The chat box has dominated AI product design for two years now. You type, it responds, you type again. It's a tool: a function with a fuzzy input. Whatever it produces, you own the next step.
This works for some things. But it's the wrong mental model for what LLMs can now do. The most useful AI we've shipped in the last twelve months doesn't look like a chat box. It looks like a colleague who's been on the team for three weeks: not omniscient, occasionally wrong, but capable of holding context and doing the thing.
The difference, in one line.
Responds to requests.
You ask. It answers. You decide what to do with the answer. Every interaction starts from zero: no memory of the last one. The user is always the one driving.
Holds context. Does the work.
It knows about the customer, the order, the policy. It uses tools on its own. It hands off to a human when confidence drops. It owns the outcome, not just the answer.
The shift isn't bigger context windows or smarter models. It's a system design shift. A partner-shaped AI has state, tools, evals, and a hand-off protocol. A tool-shaped AI has a prompt.
What changes when you treat it as a colleague.
Three things change, in our experience building these systems.
1. Onboarding becomes real. You don't prompt-engineer your way to a competent colleague. You give them documents, examples, an eval set. The good ones improve over weeks of feedback, not minutes of prompting.
2. The interface disappears. The best agent we shipped this year doesn't have a UI. It reads support tickets and drafts replies inside the existing ticketing tool. The human reviews, edits if needed, and sends. The “AI” is invisible. The value is the shorter response time.
3. The hand-off becomes the product. Tools never need to admit they don't know. Partners do. The interesting product work is in the moment of confidence drop: when the agent says “I'm not sure about this one” and routes to a human. That's where trust gets built.
The most useful AI we've shipped doesn't look like a chat box. It looks like a colleague who's been on the team for three weeks.
The honest limits.
This isn't the “AGI is here” argument. It isn't even the “AI will replace your team” argument. Both of those are bad arguments made by people selling things.
What we've learned: a partner-shaped agent is good at the same things a competent junior colleague is good at. Routing tickets. Drafting first-pass replies. Pulling the right policy doc. Spotting outliers. Asking the right clarifying question.
It is not good at: making judgement calls under ambiguity. Holding a relationship over years. Knowing when the customer is angry-because-frustrated vs angry-because-wronged. Those things stay with humans. That's a feature, not a limit.
What we won't let an agent do.
None of this is new. Stripe's engineering posts have been saying it for a year. What's new is that the gap between “tool” and “partner” is now mostly a system-design gap, not a model gap. You can build a partner-shaped AI today, on top of last year's models, and it will out-perform a chat box.
The next two years of AI products are going to be won by the teams who get this. The chat box was a useful demo. It is not the product.
— A.B., from the studio.