179 people, one Mac mini, one WhatsApp number

Mastan is an AI assistant with no app. No sign-up, no instructions. It's just a contact on WhatsApp and Telegram.

My family used it. So did colleagues, customers of my small business, and strangers who found the number on this website. Since it started, 179 people have talked to it.

I wanted to know one thing: what do people do with an AI when it is just another chat in their phone? Not in a lab, and not in an app they downloaded on purpose. In the same list as their mother, their boss and their group chats.

So I analysed everything that was still on record on 30 September: the conversations of 135 of those people. Here is what I found.

The setup

Nothing fancy. The whole thing ran on one small computer that never switched off.

It broke in ordinary ways. Duplicate messages. Slow replies. And, more than once, an assistant that cheerfully reported it had done something it hadn't. More on that below.

Most people didn't test it. They came with a job.

I expected people to poke at it first. Who are you, are you real, tell me a joke. Some did. But three in four people (101 of 135) asked it for real work at some point, and 59 of them did it in their very first conversation.

47 people started with small talk, a quick question or "what can you do?". About half of them (24 of 47) came back later with something real to do.

More than half of all conversations (54%) asked for a finished piece of work: a document, a report, a translation, a piece of code, an image. People were not looking for someone to talk to. They had something to get done.

Some found one thing and kept doing it

He wasn't alone. Another person needed the same few lines again and again: a short objective for the next activity plan, often sent as a voice note, for four months. Thirty-two of their thirty-three conversations were that one job. Someone else kept checking in with a quick spoken question about where the market was heading. Same question, different day.

Among the 61 people who had five or more conversations, 16 used it almost entirely for one job. They didn't want a general assistant. They wanted one thing done well, in their own language, in a chat they already had open.

Some built things with it

Another regular, after months of using Mastan for everyday work, built a separate assistant of their own with a different tool. Then they came back and asked Mastan what their new assistant was missing, and how to improve it.

Not every build worked. One person wanted a repetitive copying job gone: send a photo, get a filled table. The first version still left the copying to them. The next one copied it wrong. They kept asking for the thing they actually wanted, which was fair.

Some brought their team

24 people used Mastan inside group chats. There, the most common job was coordination: status updates, trackers, notes from meetings.

Trust came fast. Sometimes too fast.

This is the part that surprised me most. People sent Mastan things they would think twice about sending to someone they had just met: bank statements, payment exports, ID documents, passwords, medical reports, spreadsheets from work with other people's names in them.

• 54 of 135 people (40%) shared something genuinely sensitive: financial records, ID documents, passwords, health information or other people's data.
• 23 of them did it in their very first conversation. Half of them did it within a day and a half of saying hello.
• 46 people shared data about someone else: customer lists, contacts, someone else's documents. Some sent spreadsheets straight from work.

One person's very first message was a spreadsheet from work and a request: turn this into a dashboard. Two months later Mastan was producing their recurring reports and presentations.

In WhatsApp you send files to people you trust. An AI in WhatsApp inherits that trust before it has done anything to earn it.

I don't think this is carelessness. It's habit. If you build something like this, plan for data you never asked for, because it will arrive.

People talked to it like a person

• 49 people thanked it. 24 treated it as company: chatting, checking in, telling it about their day.
• In 62% of conversations people used the informal "you" (sən, ты). Only 2% were formal.
• Eight conversations include an apology to the assistant.

Someone gave it a new name of their own and asked if they could vent about a person. Not gossip, they clarified, just to say it out loud. Then they asked a normal question and moved on. Someone else used it as a rehearsal room for personal messages: a little warmer, no flirting, then a sharper reply when the other person took too long to answer.

When it went wrong

One person came because someone had told them Mastan never makes mistakes. Two conversations and a run of confident wrong answers later, they said so. A job seeker loved the application drafts but kept asking Mastan to press submit for them. The writing wasn't the hard part. The last form, on a phone, was.

The biggest group is silence. The job was done, or looked done, and nobody said anything. Silence isn't success. The most reliable warning sign was repetition: people had to ask the same thing again in 23% of all conversations, and in half of the ones that failed.

And the assistant was not a reliable witness to its own work. It sometimes said "done" when nothing had happened. The rule I ended up with: an action only counts if there is proof outside the AI's own words.

Voice notes, nine languages, and 2 a.m.

People wrote and spoke to it in nine languages. Azerbaijani and English carried most of the conversations, followed by Arabic and Russian. Turkish, Polish and Urdu came up much less often, and German and Bengali came from one person each. Many people moved between languages, sometimes within one conversation.

Counting languages turned out to be harder than it sounds. The first automatic pass found 25. Checking each one against what people actually wrote cut that to nine. Most of the extras were Azerbaijani voice notes that the transcription model had written down as Turkish, Kazakh, Persian or even Armenian.

• About one conversation in ten arrived as voice notes. In the family circle it was almost one in three. One person asked for the answers as voice notes too, because reading was hard for them.
• The busiest time was weekday mornings, 9 to noon, and the busiest day was Tuesday. But 6% of conversations started between midnight and 6 a.m.

What I take from it

1. People want a job done, not a chatbot. In the chat they already have open, in the language they already speak.
2. Many need one thing, not everything. A narrow job done well was enough to bring people back: the one-job users returned on a median of 12 different days.
3. Trust arrives before the AI earns it. That is a responsibility, not a feature.
4. Silence isn't success. Repetition is the signal worth watching.
5. Don't believe the AI's own report. Check what actually happened.

An independent research project

Mastan started as a personal study of how everyday people use AI. It is built and run independently, and it grows slowly, on purpose, learning from real conversations.

Reach out to me on LinkedIn