The Playground Was Never Meant to Be Your Office

The chat interface you use every day was built as a showroom, and you moved in.

That sounds like a criticism of the people who did it. It isn't. The move was completely rational, and nearly everyone made it, including us. But it's worth understanding what those products were designed to be, because the design intent explains almost everything that frustrates people about using AI at work.

When the model providers launched their chat interfaces, the purpose was to let you try the model. Type something, see what comes back, form an opinion, come back when there's a new version. They are, in the most literal sense, playgrounds — a place to experience the capability. That is a genuinely useful thing to have, and it's why they're free or nearly free. The intent was never that you would relocate your entire working life into them.

Which is exactly what happened. People started bringing the work to the model instead of bringing the model to the work. The analogy I keep coming back to is that we've moved all the furniture out of the house and into the arcade, so we can be closer to the games. It's not that the arcade is bad. It's that you now live in a place with no kitchen.

Here's what that costs, concretely, and none of it is about the quality of the model.

You re-supply the same context every single time. Every session starts cold. Who we are, what we sell, who we sell it to, what we already tried and why it didn't work, what our pricing looks like, what our client said last month. A person on your team retypes some version of that briefing several times a day, and gets a slightly different result each time because they phrased it slightly differently. The intelligence is excellent. The system has amnesia, and you are the memory.

Nothing happens on a schedule. Work only occurs when a human remembers to open a tab and ask. Which means the work that gets done is precisely the work somebody remembered to think about — and the entire category of work that matters but doesn't announce itself never gets done at all. Nobody wakes up on a Tuesday and thinks, unprompted, "I should check whether any of my dormant accounts had a leadership change this week." That's not a discipline problem. It's an architecture problem: you cannot schedule a thought.

And nothing is inspectable by anyone else. A genuinely useful analysis lives in one person's chat history, invisible to the colleague who needs it, unbuildable-on, unauditable. Multiply that by everyone on the team and you have a company that has done a great deal of thinking and retained none of it.

Put a number on the ceiling this creates. Someone who spends a solid twenty minutes a day in a chat window is buying roughly eighty hours a year of assistance. Every one of those hours was initiated by them, shaped by how well they framed it that day, and discarded afterward. That's the whole return, and it's capped by human attention no matter how good the model gets. A better model makes each of those eighty hours better. It does not make there be more of them.

The alternative isn't better prompting. The prompt-engineering framing is a symptom of living in the playground — it assumes the human will always be the one initiating, so the goal becomes initiating more skillfully.

The alternative is to configure once and let it run. Define the standing context — who we sell to, what a good account looks like, what a real buying signal is versus noise, how we talk, what we will and won't say. Then define workers with narrow jobs and clear boundaries, and give them a schedule. The research runs whether or not anyone remembered. The monitoring runs overnight. What arrives in the morning is a short, ranked list of things that changed and what you might do about them.

And then the relationship inverts in a way that I think is underappreciated. You stop prompting and start interrogating. Why is this account at the top? What did you look at to conclude that? Which sources did you check, and what did you skip? That question is unanswerable in a chat window, where the reasoning evaporates with the session. In a system built for the work, it's just a question you can ask, and the answer is a record you can disagree with.

I want to be careful about one thing, because this is where these arguments usually overreach. None of that means the machine runs the company. Everything that leaves the building — every message to a client, every commitment, every number in a proposal — goes through a person. The point of moving out of the playground is not to remove the human. It's to stop spending the human on retrieval, context-setting and remembering, so the human is spent on judgment.

Here's the diagnostic I'd offer, and it takes about ten seconds. Name the last thing AI did for your business that you didn't personally sit down and ask for.

If nothing comes to mind, you're living in the playground. That's not a failure — it's where the entire market currently is, and the tools were built to put you there. But it does mean your ceiling is set by how many times a week your people remember to ask, which is a strange thing to have bet your operating leverage on.

Moving in is less dramatic than it sounds. It's three things: standing context the system keeps so nobody retypes it, a schedule so work happens without being remembered, and a trail so anyone can see what was done and why. None of the three is a model problem. All three are the reason a company with mediocre tools and good structure will outrun a company with excellent tools and none.

If you want to work out what your version of that looks like, we're glad to walk through it. Bring the three or four things your team currently retypes into a chat window every week — that list is usually the whole answer.