Chat Is Dead: AI Belongs in the Process

Chat Is Dead — Why AI Belongs in the Process, Not the Process in the Chat

In early June, a report made the rounds in the trade press: a senior employee at the company behind ChatGPT told the Financial Times that “chat is dead.” The product that triggered the AI wave is to be rebuilt — away from the plain input line, toward a platform of agents, coding tools, and connected services. Why the most successful chat product in history is burying the chat is something I can only partly judge from the outside. What I find more interesting is something else: the reasoning touches on a point I have been observing in projects for quite some time.

What Is Officially Cited

The stated reasons are primarily economic. The product has over a billion users, but the majority pay nothing — as a free mass-market product, chat is hard to monetize. Growth today comes from other areas: coding tools and business customers account for a rising share of revenue, and that is exactly where the focus is shifting, shortly before a planned IPO and in the middle of a race for paying enterprise customers. The product vision behind it is a personal agent that leaves today’s interface behind and handles tasks on its own.

That is an understandable business decision, and I would not presume to judge it better than the people who made it.

The Lingering Question

On one point, this reasoning aligns strikingly with what I see in practice: the chat window does not carry as far as a universal interface as the initial enthusiasm suggested. Even the reporting contains a hint — that conversion is weak for more demanding tasks.

But why, exactly? The economic explanation says that the product monetizes poorly. It does not say why the more complex tasks in particular so often fail to reach their goal in the chat. I have no answer to that from Silicon Valley. But I do have an observation from project work — and it has two sides.

The First Side: The User vs. the Empty Field

A chat window demands more from the user than is visible at first glance. You have to phrase precisely what you want. You have to keep the thread across a longer dialogue. You have to read longer answers, make sense of them, and recognize for yourself when enough context has come together to move forward. These are demanding skills.

A small group masters them well — and for that group, the tool feels powerful. The vast majority of users, however, stand rather helplessly in front of the empty field. From my point of view, that explains the gap between “over a billion users” and “hardly anyone pays” better than any pricing model: the tool presupposes an operating competence that cannot be scaled across the board at will.

In a B2B context, this becomes very concrete. The person at the tool there is not the AI enthusiast, but the case worker who wants to close out a task cleanly — a complaint, a quote request, a status inquiry. She has no interest in optimizing prompts. She wants the task to work.

The Second Side: The Tool Gets Worse When It Matters

On top of that comes a property of the models themselves. In long conversations, quality often declines. Context becomes diluted, earlier commitments slip out of view, and with repeated attempts the results not infrequently get weaker rather than better.

That is the uncomfortable punchline: the more extensive and multi-step a task is, the more the chat interface would actually need to carry it — and the worse it carries. Person and model are therefore overtaxed at the same point: where things get long and complex. Precisely where a business process becomes demanding.

What Follows for Business Processes

It does not follow that chat is worthless. On the contrary: for clearly defined sub-tasks it works very well. Summarizing a matter, drafting a text, preparing a piece of information — these are bounded tasks with manageable context, and here the interface plays to its strength.

The expensive thinking error begins one step later. Because a single sub-task can be solved well in the chat, the idea quickly arises to map the entire process around it there as well — or to shift it into a purely dialogue-driven agent setup. That is exactly what fails on what has just been publicly admitted. A good piece of sub-task preparation in the chat is no argument for pulling the entire task into the chat window.

What I See Instead

In my experience, the more viable direction lies elsewhere — and it is almost the opposite of the “away from chat, toward the fully autonomous agent” logic. The process does not move into the AI; the AI moves into the process: embedded where the task already runs, to condense information and support decisions — not to replace the interface where the work is done.

In that sense, “chat is dead” hits a true core — only perhaps for a different reason than the headline suggests. Chat has not failed. It was simply never meant to be the interface for the entire process.