Your AI Agent Still Needs a UI

There’s a question we’re hearing more often from clients building AI-powered applications: if users can interact with an AI agent through Claude or ChatGPT, do we still need to build a front end?

It’s a reasonable question. If an agent can query databases, retrieve information and take action through connections to external systems using protocols such as Model Context Protocol (MCP), why spend time and money building a dedicated product interface? Why not let users work with the product entirely through conversation?

The answer comes down to something that’s easy to overlook: a user interface doesn’t just help people use software. It helps them understand what the software can do.

You Can Only Ask for What You Know Is Possible

Think about the software you use every day. Perhaps it’s your CRM, project management platform or analytics tool. You know what to ask for because you’ve spent time using it. You’ve explored the interface, discovered the available reports, learnt which metrics you can track and developed a feel for how the system works.

Now imagine opening a blank chat window connected to that same software, with no prior experience of the product. What would you ask it to do? You might start with the obvious things, but how would you know which reports were available, which workflows could be automated or which less obvious capabilities might be useful? The agent could be perfectly capable of doing all of it, but you wouldn’t know to ask.

Of course, an AI assistant can explain its capabilities, suggest prompts and guide users through a conversation. But that relies on people knowing what to ask, or the agent knowing when and how to introduce useful options. A visual interface can make those capabilities visible from the outset, letting people explore without needing to know exactly what they’re looking for.

This matters particularly when you’re building a new product. A well-designed interface introduces users to what’s available, surfaces useful actions and shows how different parts of a system fit together. As we’ve explored in The Validation Surface, understanding how people interact with a product is central to making the right design decisions. Replacing a front end with a conversational interface doesn’t remove the problem of discovery and onboarding. It changes how you solve it.

Not Everything Works Better Through Conversation

Discoverability is only part of the argument. There are also situations where a visual interface is simply a more effective way to work with information.

Consider an analytics platform. An agent might be excellent at investigating a trend, comparing performance across periods or explaining why a metric has changed. But users still want to explore dashboards, examine charts and spot patterns they wouldn’t have thought to ask about.

The same applies when an agent takes action. Imagine asking an assistant to update hundreds of customer records or prepare a financial forecast. The agent might be fully capable of doing it, but people may want to review the proposed changes, adjust assumptions and approve the outcome before anything happens.

These aren’t limits on what the agent can do. They’re about how people understand information, make decisions and stay in control. A dedicated interface can make complex information easier to explore and important actions easier to review. Conversation makes it easier to request work and move between tasks without following predefined steps. The two can complement each other.

An obvious response is that the interface can simply live inside the assistant, with interactive screens rendered within the chat itself. That’s a real option, and a promising one. But it doesn’t remove the design work. Someone still has to decide what those screens show, how people review and approve actions, and how users discover what’s possible in the first place. The question isn’t whether an interface is needed, but what form it should take and where it belongs.

The Case for Offering Both

None of this means every AI-powered application needs a traditional front end. There are clear advantages to working with software through Claude, ChatGPT or another assistant. Users can work across multiple systems through a familiar conversational experience, ask questions in natural language and complete tasks without moving between applications. For narrowly defined agents performing straightforward tasks, that may be all the interface they need.

But there’s a difference between making a product accessible through an AI assistant and assuming the assistant should be the only way to reach it. The two can serve different purposes, even when they’re connected to exactly the same underlying systems.

That isn’t just theory. In a recent conversation with a company developing AI-powered software, we heard that usage of its conversational AI features was split roughly equally between external AI assistants and its own product interface. That’s one company’s experience rather than a universal rule, but it suggests people value having different ways to interact with the same product.

The right approach will depend on what the product does, who uses it and what they’re trying to achieve. Some experiences will be better suited to conversation, others to visual interaction, and many will benefit from a combination of the two. It’s an important consideration when developing AI-native products, where the experience should be designed around what AI makes possible rather than simply adding AI functionality to an existing application.

Designing for How People Actually Work

As AI agents become more capable, it’s tempting to think less about the interface and more about what the technology can do. But capability alone doesn’t determine whether a product is useful.

Good AI-native product design starts with understanding what people need to achieve, how they work and what information they need to make decisions. That should shape where AI fits, how users access it and what kind of interface supports the experience.

For some products, that might mean building a conversational experience from the ground up. For others, it could mean adding an AI assistant to an existing interface or allowing users to access functionality through external assistants as well. As we discussed in Adding AI Features Doesn’t Make a Product AI-Native, the distinction is about how AI shapes the product itself, not simply whether AI functionality is present.

The question isn’t whether AI agents will replace interfaces. It’s which interactions benefit from conversation, which need a visual experience and how the two should work together.

The best products won’t necessarily be those with the most capable agents. They’ll be the ones that make those capabilities useful, understandable and accessible to the people who need them.