A client commissioned an AI chatbot, but the solution came from a completely different technology

There's an old engineering maxim: don't automate chaos—first bring order. A story published on Habr perfectly illustrates this and simultaneously shows where artificial intelligence can genuinely provide value, and where it becomes an expensive toy solving the wrong problem.

The situation is painfully familiar: a company turns to developers with a specific request—they need an AI chatbot for customer support. The numbers on the table are convincing: 200 requests per day, four operators who can't keep up, and constant staff turnover that constantly degrades service quality. It seems like a textbook scenario for deploying a large language model—train the bot on chat history, launch it to production, reduce the load on people. This is exactly how dozens of integrators across the market sell AI solutions.

But the development team took a different approach. Instead of immediately setting up chatbot infrastructure, they sat down and manually read five hundred tickets. They didn't skim them, didn't feed them to a neural network for quick classification—they actually read them with their eyes, understanding the substance of each request.

And the result was sobering. It turned out that sixty-eight percent of all requests could be closed by a simple API call: check order status, update data, initiate a return. These are tasks for which AI isn't needed at all—a well-designed interface and backend integration are sufficient.

Another fourteen percent of requests were solved by a wizard form—a step-by-step scenario where the client goes through a series of questions and gets the result without operator involvement.

In other words, more than eighty percent of support load existed not because tasks were complex, but because customers didn't have a convenient self-service tool. The problem wasn't the absence of artificial intelligence, but the absence of basic automation. This is a fundamental distinction that's very easy to overlook in the hype surrounding generative models.

But the story doesn't end there, and this is where it becomes truly interesting. When routine work was removed from the equation, the remaining requests—those complex, non-standard ones requiring human attention—were subjected to AI-powered clustering. And the model discovered something that neither operators nor managers had noticed: a statistically significant group of complaints pointing to a defective batch of goods. The problem hadn't become widespread yet—individual requests were lost in the general stream, and no single operator could spot the pattern by handling tickets one by one. But the clustering algorithm, processing the data in aggregate, identified the anomaly and essentially warned the company of an impending crisis.

This case is valuable because it inverts the conventional logic of implementing AI in business processes. The market is currently obsessed with the idea of chatbots based on large language models. Companies spend hundreds of thousands implementing solutions that essentially do what a well-written FAQ page with a search bar does. Meanwhile, the true power of machine learning—the ability to find non-obvious patterns in large datasets—remains unutilized, simply because they never get there, mired in automating routine work.

In broader context, this is part of a trend gaining momentum in the industry. More and more practitioners are saying that successful AI implementation doesn't start with choosing a model or writing prompts, but with process auditing. Before asking "which neural network should we connect," it's worth asking "what exactly is broken and why." Often the answer turns out to be mundane: what's broken isn't the intelligence of support, but request routing, knowledge base structure, or the simple absence of a "check order status" button in the customer account.

The defective batch story is also an argument that AI is most effective not as a replacement for humans, but as their amplifier. A support operator is physically incapable of holding thousands of requests in their head and identifying statistical anomalies among them. But they're perfectly capable of making a decision when the system highlights a problem. It's in this combination—machine analytics plus human judgment—that real value is born, not in attempting to replace a live operator with a generative model that sometimes hallucinates.

The conclusion is simple and uncomfortable for those selling "turnkey" AI solutions: sometimes the best thing a development team can do is honestly tell the client they don't need a chatbot. And then show where artificial intelligence will actually change the rules of the game.