Anyone who sets out to start sorting enquiries automatically asks the same question first: how accurate is the machine? That is the wrong question. A classifier that files 95 out of 100 messages correctly sounds like a decent rate, until you look at the five it gets wrong. The two mistakes it can make do not cost the same. A marketing email in the enquiry lane costs you twenty seconds. A customer enquiry in the bin costs you the customer, and you will never find out. This piece is about building the sorting so that it only ever makes the cheap mistake.
How much actually arrives?
Enough that doing it by hand becomes a job. On 13 January 2026 the German digital association Bitkom published new figures on business email volume: people who use the internet at work receive an average of 53 business emails a day. In 2024 the figure was 40, in 2021 it was 26. Volume has doubled in five years. Fourteen per cent of respondents receive 100 messages or more a day, another 22 per cent receive between 50 and 100.
The shared mailbox is where everything that is not addressed to a named person ends up: enquiries, job applications, invoices, newsletters, booking confirmations, advertising. At 26 messages a day a person sorts that while doing something else. At 53 it is an activity, and at 100 it is half a role. That is the reason automation comes up at all - not because it is modern, but because the volume eats the manual work.
Why is the accuracy rate the wrong number?
Because it says nothing about the direction the errors run in. The largest sorting machine in the world makes the point. Google writes in the Workspace blog that its filters keep more than 99.9 per cent of spam, phishing and malware out of users' inboxes. That is an impressive figure, and it measures exactly one direction: what stays outside.
Somebody else measures the other direction. The Email Deliverability Benchmark Report 2026 by Validity puts the global inbox placement rate at 87 per cent. Put differently: roughly one in eight legitimate, permission-based marketing emails never reaches the inbox. These are two different studies on two different populations, and they sit side by side here deliberately. Together they say this: the same kind of apparatus that is close to perfect in one direction makes enough mistakes in the other direction to keep an entire industry busy fixing them. What that looks like from the sending side is covered in our piece on deliverability.
For your inbox that means a single accuracy rate tells you nothing. Two systems with an identical 95 per cent can behave in opposite ways. One clears away when in doubt, the other lets through when in doubt, and for you that is the whole difference.
Which mistake actually costs you something?
Only one of the two. If a newsletter lands in the enquiry lane, you see it in the same minute you do not need it, and the damage is one glance. If a real enquiry lands in the archive, you see nothing at all. Nobody misses a message they never saw. The only person who notices is the sender, and as a rule the sender does not complain. They ask somewhere else.
There is a second point that automation projects routinely miss: deadlines keep running regardless of which folder a message sits in. Article 12(3) GDPR requires a response to a data subject request "without undue delay and in any event within one month of receipt of the request". What counts is receipt, not the moment somebody reads it. The same holds for declarations with legal effect: a termination or a withdrawal takes effect on receipt, and an email sitting in your mailbox has been received, whatever a filter did with it afterwards. The classifier changes the location, not the clock.
What happens to the case the machine cannot place with confidence?
That single question separates working triage from a pretty demo. Most set-ups know two states: lane A or lane B. Yet a classifier always returns a confidence value as well, and throwing it away means throwing away the only information the system holds about its own uncertainty. Three things follow.
First, set the thresholds asymmetrically. The system may only clear a message away at high confidence. Into the expensive lane, meaning the enquiries, it may file at low confidence. That deliberately produces more of the cheap kind of error, and that is precisely the point of the exercise.
Second, build a third lane. It is not called "other", it is called "uncertain", and it is not a siding but a mailbox with a name next to it and a deadline attached. If nobody collects from it, the sorting has not been automated, only moved. This is where human in the loop stops being a phrase and becomes a responsibility.
Third, separate sorting from answering. Intent recognition decides the route of a message, not its content. A system that files and replies in one step turns a sorting error into a wrong answer to a customer. A silent mistake becomes a loud one. Keep the two steps apart, even when the same tool could do both.
How do you know it is working?
Not from the accuracy rate, but from two other numbers. The first is first response time per lane, taken as a median rather than a mean. An average across 53 messages a day reliably hides the exact case you care about. The first response time of the uncertain lane is worth watching too: if it sits well above the enquiry lane, your third lane has turned into a siding.
The second number only comes by hand. Once a week, take twenty of the messages the system filed away and read them. That is the only measurement that makes error A visible at all, because this error appears in no statistic your system produces about itself. A short piece of arithmetic: at 53 messages a day and an error rate of two per cent in the clearing-away direction, that is roughly one message per working day, around 260 a year. How many of them were enquiries is something nobody knows without looking. That is a worked example from our editorial practice, not a measurement of your mailbox.
The side effect of that sample is the real prize: it hands you the cases you use to sharpen the rules. After four weeks you know which senders, phrasings and subject lines your system systematically misreads, and that is a better basis than any preset a tool ships with.
Three levers for the next fortnight
1. Write down your lanes before you open any tool. Three to five lanes are enough for almost any business: enquiry, existing customer, administration, advertising, uncertain. Starting with fifteen categories builds a system nobody can maintain, and spreads the errors across so many lanes that they stop being visible.
2. Set the thresholds asymmetrically and set up the uncertain lane. With a name, with a responsibility, with a deadline. As long as no person stands next to it, it is not a lane, it is a folder.
3. Measure first response time per lane, and read twenty filed-away messages every Friday. Twenty minutes a week. That is the entire price of no longer being blind to the expensive mistake.
If you want to know which lanes your inbox actually needs, and where your current process quietly loses enquiries, let us look at it together. 📥