The AI feature nobody is shipping: a system that asks you questions

Open any AI product and you get the same shape: you type a task, it executes, you get a result. Confident and fast. It's also missing the one thing that would have made it correct, which is a question.

I've spent the last year building agents for reporting, CRM, and support workflows. What keeps breaking isn't the model. It's the assumption that the person in front of it already knows how to describe what they want.

The execution trap

We've optimized hard for execution. Give the model a clear task and it will do it, often better than the person who asked. That's real, and it's also where most teams stop thinking.

The trouble starts the moment the request is underspecified, which is most of the time. A real example from my own life. I asked an assistant to sort the email from my kids' schools. Three kids, three schools, a flood of messages about events, forms, and payments. The system said "task received" and started categorizing. It never asked who actually manages the schedule. It never asked whether the information matters differently for a younger child and a teenager. It never asked who picks the kids up on which day.

I got back exactly the task I typed. Not the task I had.

The person who actually owns that calendar is not me. They read those emails first, they know the schedule, they handle the forms. Any system that wants to help with school logistics has to know that, and it can't know it unless it asks. Instead it treated my typed sentence as the whole context and produced a tidy, useless summary.

Why nobody ships the question

Asking is harder than answering, for reasons that are mostly economic.

An assistant that asks four questions before acting feels slower than one that answers instantly. Product teams optimize for the demo, and the demo rewards a fast, confident answer. A model that replies "before I do this, tell me who owns the schedule" looks like a worse product in a 30-second clip, even when it's the better one.

There's a cost angle too. Every clarifying turn spends tokens and time. At the top of the market that's real money: serious agent work runs into tens of thousands of dollars a month, and a single hour of a strong model working a hard task is priced like a junior contractor. Asking questions multiplies the number of turns, so the incentive is to guess and move on.

There's a safety angle vendors rarely say out loud. A system that interrogates you is a system that collects your context. Once it knows who owns the calendar, where the kids go, and what you care about, it holds a profile of your life. That's the data that makes the product valuable, and the data that makes it a liability.

So the current generation of models does the easy half well. It executes, and it rarely asks.

Why the asking model wins

The lab that ships a model that genuinely interrogates you takes the market, and I don't think it's close.

Execution is turning into a commodity. Every serious model can now take a clear instruction and carry it out. What none of them do well is build the picture of your world that makes the instruction meaningful. The model that asks good questions, and remembers the answers, compounds. Each conversation makes the next one better. You can't copy that by releasing a slightly smarter model.

If even a fraction of the people already using AI assistants were asked a few good questions and had the answers stored, you'd have hundreds of millions of filled-in profiles no competitor can replicate. The value sits in the memory of who the user is, not in the model weights.

Today's models still wait for you to volunteer context. They treat a short prompt as a complete specification. The gap between "it did what I said" and "it understood what I meant" is where the next product cycle lives.

How to build it

If you're building an agent, the question step isn't a nice-to-have. It's a node in your graph, and it deserves the same care as your tool calls.

A rough shape:

intake ──> assess_context
             │
             ├── context sufficient ──> plan ──> execute ──> respond
             │
             └── context missing ──> ask ──> (wait for answer) ──> assess_context

Two rules I'd put on that node:

1. Ask before acting when the stakes are non-trivial. A password reset is fine to just do. Anything touching money, access, schedules, or other people should trigger a question first.
2. Persist what you learn. A question you ask once and forget is worse than no question at all. The answer belongs in durable memory, keyed to the user, and it should change how the next request is handled.

The loop is the easy part. Deciding what to ask is the hard part, and it's a product problem, not a model problem. The best questions resolve the most ambiguity with the least friction: who owns this, what does "done" look like, what should I never do without asking. Get those three right and the rest follows.

The wider shift

The asking model is one piece of a larger change in how I build with AI.

Stop chasing the word "agent." It's too abstract, and it pushes you to optimize one small piece of a system instead of seeing the whole thing. "An agent that spawned 8,500 sub-agents" sounds impressive and means nothing. I expect the term to fade, the way "prompt engineering" is already fading. Teaching someone to write prompts today is like teaching them to change the screen resolution on an operating system from a decade ago. The models are being built to read your context themselves.

Rebuild the process around the model instead of bolting the model onto the old process. If you're adding an AI module inside a CRM that exists only because software used to be hard to change, you're automating a container that may not need to exist. The useful question isn't how to add AI to this workflow. It's why this workflow has this shape at all.

Pick one primary system and go deep, two at most. Scattering across five half-learned tools is how you end up fluent in none of them.

Access to the strong models will stay expensive and limited. The best systems already sit with a handful of corporations and states. That's not a reason to wait. Build the memory layer, the part that's yours, while the model layer keeps getting rented from someone else.

What I keep coming back to

If I had to bet on one feature for the next year, it isn't a bigger context window or a faster model. It's an assistant that asks me three good questions before it touches anything, and remembers the answers forever.

The model that asks is the model that understands. Everything else is execution, and execution is getting cheap.

---

I build AI agents for reporting, CRM, and support workflows, with human review where it counts. Based in Potsdam, Germany.