I automated 70% of my work. Here's what remains in that 30%

I am involved in the integration of AI into business. Over thirty projects in recent years: chatbots, RAG systems, document flow automation, analytics. Every time I explain to clients that the neural network will take over the routine and free up time for what matters.

At some point, it became awkward. I sell automation to others, but I sit here writing emails by hand, compiling reports, researching competitors. A shoemaker without shoes. I decided to fix this.

In six months, I systematically transferred everything I could onto neural networks. And here’s what came of it.

What has gone to neural networks

Communication with clients

Previously, I spent about an hour and a half a day on emails. Incoming messages require understanding the context, formulating a response, rereading, and sending. Multiply that by fifteen emails. A routine, but it eats up time.

Now I have Claude with the project context. I drop in the incoming email and receive a draft response. I edit about 20 percent, sometimes less. The neural network knows the tone (businesslike but without bureaucratic jargon), understands the project, and knows the history of communication. I spent one evening gathering prompts with examples of my actual emails. Since then, it has been working.

An hour and a half has turned into twenty minutes. Not because the responses have become worse. The time spent on "staring at the screen and thinking about how to phrase" has simply disappeared.

Commercial proposals

I used to compile commercial proposals manually. You take a template, input the client's tasks, select cases from the portfolio, calculate costs, and format it. Three to four hours for one proposal. And sometimes, the client receives the proposal and disappears. It's frustrating to spend half a day on a document that no one will read.

Now the process is different. After a call with the client, I record key points verbally: what they want, what budget they mentioned, what their pain points are. The neural network receives this recording, my template for the proposal, and a database of past cases. The output is a ready document with relevant examples and calculations.

All I have to do is check the numbers and adjust the phrasing where the neural network overcomplicated things. It likes to write "comprehensive solution" and "synergy of approaches." I have to revert to human language.

Four hours → forty minutes. And it’s not so disappointing when the client disappears.

Research and analytics

A client from logistics asks: "What do your competitors offer for route automation?" Previously, I would have spent half a day googling, reading articles, and compiling data into a table.

Right now, I'm giving a task to Claude or Gemini: gather information on competitors, compare approaches, and produce a table. The result is not perfect. The neural network sometimes confuses companies, attributes one feature to another, and invents products that do not exist. About 30% of it needs to be verified and supplemented. But instead of six hours — it's one and a half. And I have a structure to start with, rather than an empty file.

Drafts of Articles and Posts

I also started this article not from a blank page. I dictated the theses, fed them to the neural network, and got a structure. Then I rewrote about half, added real stories, and removed things that sounded too smooth. The neural network loves to write "in today's world" and "it's no secret that." All of this gets cut out.

Without the neural network, I would have spent the whole evening staring at a blank document. With it — I sat down, made corrections, added my own. Two hours instead of five.

Code Review and Refactoring

Small coding tasks: check someone else's PR, find a bug, rewrite a piece. I used to open the IDE myself and figure it out. Now I throw it into Claude Code — it finds the problem and suggests a fix. I check it visually and accept or reject it.

It doesn't always work. On complex business logic, the neural network struggles. Recently, it suggested an "optimization" that broke the discount calculation logic for one client. The tests passed, but it would have failed with real data. So, blind trust is not an option.

But for standard tasks — linting, minor refactoring, finding obvious bugs — it really saves time.

In Summary: 70% of Routine Tasks Gone

If we count in hours, my workday used to look something like this:

• Correspondence: 1.5 hours → 0.3 hours

• Proposals and documents: 2 hours → 0.5 hours

• Research: 1.5 hours → 0.5 hours

• Content: 1 hour → 0.3 hours

• Code review: 1 hour → 0.4 hours

It was about 7 hours of routine. Now it’s about 2 hours. The other 5 hours have been freed up.

And this is where it gets interesting.

Stumbling Blocks Along the Way

Before talking about those 30% — let’s discuss the mistakes. Because the transition was not smooth.

In the first month, I was enjoying it. Everything was flying, I had plenty of time, I felt like a productivity genius. In the second month, reality hit.

The client received a letter in which the neural network wrote, “as we discussed in our last meeting” — but there was no meeting. It inferred this from the context of previous correspondence. The client said nothing, but I noticed that the tone of communication changed. Later it turned out: he decided that I don't even read my emails.

Since then, the rule is: I read each external email in full before sending it. Yes, that adds three minutes. But the alternative is to lose the client's trust.

Another story. The neural network generated a proposal for a medical client. It pulled a case from the database where we created a chatbot for a bank. Formally, it fits — both involve a chatbot. But medicine and banks are different universes. The client looked at it and asked, “Do you even understand our specifics?” I had to redo it manually and apologize.

The third snag — research. Gemini produced a nice report on competitors, with links. I included it in the presentation. In the meeting, the client opened one of the links. Page 404. Because the neural network generated a URL that did not exist. After that, I check every link manually. Each one.

In short. The neural network is a draft. Good, fast, but a draft. It cannot be sent directly. Period.

Those 30% that cannot be automated

First call with the client

The client calls, describes the task. Says one thing, means another, and needs a third. I understand this from the pauses, from how he formulates things, from what he DOES NOT say.

“We need a chatbot for support.” But in reality, their managers are quitting because they are overwhelmed with repetitive questions, and the bot is not the main issue. The main thing is to lighten the load on the people. A bot — maybe. Or maybe the FAQ on the website and a decent knowledge base are enough.

Recently there was a call. The person talked for half an hour about "the AI transformation of the company." Beautifully, with a presentation. I listened and felt: he himself does not believe what he is saying. I started asking questions. It turned out — the management set the KPI "implement AI," and he is looking for something to implement just to tick the box. There is no real task.

I could have sold him anything. But that would mean getting a dead project and a bad review in three months. Better to honestly say: "You don’t have a task for AI. Let's first find the pain, and then the solution." He was taken aback. Then he thanked me.

The neural network will not do this. It will produce a tidy summary of what was said. But it does not catch what is between the lines.

The Moment of "Something Is Off"

The project is going well, the metrics are fine, the client isn't complaining. But something is gnawing at me. I can't explain what it is. Just a feeling.

I dive in to investigate. It turns out the team has been working on a feature for two days that the client doesn't need. They took the requirements literally rather than understanding the essence. Or the client is silent not because they are satisfied, but because they have lost interest and are preparing to leave for competitors.

One time this saved the project. The client stopped responding to messages throughout the day, although before they would reply within an hour. Formally, there's nothing to worry about; the person is busy. But I called. It turned out that their boss decided to "try ChatGPT without intermediaries" and had already assigned the task to their team. If I had waited another week, the project would have just quietly died.

The neural network doesn't see this. It works with what it's given. And the "something is off" lives in the intuition that has been built over years of similar situations.

To Tell the Client "No"

"Sergey, let's add emotion recognition to the chatbot. And image generation. And let it call clients by itself."

The neural network will respond: "Great idea! Here's the implementation plan..." It doesn't know how to say "no, that's a bad idea, and here's why." It can't explain that emotion recognition in text works at best 60% of the time, and clients will get angry when the bot starts to "calm them down" at the wrong moment.

I got burned on this in one of my first projects. The client wanted everything all at once. I didn't refuse. In the end—six months of development, a budget twice as high, and a result that was mediocre. Since then, I calmly and reasonably say "no." Not "no, because it's complicated," but "no, because here specifically we will lose money and time, and here's why."

No one to delegate this skill to.

Architectural Decisions

RAG or fine-tuning? One large service or three small ones? Claude or GPT for a specific task? These decisions determine whether the project will take off or not.

I can ask the neural network. It will produce a neat list of pros and cons for each approach. But it doesn't know that this particular client has a sysadmin leaving in a month, and there will be no one to support the complex infrastructure. Or that the budget is formally available, but the financial director is cutting everything indiscriminately, and it's better to create an MVP in a week than an ideal system in three months.

Context. All this dirty, unstructured, human context — a neural network does not have it and cannot obtain it. And without it, a technical solution may be correct according to the textbook but fail in real life.

Team Motivation

I have a small team. Sometimes a developer slows down not because the task is difficult, but because they are burned out. Or had a disagreement with a colleague. Or no longer believes in the project.

This is evident in small things. In how a person writes in the chat: earlier they would share solutions, now it’s just “ok.” In how they respond in daily meetings. In that they stopped suggesting improvements.

Claude can write a great motivational letter. But to feel that it is needed right now, for this particular person, in exactly this form — that’s not its job. Once, it helped just to invite a person for coffee and ask, “so, how’s it going?” Without an agenda, without tasks. Twenty minutes of conversation — and the person came alive. Try to automate that.

What I Understood

When the routine left, the essence remained. And this essence turned out to be entirely about people. About hearing what is not said. Feeling what is not shown. Making decisions when there is not enough data, and the deadline was yesterday.

I thought for six months that I would automate my work. But in reality, I cleared it. I removed the husk and saw the core.

Those 30% are not the remainder. They are what people actually pay for. The other 70% were just the wrapping. Necessary, but still wrapping. Like scaffolding around a building: you cannot build without it, but people do not live in scaffolding.

Well, and an unpleasant thought to end with. If 70% of my work is done by a neural network for twenty bucks a month, what does that say about the cost of routine? All this correspondence, reports, research — I thought that was the work. But it turned out that the work is those moments when I put down the laptop and just think. Or call someone. Or say “no.”

How much is that worth? More than I thought before. Or less. Honestly — I still haven’t figured it out.