Why OpenAI is Training New Models on "Real Office Work"

OpenAI is paying people to share examples of past work from former employers. They are asking people to anonymize it. But they are explicitly asking for real work, real tasks, and ones that were at least 1-2 days long.

If you’re paying attention, that should tell you a lot about where 2026 is headed.

AI model companies aren’t optimizing for clever demos anymore. They’re optimizing for work. And that shift quietly changes what matters for all of us.

What it means: Being an AI power user is the wrong goal

Many people we talk to feel “done” learning AI because they use it all the time and consider themselves power users. Totally reasonable. And for a while, that was enough.

But that phase is ending.

Because the limiting factor is becoming less and less about just using AI, but more about what you can get AI to do. 

Experiment to see it for yourself

One of the things we are constantly doing as a team is testing out new tools to continue getting a pulse on AI's capabilities. If you're not regularly doing this, let us tell you about something we did just last week:

We were testing Claude’s new browser extension on a very normal task:

Review a 10-page form and fill in our company details.

And it got about 80% right.

Some people look at that and say, “See? AI still isn’t reliable.”

We see the opposite. 80% is higher than most people likely would get on their first try. Or you’d have people asking tons of questions not sure how to fill it out. 

In fact, what it got wrong, ended up being more obviously a shortcoming in the form rather than the AI itself. 

As models get better, intelligence stops being the bottleneck. How the work is structured becomes the bottleneck.

This is the real skill shift

For the last few years, being “good at AI” mostly meant being a good user. Knowing what to ask. Knowing which tool to try. Knowing how to iterate.

That’s still useful but it’s no longer where the leverage is.

The real skill now is learning to think like an operator.

Power users ask: “What can this AI do?”

Operators ask: “How does my work need to be structured so AI can meaningfully participate in it … or eventually do it on its own?”

That’s a very different question.

And that’s the #1 reason why so many AI efforts stall out inside organizations. Not because the models aren’t capable…

…but because the work was never designed to be shareable with a machine.

A quick way to stretch your AI Operator thinking: Be a QUEEN 👑

When we teach teams how to work with AI, we don’t start with tools. We focus heavily on how you are structuring your work, which ultimately means how you structure your conversation, prompts or workflows. 

One simple shorthand we use is QUEEN – 5 ways to architect how you engage with AI, depending on what the work actually needs.

• Questioning: This is where you are asking AI questions. It can be helpful for researching or thinking through a new topic.

• Uncovering:  You ask the AI to interview you – to pull assumptions, constraints, and half-formed ideas out of your head and into the workspace.

• Executing: Asking AI to do tasks for you. This requires you to delegate clearly, with defined context, instructions and success criteria.

• Emulating: An underrated method – asking AI to act as a persona, and shift its perspective. Think of this as having AI role-play, and can be used to mimic getting feedback from a stakeholder, reviewer, etc.

• Navigating: Using AI as a thought partner to compare options, critique, or decide what to do next.

If you find yourself stuck in a rut with AI, try mixing up how you are approaching and structuring your work and the conversation.

Why this matters more every month

Agents, agentic browsers, autonomous workflows… all of these are getting better, and fast. 

But they don’t remove the need for structure. They amplify it.

The more autonomy we want AI to have, the more explicit our work has to become. And the people who learn to see it that way won’t just “use AI better”…

They’ll shape how work gets done.

If there’s one thing I hope you take from this:

Being an AI user is becoming table stakes now. Now it’s more about how you structure your work, and how you think like an operator.

—

P.s. We are going to work with a small group of teams next month that want to learn to think more like an AI operator. To learn to free up hours of their time, on repeat. Want to join us? Send an email to [email protected] and we’ll see if you’re a fit.

LINKS

For your reading list 📚

• Grok has hit the headlines with their controversial response to their AI model undressing people… good thing this happened after the holiday family dinner tables!

• Google announced their new AI Inbox - reviews have been mixed to say the

  least! What do you think?

• Did you see OpenAI’s new job posting for the Head of Preparedness? Interesting…

That's all!

We'll see you again soon. Thoughts, feedback and questions are much appreciated - respond here or shoot us a note at [email protected].

Cheers,

🪄 The AI Exchange Team