All insightsAI Automation

An AI agent in the inbox: how case intake works

Enquiries, damage claims, complaints: an AI agent reads, sorts, opens the case and drafts the reply. Your team makes the decisions.

· 6 min read · Stefan Parge

In many companies the work starts in a shared inbox. Enquiries, damage claims, complaints, questions about quotes, it all lands in info@ or service@. Someone reads, sorts, looks up the customer number, opens a case, forwards it and writes a first reply. It matters, and it’s also the part of the job nobody was hired for.

This is where an AI agent can take a lot off your plate without you giving up control.

What the agent does

An agent for case intake does five things at its core, for every incoming email:

  1. Read. The email and its attachments, including scanned PDFs and photos.
  2. Classify. Is this a new enquiry, a complaint, a question about an open case, or advertising?
  3. Extract the details. Customer, contract or order number, date, what it’s about, what’s missing.
  4. Open the case. In the ticket system, CRM or ERP, with the right team, with everything it found.
  5. Draft a reply. By your rules: what may be promised, which documents are missing, which deadline applies.

Your team no longer sees an email. It sees a prepared case with a draft. The work shifts from typing to deciding.

Why rules matter more than the model

Any current language model can read and summarise an email. What it doesn’t know is how your company handles that email. Which claims you settle yourself and which go to an assessor. Up to what amount a goodwill offer is fine. Which phrasing you never use with customers.

Today these rules often live in people’s heads, old emails and a handbook nobody opens anymore. Before an agent writes good drafts, they have to be in one place, with someone responsible for keeping them current. That’s why I built CtxCore Brain: the agent reads the approved version of each rule there and shows in every case which rule applied.

Where a person stays

At the start every reply goes through a person. That’s not distrust of the technology. It’s the fastest way to find mistakes. After a few weeks you can see which case types the agent prepares reliably.

For some you can relax approval, for example an acknowledgement of receipt with the list of missing documents. For anything involving commitments, deadlines or money, a person stays involved. And anything the agent is unsure about lands with your team anyway, with a sentence saying why.

How to start

Not with the whole inbox. I start with the one case type that comes up most often and takes the most manual work. Usually that’s obvious as soon as you ask the team what annoys them.

Then real past emails run through the agent, and we compare its result with what your team decided at the time. Only when that matches does it go to the live inbox, first in shadow mode next to your people. How the testing works in detail is in this article.

Is this for you?

An agent in the inbox pays off when enough emails of one kind come in each day that intake costs real time, when the cases have to get into your systems, and when there are rules your team already decides by, even if they’re not written down anywhere yet.

If you want to know which case type would be the best start for you, let’s talk about it in a free first call. More about my work: AI agents and automation.

Short answers

What does an AI agent in the inbox do?

It reads incoming emails and attachments, recognises what they're about, pulls out the key details, opens the case in the right system and drafts a reply according to your rules. A person checks it and sends it.

Does the AI agent reply to customers on its own?

Not at the start. It writes drafts that a person approves. For simple cases that are always the same, like an acknowledgement of receipt, you can relax that later. For anything involving commitments, deadlines or money, a person stays involved.

What happens to emails the agent doesn't understand?

It passes them to a person with a short note instead of guessing. A good agent knows when it's unsure, and exactly those cases land with your team.

Read something that raised more questions?

Good. That's usually where the interesting work starts. Book a free 30-minute call and we'll talk it through, without a sales pitch.

Book a call