How I accidentally became an AI automation engineer

I never planned to change profession. I simply spent too long automating my own work.

When I joined ChatMe support, I had no plan to become an AI engineer.

There were tickets, client bots, scenarios, integrations and the ordinary support task: work out what broke, and either fix it or gather enough information for the people who can.

The problem was elsewhere. The longer I did this, the more I noticed how much time went not into the solution itself, but into the road towards it.

Find the right dialogue. Pull the analytics. Reconcile data from several sources. Work out which scenario was running at that moment. Check the config. Remember where the relevant logs live. And a week later, repeat roughly the same thing for another project.

At first I was just trying to make my own work slightly less stupid.

First the reports disappeared

One of the first big targets was analytics.

Instead of handling exports manually, I began assembling a process where data could be fetched, processed and checked automatically. Then there were several such projects.

At some point it was hard to call this simply “a handy script for support”. LLMs appeared, along with separate processing scenarios, work with APIs, automatic checks and tools for investigating incidents.

But in my head I still saw it all the same way:

there is repeating work → let us try to remove it

That is probably why the transition turned out to be so unnoticeable.

Then it turned out the answer was not the thing to automate

Once you start working with LLM bots in production, an unpleasant thing becomes clear fairly quickly.

The model’s answer is only the last point of a long chain.

Before it there is scenario state, input data, memory, routing, external APIs, configuration, a knowledge base. After it there is always something too: message delivery, a channel switch, a human operator, a webhook, or one more system that has decided to live a life of its own today.

So gradually I stopped seeing the problem as “the bot answered incorrectly”. The question became a different one:

Why did the whole system arrive at exactly this state?

That already required logs, analytics, configs and an understanding of how a request travels from the user to the final action.

The work shifted quietly from supporting a result to engineering a process.

Then I started building tools not only for myself

I noticed the next shift recently.

Until then most of the automation mainly helped me work faster with projects. But if a tool is genuinely useful, it is rather strange to keep it as a personal collection of spells.

That is how the internal kit for working through tickets appeared. Instead of manually reconstructing a project’s environment every time, a working agent gets a consistent path: identify the right bot, look at the config and the analytics, check the scenario, and only then, if the cause is still unclear, go deeper into the logs.

I wrote about it separately here: I put support into a folder.

What interests me most in it is not even the automation. It is the memory.

Every case worked through leaves behind a little more context about a specific project. The next investigation no longer starts from zero.

At some point I realised I was not building a tool that “knows the answers”. I was building an environment in which a person and an agent can reach an answer more easily.

The job title arrived a little late

The funniest part of the whole story: I do not remember the moment of transition.

There was no morning when I closed the laptop as a support specialist and opened it in the evening as an AI automation engineer.

I simply spent less and less time on individual tickets and more and more on what sat between the ticket and the solution:

• analytics automation;
• LLM scenarios;
• integrations;
• debugging complex chains;
• tools for investigations;
• agents that can work with a project’s real context, rather than only converse nicely.

At some point the name of my role simply caught up with what I was already doing.

And that, I think, is what I like most.

I did not pick a new profession from a list of vacancies, and I did not build a perfect transition plan. I just wanted, over and over, to remove the next meaningless part of my own work.

And when enough of those parts were gone, it turned out there was almost none of the old work left.