AI in the office: the super-secret race between anthropic and OpenAI to build your next “virtual assistant”
Imagine an AI that doesn’t just talk to you, but actually does your work for you. That’s the big race leading artificial intelligence companies like OpenAI and Anthropic are in right now! The tech world is buzzing, especially after Jensen Huang (the NVIDIA boss) joked that IT departments might just become “HR for neural networks.” This whole idea of AIs becoming autonomous office workers? It’s quickly becoming the defining trend.
A whole new way to teach AI
Traditionally, teaching AI was a two-step process: first, it gulped down tons of data, then it got a little fine-tuning for a specific job. But for real office work? That’s just not enough. An AI needs to learn how to actually use applications, just like a person would – clicking around, understanding what happens when it does something, and working towards a goal.
This huge jump from merely “knowing” things to actually “doing” them has sparked a whole new approach to AI training. Instead of just learning from static piles of data, these AI models are now basically going on “internships” inside virtual copies of real office applications.
It’s like a flight simulator, but for AI!
These special training grounds are called “RL environments” (Reinforcement Learning environments), and they’re kind of like flight simulators for AI agents. The top labs are already building them. Take Turing, for example, an early startup that created over 1,000 virtual trainers – everything from fake Airbnb and Zendesk setups to Microsoft Excel sheets. OpenAI itself uses some of Turing’s tools, and Turing recently raised a cool $111 million in June 2025.
Others are jumping in too. Scale has snagged a massive $14 billion from Meta, and their competitor, Surge, is reportedly in talks to raise a billion bucks!
So, what does an AI agent’s “training day” look like?
- The instruction: It gets a natural language command. For instance, “Analyze the Salesforce database, find clients we haven’t contacted in over six months, and email them to set up a meeting.”
- The action: The AI “wakes up” inside the app’s virtual interface and starts trying things out, often by trial and error.
- The feedback: A detailed checklist of correct actions ensures every step is covered, and a system verifies if it finished the task. If it nails it, that strategy gets reinforced. If it messes up, it gets an error analysis to help it adjust for the next try.
The brilliant part about this? Scale and safety. An AI can repeat the same task for hours, turning actions into “muscle memory” – all without accidentally spamming real clients, messing up a live database, or deleting crucial info.
The high cost of good teachers
To make this training as effective as possible, companies are hiring super-specialists from all sorts of fields – biology, software engineering, medicine. These experts show the AI how to use workplace tools correctly. The AI doesn’t just copy the steps; it tries to understand the expert’s decision-making logic.
As AI gets smarter, the bar for these teachers gets higher. Early on, “student-level” knowledge was fine, but now labs are recruiting top-tier professionals from places like NASA and other government projects!
Naturally, demand is pushing prices way up. According to Labelbox, which supplies specialists to giants like OpenAI, about 20% of their contractors earn more than $90 an hour, and nearly 10% pull in over $120. They’re predicting rates for the absolute best experts could jump to $150–250 within the next year and a half!
OpenAI is planning to shell out about $1 billion on these experts and RL environments just in 2025, and a staggering $8 billion by 2030. Anthropic, rumor has it, might pour up to $1 billion into creating and using these virtual applications in the next year alone.
The real prize: beyond chatbots
You might think Anthropic and OpenAI are just spending billions to make their AIs a little bit smarter. But the truth is, the goal is much, much bigger: they want to break through the current limits of AI and create a whole new business model.
First off, simply feeding huge amounts of internet text to Large Language Models (LLMs) to predict the next word is starting to hit its limits. RL environments offer a completely different path. They allow AI to do more than just generate text; they can act within complex, multi-step processes – and that’s the key to true autonomy. As Edwin Chen, CEO of Surge, puts it, the methods Anthropic and OpenAI are using “mirror how people learn,” putting AI models in environments as close to the real world as possible.
Most importantly, these RL environments promise a huge payday. For these AI giants, just selling access to a chatbot API is only the first step. The next, far more valuable model? Renting out “virtual employees.” In this new hybrid world, AI agents will mostly handle data processing and all that administrative grunt work.
On one hand, this shift makes us optimistic about huge boosts in productivity and making routine tasks much smoother. On the other hand, it understandably sparks anxiety about jobs disappearing.
Will they take all our jobs?
Startup Magazine argues that the ideal future isn’t about replacing people, but about augmenting them – giving them powerful tools. They use customer support as a great example:
- From the employee’s side: Imagine a digital assistant, super-smart on your company’s knowledge base and able to sound just like your brand, taking care of 80% of those repetitive tasks – tracking orders, answering FAQs, generating standard reports. That frees people from the boring stuff, reduces burnout, and lets them focus on the truly complex, emotionally charged cases that need human empathy and creative thinking.
- From the client’s side: They get instant, accurate answers to simple questions anytime. And when they do reach a human, that service is much deeper and better because the specialist isn’t bogged down by drudgery.
Experts also note that success will largely depend on how companies approach this. Leaders need to prioritize how employees feel, invest in retraining, and keep experimenting. The fact that big firms are betting so heavily on AI agents indirectly supports this view: the goal isn’t to replace humans, but to give them a powerful partner to handle most of the day-to-day tasks. Plus, studies show that despite the rapid advances in AI, it hasn’t really shaken up the job market significantly yet.
An economy-sized simulator?
The vision for the future inside OpenAI’s walls is even bolder. According to The Information, one of the company’s senior executives privately shared his expectation that the “entire economy” could eventually become one big “RL machine.”
That would mean AI learning not just from fake simulations, but from actual recordings of professionals’ real-world workflows around the globe: how a doctor makes a diagnosis in a medical system, how a logistics expert optimizes supply chains, or how a lawyer drafts a contract. Mind-blowing, right?
