No UI is bad UX

When creating AI products, don’t be like Alexa. No UI is bad UX.

When Amazon launched Alexa¹ , it seemed like a breakthrough. Today, it’s basically an overpriced clock radio that has a kitchen timer and can read you the weather. I think a major reason why is that it has no UI, so people don’t know what they can do with it. Instead, users stick to the basics that they know will work.

Chatbots have the same problem: Users stare at a blank chat and have no idea how best to achieve their goal² .

“What can it do? What should I say? How do I even get started?”

Here’s some tips that can help:

1. Hide your prompt engineering behind a button³. Traditional UX is so great for communicating what can be done and what will happen. Tapping the CTA launches chat, but it’s chat with a purpose.

2. Give people options in chat. “Would you like to learn more about X or Y?” demonstrates subject matter expertise, provides users inspiration and education about what is possible, and helps avoid UX dead ends.

3. Use traditional UX, powered by agents. Search is a great example here — search UX has been around for 25+ years, but most search results are frustratingly bad. A bunch of agents orchestrated together can massively improve the quality of search while still offering a familiar and high converting UX⁴ .

There’s a reason why, 13 years after the Tesla Model S launched with an interior dashboard that was entirely a giant screen with no physical buttons, buttons are coming back. The point of great UX is to help users progress towards their goal as quickly as possible. Just because your UI could be chat-only doesn’t mean it should be.

The Workshop

This is a newsletter-only section where I share a half-baked idea in hopes that y’all who are smarter than me can work it out with me.

We’re going to talk a lot more in the future about MCP servers and the power of giving agents access to documents, so instead I wanted to share two new product announcements by friends of mine that I thought are spot on.

1. Jasper MCP Server. This is such a perfect use case for MCP — the subject matter experts (e.g. legal, brand marketing) within your company maintain, in Jasper, documentation that outline all of your guidelines, compliance policies, and style guides. Your content creation teams continue to use whatever tools they want; AI agents can pull down from Jasper all of the guidelines, run the content through, and evaluate (or rewrite) accordingly. The abstraction here allows both teams to focus on what they do best, and MCP is the interconnect that enables one to impact the other without forcing the adoption of specific tools and specific workflows.

2. Gather Grapevine. The big unlock over the last 6 months with agents was getting out of trying to shove everything into a single prompt’s context window and instead providing agents with access to documents that provide the necessary context. This just launched yesterday so I haven’t had the chance to try it myself, but I’m excited by the idea that Grapevine does the work of assimilating all the knowledge across the company into a single source, and then that becoming one of the tools other agents can call upon to do their job.

1  Ten years ago! That’s crazy how time flies.

2  Some users will say nothing or very little, and some users will assume it can do everything. Which is actually worse than Alexa in some ways, as LLMs will always respond (with confidence!) regardless of the quality of its answer.

3  This layer of abstraction has the benefit of allowing you to iterate on your prompts independent of the UX.

4  Multi-agent orchestration is too big of a topic to get into today. But suffice to say, there’s a way to break down the search problem into component “jobs” and give a different AI agent each one of those “jobs”. Make sure to solve for structuring the output data in a way that can still offer users filtering controls. It’s a way to give users way more degrees of search freedom than they’ve ever had before.