In this article I would like to share my opinion on agentic AI, on job losses, and on the current state of the technology. Since this is my first article and the "About Joe" section is not finished yet, first a few sentences about myself. I’m Joe, 42 years young, German, living in Ireland. In private I work a lot with AI, I test models locally, I own almost every subscription to the current flagships, and I also have an OpenClaw network. I’m not a professional, not an IT specialist, but I do have a bit more experience with how AI models work than most people who only use Gemini or ChatGPT. It is exactly this small advantage that hopefully allows me to speak in a way that is not too technical, but still reaches people who do not have many touchpoints with AI. This article is about the huge hype currently shaking up the world of work - will AI replace my job? Why are so many people already being laid off? What is agentic AI, and what does it cost? Because there is quite a lot to clear up here.
What is agentic AI and how powerful is OpenClaw in practice?
Agentic AI means language models that work through tasks independently. In a normal subscription with Gemini, ChatGPT, Claude or whatever they are all called, the user enters a prompt and receives an answer. The more context the model has, the better it can answer questions. An agent handles tasks autonomously. Of course it also needs a prompt - but then it does not simply generate an answer, it can work through an entire project from start to finish without further input. It can also assign tasks to other agents. OpenClaw really exploded at the beginning of this year and yes, OpenClaw is brutal. I myself have an agent in my house that monitors my cameras and reminds me of routines. Her name is Daphne, inspired by the Daphne episode of American Horror Stories, and she sends me a WhatsApp message as soon as she notices anything. She has her own phone number, her own laptop - important: for security reasons, never install something like this on your main computer. But also please do not fall for all that hosting nonsense - you only need an old laptop, ideally Linux, and a heavily loaded bank account. Because this is where the core of this article is hiding, but before we get there, let’s first talk about why a small 9B model is not capable of handling this task.
Why small language models fail and real AI is extremely expensive
A small model will not deliver usable performance as an agent. By small model I mean both inherently small 9B, 12B and 27B models. They are good for local AI setups via OpenWebUI with your own llama.cpp server, but they are not capable of functioning as agents. So-called Mixture of Experts models are also only useful to a limited extent here. On paper, these models often have many billions of parameters, but only a small part of them is ever actively used. So from my experience, Mixture of Experts, or MoE, models are out here as well. To be truly useful as an agent, you need computing power and an extremely large model. A wild comparison: you also would not expect a primary school child to run a global corporation. “Oh, a little bird,” and the child CEO is mentally somewhere else. And this is exactly where it becomes extremely uncomfortable for the companies that are already laying off employees. Because the claim that AI can currently do the work of humans more cheaply is simply a lie. Let’s take a look at what my experience with an AI agent and its token consumption shows.
What are tokens, anyway?
Honestly? I cannot explain it in technical detail. But very, very roughly speaking, a token is about one word. Of course that depends on the language and the average word length, but let’s assume one token is one word. That is basically the unit AI models work with, and the unit AI providers use for billing. A distinction is made between input tokens, the prompt you send to the model, and output tokens, the text the model generates.
How many millions of tokens does an autonomous AI employee consume per day?
With a powerful model, a daily working time of 24 hours, complex tasks, possible customer contact and a very large context window, you have to assume roughly 4 to 5 billion (!) tokens per day. Powerful models are currently listed at around 15 dollars per million output tokens on OpenRouter.ai. So today we are talking about costs in a single day for which a company could employ a human worker for an entire year. These numbers are rough estimates, and of course you would have to look more closely at how much of that is output and how much is the somewhat cheaper input tokens, and surely optimisations can be made here and there even with strong models - but current models are simply many times more expensive than a human employee. My local Daphne had a consumption of roughly 100 million tokens in one day with low load and no active tasks. And as I said, she was only running in the background and had no conversations to conduct. And then comes the next point that still absolutely slows down agentic AI - in other words, an AI that could truly replace an employee - today.
What is a context window?
The context window is basically the short-term memory of an AI model. If you talk to an AI for a longer time or work on code, it is important that it can remember far back. Google was the undisputed king here with Gemini for a long time, until they started compressing the context window beyond a certain length. That is a shame, because in the subscription it was for a long time the absolute king. I even wrote books with it, and it kept the thread over hundreds of pages. But not only that disappeared - in my experience the max tokens, meaning the answer length, were also massively throttled. That is sad, but on the other hand understandable if you are a hardcore user like me who regularly hits the limits even of the paid tier. Because a long context window costs enormous computing power. You have to imagine it like this: you ask the AI something and the moment it responds, it goes through the entire conversation so far in order to make the exchange feel alive. Well, not the entire conversation, but exactly this context window, which is different from model to model in size or in how real it is - meaning compressed or truly real. And going through this context window consumes huge amounts of tokens.
Why the limited context window slows down every virtual colleague
Today’s models use what is called a sliding context window. That means once the context window is full, the oldest part at the back gets cut off and the model no longer has any idea that you talked about it. And even though a 2 million token context window may sound insanely large at first, with an AI agent that has customer contact, this window fills up extremely quickly. Let’s take an AI agent that has phone contact with customers. Of course a separate session would be created for every customer interaction. The model knows via a system prompt what its job is, how it should behave, what it is allowed to do, what it is not allowed to do. If you worked in one and the same session, the AI currently speaking to Mr Wegemann would suddenly be far more likely to talk about the annoying previous customer, Mr Joe. That must not happen, of course. But customer conversations can become extremely long, and if the agent suddenly forgets what was discussed at the beginning, then - well, then you simply see the limits of AI.
Why does the context window fill up so quickly in a real customer conversation?
Because the AI is not only conducting the conversation, but also running background processes. The agent will look at the customer history, it will perform actions on the customer account. Let’s use a car rental company as an example. A returning customer calls for the ninth time because the refund of his deposit still has not been released. The conversation becomes long, the discussion heated - the agent first has to load all previous conversations, or at least the conversation notes, into its context window, and then conduct the current conversation. Then naturally tokens are also consumed for text-to-speech, or TTS, meaning the voice of the agent, and likewise for recognising the customer’s voice, speech-to-text. And with every interaction, all the data in the context window has to be loaded into the model again from scratch. That costs money. A lot of computing power and a lot of money.
But aren’t there special AI models for voice chats?
Ooooh, that is a very delicate topic. The providers of AI models offer voice chats, yes. But these voice chats are stripped-down models that are nowhere near the capabilities of real language models. Whether it is Gemini Live, ChatGPT Advanced Voice or Grok in voice mode - these are models with massively restricted capabilities designed to minimise latency. On social media I often see people mocking AI and its stupid answers. And what they are showing is not the actual model, but those tiny voice chat models. I do not understand why OpenAI, Google, X and all the others offer such models - because even for small talk they are hardly usable and, God forbid, certainly not for business use in customer service.
Can Microsoft’s 1-bit LLMs solve the current cost trap?
As you can see, research has exploded in recent years, but we are still very far from the point where people will lose their jobs to AI on a massive scale. If you look at the development so far, AI could in a few years be at a level where much longer context windows become possible and the costs of token generation decrease. In this area I am watching Microsoft’s BitNet research with interest. Anyone interested can find more information directly from Microsoft (click here) - summarised in simple language: Microsoft is working on making models far less computationally intensive so they can run on smaller hardware. However, this is its own infrastructure - meaning you cannot simply throw an existing model into this new infrastructure. You need a different base, bitnet.cpp instead of llama.cpp, and the models have to be trained specifically for it. I hope we will soon see progress here, including with large models.
The big AI lie: why corporations are not really hiring language models
Because powerful AI models acting as agents in 2026 cause costs many times higher than a human employee. The argument “AI hallucinates, AI must be monitored” is, in my view, bullshit, because development has almost eliminated that issue in advanced models. The only reason why real AI is not taking over real human jobs in 2026 is simply the cost factor.
Algorithm instead of intelligence: the excuse for massive job cuts
And here we are at the real scandal: what companies are currently using has about as much to do with agentic AI as my pug has to do with a professional athlete. Absolutely nothing. Companies sell the use of AI to their shareholders as an innovative thing, as a cost-saving development. But what are they using instead? The simplest algorithms. Yes, exactly - you know it - you call your car rental company or your trusted taxi dispatch and instead of speaking to an employee, you land in a dialogue.
“Are you calling about A, B or C?”
You say C.
“Wonderful, I can book the requested car for you directly. Please enter your account number now so I can find your account.”
And so on and so on. We have known this for decades. That was what AI used to be seen as in the past - an algorithm that answers, but has to choose rigidly from pre-written responses. AI can do much more today. AI can think along. Yes, AI can be empathetic.
Why are so many jobs already disappearing because of AI in 2026?
Because companies think economically. Because companies have marketing departments. Because companies want to save costs. AI can currently support people, but economically AI cannot replace an employee. Not in 2026 and not in 2027. Companies that talk about AI while cutting jobs are using the hype around AI to save costs, but they are not using an advanced algorithm for a customer conversation.
What do you think about this? Feel free to tell me in the comments below this article what you think.