Agentic Data Discovery

Overview

 

Agentic Data Discovery reimagines how agentic AI integrates into an existing platform UI. Rather than relying solely on conversation threads, the design brings structure, persistence, and navigability to the agentic experience so analysts can focus on finding insights rather than managing noise. The MVP is launched to 30 users as a feature flag, with user testing validating that structured UI significantly outperforms pure chat for goal-oriented data discovery tasks.

 

 

Problem

 

Our engineering team had built MCP-powered agentic data discovery tools that were gaining real traction. But momentum created a new problem: there was no clear path to integrating these tools into our existing UI. Bolting AI onto an established product risked confusing users, disrupting familiar workflows, and undermining trust in features they already depended on.

 

 

A design existential question loomed: Is UI dead?

 

 

Deep down, I wanted to believe that UI is not dead and that there is still value in showing the UI.

 

 

 

Research

 

I conducted 9 follow-me-home research with users across data scientists, product managers, and analysts, observing them as they used MCP data discovery tools in their actual work environments. 

 

 

I identified three specific failure modes before they became a critical bottleneck:

 

1.  

1. 1. One-off responses. Conversation threads produce impromptu, disposable answers rather than persistent, reusable artifacts.
   2. Verbosity and noise. Long, text-heavy responses caused users to skim and miss critical information.
   3. Buried artifacts. In non-linear workflows where users need to refer back to something, artifacts get lost in walls of text.

 

As one user put it: “If it outputs this massive message, it’s like overload. I’m looking for very specific pieces of information. In the UI, I know exactly where I can find the information.”

 

 

Watching users in context revealed a fundamental mismatch: data discovery is a non-linear, goal-oriented task, but conversation threads are built for linear, open-ended dialogue. The tools users were adopting didn’t match the way they actually worked.

 

 

Ideation

 

The design challenge I faced was integrating the goodness of agentic AI experiences into the existing data platform UI. But how do I do that seamlessly without confusing the user and overriding existing features?

 

1. How to differentiate Search from Agentic Input?

DDE already had a Search function. The agentic input looked visually similar but functioned completely differently, creating a dangerous source of confusion when trying to introduce users to a new behavior. 

 

I looked at popular or similar products that analysts used with Search functions and agentic experiences to see how they did it: Google Drive, Databricks, Cursor.

 

 

 

 

 

 

I noticed that these products had Search at the top, while the agentic chat was open on the right as a side panel, so I designed similarly for DDE and tested it out with users.

 

 

When I tested this with users, they were paralyzed on which direction they wanted to go: Search, or agentic AI? With the agentic chat showing in the UI, users questioned how much they should trust the agentic AI. They felt that it was just another AI feature gimmick that functions below expectations, so maybe they should just continue with Search instead?

 

I decided that showing two open text fields is too confusing and looked at other search engine products with agentic AI built in: DuckDuckGo, Yahoo Search, and Google Search.

  

 

 

 

I first tried placing Search and the agentic input side by side, giving users a clear choice. But in testing, users automatically reached for the Search bar every time, bypassing the agentic input entirely out of habit. I tried keeping both inputs separate but equally visible and got the same result.

 

 

 

The insight was that familiarity was working against adoption. To introduce users to the agentic experience, I needed to shift the visual hierarchy. I minimized Search to a button icon and made the agentic input the dominant element on the page. Users defaulted to the agentic input while retaining the option to use explicit search when needed.

 

 

 

 

When I tested this with users, their eyes naturally scanned from top to bottom. They still saw the plain search button but saw the open text field for the agentic experience. Having the agentic input as front and center gained user’s trust that the AI was stronger than search. 

 

 

2. Where to place AI responses?

The second challenge I ran into is what happens after the user clicks on “ask AI”? Do the AI responses show at the same time as search results? Should they show separately? Should it be a side panel or immersive?

 

 

 

I explored layouts from the following reference points: Google Search’s AI Overview, Amazon Rufus, and Claude Desktop. Google Search had their AI overview at the top of their search results and users can view the AI experience immersively under the AI Mode tab. Amazon Rufus had a side panel, while Claude Desktop had an immersive whole page experience.

 

 

 

 

 

Google’s approach placed an AI summary at the top of search results and tested poorly. Users couldn’t distinguish it from regular search results, leaving them confused about what they were looking at.

 

 

With Amazon Rufus layout, users said that this looked too cluttered with the UI content on the main page. 

 

 

They actually preferred the immersive chatbot experience better because it didn’t disrupt their workflow and looked cleaner.

 

 

 

This informed a layout where the agentic chat opens in an immersive view, with the option to collapse into a side panel. Users can control how much space the AI occupies depending on their task.

 

 

3. How to show artifacts without disrupting flow?

For data discovery tasks, users frequently need to act on artifacts mid-conversation, such as accessing a table, reviewing a query, or submitting an access request. The question was where those artifacts should live. I saw in products like Databricks, Palantir, and Qliksense that they showed artifacts as in-line objects within the chat, so I incorporated this pattern into my designs.

 

 

 

 

 

I also noticed instances where the chat panel existed next to the main UI workspace, like Google Drive:

 

 

 

 

I noticed that the AI is in a side panel when the user can ask further questions or want to change something in the UI, so I mocked follow-up questions that the user can ask based on the content of the UI. 

 

 

 

Some users found that viewing the UI content by the chatbot panel helpful, like query insights or the table details page

 

 

 

 

 

 

So I added a secondary button labeled “Open side-by-side” so that users have the option to view the content next to the chatbot if they wanted to.

 

 

 

I designed inline containers for each artifact type, embedded directly in the chat thread. When tested, users consistently preferred opening artifacts in a new tab rather than as a side panel within the interface. They didn’t want to lose their place in the agentic workflow. A new tab let them act on the artifact and return to their thread without disruption.

 

 

Minimum viable product

 

After testing these designs with users, I walked through the design prototype with engineering to see what can get implemented in the first release. 

 

At the minimum, I wanted users to be able to:

1. Type and receive an agentic response
2. Click open an artifact in a new tab 
3. Have multiple conversations and threads

 

I prioritized features that can be built using components established by the Intuit Design System and GenUX team, so luckily they already covered conversation threads and responsive chatbot with immersive and side panel views. This made it easier for the engineering team to implement and not have to build components from scratch.

 

I felt that the in-line artifacts can be added on in a future release since these will require more engineering time and effort. They were more of a nice-to-have at this point rather than a must have because users can see the artifacts in a new tab.

 

Final Designs

 

The alpha version launched in the E2E dev environment with four core features implemented:

• • Minimized search button icon to elevate agentic input

• • Prominent agentic chat input as the default entry point

• • Immersive chat view with side panel option

• • Persistent previous thread history
    
    

 

Results

 

More than 30 users were added to the feature flag to test the alpha and report bugs.

 

User responses validated the core premise that a structured UI serves data discovery better than a pure conversation interface:

“When I go to DDE, I have a very specific goal in my mind. I need to find which table exactly to achieve this goal.”

“The thing I like about data discovery is that there’s a UI. I know where certain information is going to be laid out for me.”

“I didn’t even know that I could look at other people’s queries and how they’ve interacted with tables before. So that’s something that’s extremely useful on data discovery that maybe I don’t necessarily have through the Cursor MCP as of yet.”

 

Reflections

 

 

These novel interaction design patterns with AI really challenged me to move out of my comfort zone because no other product has them totally figured out. 

 

 

I realized that other products are also experimenting like me, and constantly changing design patterns based on user feedback. For example, Amazon Rufus has been discontinued and Google Drive doesn’t open their AI chatbot in a side panel on the homepage anymore (it’s immersive now).

 

 

 

 

 

Seeing the evolution of these products and design patterns make me feel that my findings are validated, but I also remind myself that users are very complex and we won’t fully know until the product is built and being used. 

 

My main takeaway from this project is that despite uncertainty, I should not be afraid of designing novel paradigms. I can rely on user testing and competitive research to make informed design decisions, and it’s okay to release features with risk and get live feedback later. 

 

I’m looking forward to see how the 30 users with access to the feature flag can use the new agentic experience within DDE, and I can’t wait to have it expand to other workflows within the product.