Inbox

This is not support software. It is revenue recovery.

The most common question before buying is “what’s on that date?”, and it is the one most often left unanswered. Every one of those is a sale that was already made.

And what makes it invisible is that it leaves no trace anywhere. Replies get counted; what nobody answered does not. So the loss shows up in no report, and the conclusion at the end of the month is that everything got handled.

One inbox, and a person is a person

Web chat, forms and email land in the same place, become a ticket and sort themselves. Someone who writes to you on three channels is one contact, not three, with their purchase history in front of you.

And with that comes the first metric nobody in this sector measures today: time to first reply, by channel and by time of day. It is the figure that explains why bookings get lost on a Sunday afternoon.

The classifier is trained on 11,028 real conversations and recognises 48 distinct intents, with measured coverage of 81.5%. It is not a generic model that has your business explained to it in a prompt: it was built on what your customers actually ask.

The knowledge base comes out of your own history

This is the part nobody is going to assemble for you by hand. Your good answers are already written — across years of emails, messages and conversations — they are just scattered, and nobody has ever gathered them up.

That is where the knowledge base comes from: versioned per article and per language, with the variables that change — opening times, prices, capacities — held apart from the text, so that updating one figure does not mean rewriting thirty answers.

And the loop does not close there. The questions that keep coming up and are still not answered on your site surface grouped by intent. Every one you put on the site is a conversation you no longer have to have.

Autonomy is earned with evidence

First it suggests drafts against that base. Then it answers on its own out of hours, and only on low-risk queries. Moving from one stage to the next is decided by the human correction rate, not by a hunch — and every correction becomes a test case, so the same mistake does not come back.

Hard guardrails

No price that does not come from live data. It never confirms a booking. It never asks for payment details. It escalates to a person for any complaint, any charge and any minor. And it replies in the language it was written to in.

And about the data you give us: personal information is stripped on the way in, not at analysis time. What gets stored from a message is the redacted text and a record of what kind of data was removed and how many — never the value. A phone number or a card a customer of yours typed into a message never comes to exist in the system.

Where it is today

In development, and it is worth being exact, because this is the piece where it would be easiest to promise too much. What is done and verified is the foundation: isolation between clients proven with tests against a real database — not entrusted to application code — across all 21 tables of the domain, the contact model, the knowledge base and the measured classifier.

What is missing is the API, the interface and the connectors for each channel. Until those exist, this page says so in the future tense.