
Whether triage in your own inbox falls closer to category 1 or category 3 depends on the type of mail coming in and who the senders are - you cannot tell that from a task name alone. The free quickscan from ftetoai.com gives, with twelve questions and no account required, an initial indication of which part of the hours in your job profile can be taken over by AI today. The full work scan that calculates at task level is still under construction; we do not yet offer that here, but the quickscan already provides an honest starting point.
Vraag maar. Ik ken de kennisbank van deze site; wat ik niet weet, zeg ik erbij.
Answers come from this site’s knowledge base. Not tailored advice, and not a scan of your company.