You Still Have to Know What to Ask

A few weeks ago, we had a roadmapping meeting at work. Different executives took turns describing their vision of the product eighteen months from now. When it was the CEO’s turn, he said something to the effect of: “I want a ChatGPT-style interface. You just hold up your phone and say, ‘Tell me how my comment section is doing,’ and it figures it out and tells you.”

I love this vision. I also think it’s only half the problem.

The other half? In order to use a chatbot, you have to know what question to ask.

And a lot of the time, you don’t.

Search vs Browse

I’ve been thinking about this since I spent years building e-commerce search engines, where we called it the Search vs. Browse problem.

Imagine going to a grocery store with a shopping list. Thanks to this list that you pre-generated, you know exactly what you need. You zip to the right aisle, grab the item right off the shelf, and scoop it into your cart. Efficient and effective. You got what you needed, and you didn’t waste time on anything else.

Now imagine going without a shopping list. You’re ambling through every aisle, scanning shelves, not looking for anything in particular. Your stomach is growling. Your brain is whirling, trying to put together recipes based on things your eyeballs scan. And then you spot it, a white box gleaming in forest of mediocre grocery products: Fruit Roll-Ups. You haven’t had one since you were twelve. But you’re an adult now, dammit, and you can buy whatever you want.

Into the cart it goes.

Now, let’s be honest: would you have ever asked for Fruit Roll-Ups? No. But browsing the store surfaced a desire you didn’t know you had.

The entire current paradigm of AI – the Claudes, the ChatGPTs, the Geminis – is a search interface. You walk in with a list. You type your query. You get your answer. You move on to the next question. It’s extraordinarily powerful… if you know what you’re looking for.

But most of the time, people don’t.

The LLM Should Prompt Me

Here’s the mental model I keep coming back to, the thing that makes me wonder if we’re still in the early days of AI.

Right now, I have to pick up my phone, hold a button, and say: “Hey Siri, when’s my mom’s birthday?”

But notice what had to happen first. I had to think of the question. I had to already be worried about my mom’s birthday, already have it somewhere in my mind, already know that I didn’t know, and spend the energy coming up with the question, overcoming the inertia of whatever it is I was doing in the moment, and ask the question. The AI just answered. I did the cognitive work of knowing what to ask.

Now imagine a different world. Siri just pops up on my phone: “Hey, Brian! Don’t forget your mom’s birthday is next Tuesday. Based on things she’s mentioned to you over the last few months, here are a few gift ideas she might actually like that we can ship in time to get there. Anything jumping out at you?”

Or: “You have a board meeting Thursday. I’ve prepared a draft for you to review.”

Nobody asked, but it surfaced things I needed to know. And the system knew that.

That’s the half I’m talking about, the thing that’s missing from my CEO’s vision. The current paradigm is that I prompt the AI. The thing that changes everything – the actual killer feature – is when the AI prompts me.

The AI That Knows What I Need Before I Do

I’ve described this to a few people and the response is usually some variation of: “But that’s what agentic AI is doing.” Maybe. I’ll be honest with y’all: everything I’ve personally interacted with still requires me to initiate. I prompt it, it executes. The oven still nees to be turned on, the calendar still needs to be opened, I still ask the question.

What I’m describing is something that’s watching, understanding context, developing a model of what I care about – and then deciding, on its own, that now is the moment to surface something. Not because I asked. Because it figured out I should know.

I’m not talking about creepy surveillance and building some sort of clandestine profile of me in order to help me buy my mom Lego flowers. No, that’s not what I mean at all. This doesn’t have to be privacy-invasive. It can be based on the questions I do ask. I imagine systems that quiz me and develop a profile based on that, or things that learn from the things I explicitly do and from the questions I ask and formulate a model of what I care about.

That might already exist somewhere in a lab. But it hasn’t reached me yet. And I don’t think it’s reached most people. Which means there’s still a very large, very important problem left to solve.

The grocery store is full of Fruit Roll-Ups nobody’s thought to ask for yet.