Automating Email and Incident Handling with AI
· by Panayot Dimkov
When someone talks about automating email management with AI, they almost always picture a robot replying to customers. That is the last step and the riskiest one. Before it there are three levels — classify and route, extract and log, draft a reply — that deliver most of the benefit at far lower risk, because none of them sends anything without a person seeing it first.
1Four levels of automating email with AI, from least to most risk
| Level | What it does | Risk |
|---|---|---|
| 1. Classify and route | Reads the message, decides what it is about and assigns it to the right team or person, with a label and a priority | Low |
| 2. Extract and log | Pulls the data out of the body or the attachment and creates the ticket, order or record in the relevant system | Low |
| 3. Draft a reply | Writes a draft with the context already gathered, for a person to review and send | Medium |
| 4. Reply without review | Answers directly, with nobody seeing the message before it goes out | High |
The bulk of the saving usually comes from levels 1 and 2, with a noticeable quality improvement from level 3. Level 4 only makes sense for narrow, verifiable cases: confirming receipt of something, giving an order status, sending a standard document. And if you do reach it, Article 50 of the EU AI Act, applicable since 2 August 2026, requires that people interacting directly with an AI system are informed that they are doing so, unless this is obvious from the context.
2Start with the shared mailbox
The prerequisite is not technological. If every person receives requests in their personal inbox, there is nothing to automate: no common queue exists to act on.
3What AI does well when reading and classifying emails
Rule-based classification — if the subject contains this word, send it there — has existed for decades and falls short, because people do not write with the expected words. A language model understands the message even when it is badly written, in another language, or covers two topics at once.
- Telling the real intent. "It hasn't arrived" could mean an order, an invoice or a document. A keyword rule does not solve that; understanding the sentence does.
- Detecting urgency and tone. An angry customer or a critical issue gets flagged and moves up the queue before anyone reads it.
- Reading the attachment. Many requests arrive in a PDF or a spreadsheet, not in the body of the email. That information can be extracted too.
- Summarising long threads. A twenty-message thread forwarded three times becomes a paragraph with what you need to know.
- Fetching context before you ask. If the message mentions an order, the system already pulls its status from the ERP and attaches it to the ticket.
4Four cases that work well
5What not to automate
There are messages where an automatic reply, however good, does more harm than good.
6How to know it is working
Four indicators are enough, and they should be measured before you start so there is something to compare against: it is the same baseline an ROI calculation needs.
| Indicator | What it reveals |
|---|---|
| Time to first response | The one customers notice most and the one automatic routing lowers most |
| Percentage classified right first time | The system's real quality. Lots of internal forwarding means the classification is not tuned |
| Messages requiring intervention | It should fall over time. If it does not, the flow is handling the easy part and leaving the rest untouched |
| Reopened cases | The honest indicator. If things close fast but the customer writes again, nothing was resolved |
A warning about the last one: it is almost always forgotten and it is the only one that reveals whether the gain in speed was paid for with worse quality.
Frequently asked questions
Will AI reply to my customers without me seeing it?
Only if you configure it that way, and in most cases that is not advisable. The split that works best is for the system to classify, route, log and prepare a draft, with a person reviewing and sending. Autonomous replies are reserved for narrow, checkable cases such as confirming receipt or giving an order status.
Do I need a ticketing tool for this?
It helps a lot, but it is not essential to start. With a shared mailbox, a labelling system and somewhere to log what comes in, you can already build classification and routing. What is essential is that requests stop arriving in personal inboxes.
What about messages in other languages?
This is where the difference against keyword rules shows most. Current language models classify a message in Spanish practically as well as one in English or French, and can summarise it in whatever language your team uses. For a company serving customers in several countries, that alone justifies level 1.
Can it connect to the ERP or the CRM?
Yes, and that is where the flow goes from useful to valuable. If a message arrives with the order status, the customer balance or the incident history already attached, whoever answers does not have to open three applications. The connection is made through an API where one exists and by other means where it does not.
What if the system classifies something wrongly?
It will happen, and the design has to account for it. What matters is that reclassifying is trivial for the user and that the system uses confidence levels: what it is unsure about it does not assign, it leaves in a review queue. A flow that is confidently wrong does more damage than one that asks.
In short
Email is not automated by putting a robot in charge of replying. It is automated by removing the work that comes first: classifying, looking up the data, logging and assigning. Those are levels 1 and 2, and that is where most of the recovered time sits.
Assisted drafting comes next and improves quality more than speed. And autonomous replies, when they arrive, should be limited to what can be verified. That is how we approach AI process automation.
How many emails land in your shared mailbox each day?
With that and a sample of request types, we can estimate what share would classify and route by itself. We review it with you and tell you where to start.
Request a process analysisNexum Blog
Related articles
8 Apr 2026
Predictive maintenance in automated plants: what is it and how to implement it?
Discover what is predictive maintenance in automated plants, how to implement it step by step and what technologies you need to reduce downtime and costs.
7 Oct 2026
Process Automation ROI: How to Calculate It
How to calculate process automation ROI for admin work with your own data: what to measure first, which costs get forgotten and how to check it afterwards.
3 Jul 2026
UKCA vs CE Marking for Control Panels
UKCA vs CE marking for electrical control panels: which directives apply, the real 2026 status and what you need to export panels to Great Britain.