Lauren Tan (@poteto on X), a SpaceX engineer, recently asked her audience whether they believed chat was the final form of agentic UI. The responses were overwhelmingly no.
That's not to say chat isn't an essential stepping stone (it is). But the discussion suggests we're still nowhere near the limits of what's possible with today's technology.
https://x.com/poteto/status/2111321176769442251?embedable=true
It's tempting to think everything can be done through chat, for a few reasons:
It's natural: It's how we've communicated for millennia, so it's familiar.
It lets us go as deep into any topic as we want, given the right context.
Done right, it can perform extremely complex actions from minimal instructions, often with impressive accuracy.
That said, chat has its limitations:
Some things are simply better shown in other formats (duh): Images, graphs, or videos, for example.
Why ask? Some actions shouldn't require a prompt. You just want them done, and then explained. For example, an app could automatically fix an obvious data issue it's highly confident about.
Beyond that, I believe the software of the future will also advise users on how to do their jobs. A CRM could act as a marketing expert, helping a partner make better decisions. Contract software could suggest the best clauses to include in certain situations.
I believe the ideal UX will be determined primarily by two factors:
- Whether an action is reactive or proactive.
- How much the user knows about the topic.
Pattern #1: When the users know exactly what they want
In this case, classic chat with the user and context-based intelligence (which may be memory) is enough. Let the user ask for what they want and tell us how to do it, and we can run it on a recurring basis.
Let’s say the user is focused predominantly on controlling the app’s return rates. They are likely to ask questions about return rates and their drivers. They would have repeatable steps that help them get better. In such cases, it’s predominantly about learning and proactively improving the app around the workflows they truly care about.
Therefore, the visual patterns I believe make sense are:
Chat as a triggering input mechanism with Intelligent UI - leverage the chat option to allow the user to initiate any new research with. With the dynamic UI showcasing the information in the best possible represenation
Infinitely probe-able UX. Every answer in a classic analysis often leads to more questions. For example, understanding last month’s sales may reveal a few spikes that need a deeper look. Those spikes may come from a category, a specific sale, or a campaign. The UI should anticipate the questions the user is likely to ask, simplify their path, and coach them toward best practices. (E.g. the leading question in the screen above)
“What caused return rates to go up on the 28th of May”
Learned dashboards, pivoted around what the user cares about most. Over time, we learn what each user cares about, which is usually a specific subset of items. We want to make sure that subset is available across the board.
We also want to give users the ability to research their specific question in more depth at any point.
Pattern #2: When the user doesn’t know what they want, consultative productization is key
There will be many cases where the user doesn’t know what questions to ask. Are sales lower because of seasonality, because the seller is in a category with no demand, because the category is highly competitive, or because of a problem with the product itself? The seller may not know where to start, and the dashboard can feel like a blank canvas.
This is an even bigger opportunity: to equip sellers with the intelligence to do their job, and give them what a hardcore consultant would. In fact, this is where a partner vendor’s chance to stand out from the crowd becomes absolutely key. The patterns that make sense here are:
Surfacing missed opportunities. Many sellers don’t realize that missing article information is costing them sales, so message them proactively about it.
Answering questions they don’t know are important. Guide sellers toward what we believe they should be doing. If their sales are great, call out what makes them great and what is driving that. Likewise, if sales are low, call out what they could do to improve them.
The right balance between control and automated decisioning is the key
In my view, we’re just scratching the surface of what the ideal partner UX is likely to be. Agents make a huge amount of automation possible.
At the same time, when you’re taking 10,000 actions, it’s hard to communicate them to the seller in a way they can understand. Everyone wants control, especially over high-value actions, or at least enough visibility to shape them. You don’t want the system making decisions that lose tens of thousands of euros without the partner having any control.
But once you bring the human into the loop, you create a winning system that gives sellers real superpowers:
