Personal AI assistant

Problem

Small admin tasks are spread across three tools: keeping track of a mail, writing to someone to set up a meeting, putting it in the calendar. Each is quick on its own, but the switching adds up, and none of it needs judgment.

What I built

An AI agent in n8n with Gmail, Google Sheets and Google Calendar as its tools. You write to it in plain words.

"Save this email" logs the mail in a Google Sheet with the sender's name and the details, so it can be found again. "Set up a meeting with X on Y date" makes the agent mail the person and create the event in the calendar.

The language model only works out what is being asked and which tool to use; the actions themselves are the tools' own API calls, so it can only do what each tool allows.

The thinking behind it

An assistant like this is the simplest form of agentic AI: a language model that plans, and a fixed set of tools it may use. The model is good at the open part, understanding a sentence like "set up a meeting with Anna next Tuesday" and turning it into a mail and a calendar event. The tools are good at the exact part: the mail is sent and the event created by their own APIs, not by the model guessing.

That split is the same rule as in my other work: AI where its output is cheap to check, plain rules where things have to be exactly right. Here the check is built in, because the person who asked sees the meeting in the calendar and the mail in the sent folder.

Where it stands

A working prototype from 2025, built while learning n8n's agent tools. It is the general pattern behind most AI assistants in companies today: a model that works out the request, and a fixed set of tools it may use.

What stays manual

It works only inside the accounts it is connected to and only through the actions each tool allows. What to save, who to meet and when stays with the person asking.