In short: Nobody needs an AI that “manages” the inbox. What helps is knowing in five minutes each morning which few messages really need a reply today, and already having a draft for each of them. That is quite feasible with today’s tools. What matters is being honest about which channels can be connected and which cannot.
The real job is the morning
When business owners come with the wish to automate their email, it sounds at first like a technology project. Interfaces, integrations, agents.
The real job is a different one. Every morning, know in five minutes which three to seven messages need a reply today, and already have a draft for each of them. The rest can wait or just gets read.
That is why in the first conversation I do not talk about APIs. I talk about the morning routine. When do you look at your inbox for the first time? What do you look for first? What annoys you when you see it too late?
The system emerges from the answers. Not the other way around.
Four piles instead of a folder structure
In practice, a few clear categories work best. I work with four:
- Now: needs a reply today.
- This week: needs a reply, but not today.
- Read: for information, no reply needed.
- Unclear: the system is not sure.
The fourth pile is not an oversight, it is intentional. Without it, the model makes a decision for every email, even when it would really have to guess. Then the important email from a new customer confidently lands under “Read”. Why exactly these silent errors are the expensive ones is explained in The Most Expensive AI Errors Are the Ones Nobody Sees.
For every message in the “Now” pile, a draft reply is ready. It sounds like you, because it is based on examples you wrote yourself. And it invents nothing. Where a price, a date or a commitment belongs, there is a visible gap that you fill.
Nothing goes out automatically
The most important principle is the simplest: the assistant sends nothing. It sorts and writes drafts. You do the sending.
That sounds like less automation than would be possible. It is. But an embarrassing auto-reply to an important customer costs more trust than the whole system saves in time. And you need that trust so that you are still using the system after four weeks.
For the same reason, nothing runs on its own in the background at the start either. The first run is always a preview: this is how I would sort, does that fit? Only once the sorting has settled over a few weeks can it run automatically in the morning. An automation that assigns labels unasked before you trust the system is the fastest way to getting switched off.
Your rules emerge while sorting, not in a questionnaire
A typical question at the start of a project is: by which rules should emails be sorted?
Hardly anyone can answer that well. People cannot name their own rules in the abstract. But they can sort.
That is why at the beginning I have 20 real messages from your own inbox sorted and take notes on why. “This one is urgent because the customer has already followed up twice.” “This one can wait, it is a newsletter from the association, but I do read it.” From these reasons, a rules file emerges, in your words.
Every correction in daily use is entered there as well. That way the system does not learn in the abstract, but traceably. You can read and change the file at any time. So after four weeks, the system belongs to you, not to me.
Honest check: which channels can be connected
Many want not just email, but all messages in one place. Here one has to be honest, and this is the part many providers skip:
- Email (such as Gmail or Microsoft 365): quite feasible. There are official, stable interfaces. Nothing stands in the system’s way here.
- LinkedIn messages: limited. The messaging interface is reserved for partners. What works is limited and has to be checked case by case.
- WhatsApp on a private number: not cleanly feasible. There is no official interface for it. Unofficial solutions violate the terms of use and can lead to the number being blocked. For health, legal or HR data, I rule them out as a matter of principle.
A channel without an interface does not mean the benefit disappears. It only means that this channel is fed in by hand. That should be known before the project starts, not after.
Data protection: you are processing third parties’ messages
A point that is often forgotten: your inbox holds data from customers, suppliers and applicants. When an AI system reads these messages, it processes personal data of third parties.
That is no reason not to do it. It is a reason to do it properly. Every tool used needs a data processing agreement. Anyone who puts a no-code platform like Make or Zapier in between has thereby added another service provider with its own agreement. Summaries and briefings should be deleted after a fixed period. And the responsibility for the data stays with you as the company. I build the system, but you decide what it is allowed to read.
Start with no-code, do not forget operations
For the start, I usually recommend a no-code setup. It is up in a few days, manageable in running costs and something you can follow yourself. A custom-programmed agent only pays off once it is clear that the system is used daily and hits the limits of the platform.
What is always part of it is operations. Such systems rot quietly. A new customer, a new employee, a changed email address, and the sorting no longer fits. Without someone maintaining the rules, the project is dead after a few months. That can be a person on your team whom I train. It just has to be someone.
What this means for you
Before you automate your inbox with AI, clarify:
- What does your morning look like today? How long does it take until you know what matters?
- Which channels really belong? And which of them have an official interface?
- Is there an “Unclear” category? If a provider does not have one, ask.
- Does the system send on its own? The answer should be “no” at the start.
- Who maintains the rules once it is running?
You do not notice a good inbox system. You only notice that the morning has gotten shorter.
If you would like to know what this could look like for your inbox, we talk about your morning in a first call, not about technology.
