Natural language data interaction in a transparent and verifiable way.
Role
Lead Product Designer
Time Frame
2 wk to MVP + 2 mo to launch + ongoing
Context
CB Insights provides private company data and research to enterprise clients. Leadership wanted to establish the company as a frontrunner in generative AI by launching a chatbot capable of answering client questions using our data.
A previous chatbot attempt had failed due to poor accuracy and unclear product-market fit.
Users needed fast, credible answers, but generative AI hallucinations made trust a core design challenge.
Leadership insisted on a chatbot, but our previous research showed that AI chat lacked usability and created unrealistic expectations.
As the lead Product Designer was to design an AI experience that clients would trust and adopt, balancing business goals with usability and credibility.
Problem Statement
How do you make generative AI credible and usable for enterprise users who depend on accurate data to make high-stakes decisions?
The main design tensions were:
Trust: Clients compare any chatbot to ChatGPT but expect zero hallucinations in a B2B context.
Cognitive load: Writing good prompts is hard. Many users second-guess themselves before even hitting enter.
Speed: LLM latency often created 20–30 second delays, a lot longer than users were used to tolerating.
Research
We designed a two-track experiment:
Chatbot prototype: an out-of-the-box LLM with source annotations.
Integrated AI search answers: contextual results embedded directly into search.
We recruited 10 clients for a two-week test, split between chat and search. Half beginning with the search experience and half with the chatbot, then the other way around.
I facilitated interviews to dig into the nuances of the two tools. Collecting their first impressions, questions and feedback on the tool.
The process also gave us a treasure trove of user questions, that allowed us to dig deeper into the user intent when using our chat experience. In order to get a sense of how this fit into our workflows, I mapped each question into our broader jobs-to-be-done framework. That would help us get a picture of what kind of use cases we could focus on and also give leadership a sense of the breadth of our scope.
It also gave engineering a chance to assess what types of questions would be feasible to answer.
Key Findings
Users trusted answers more when sources were shown even if they didn’t actually click them to verify information.
Chat lowered friction for complex queries (e.g., “What companies has a16z invested in that are based in Tel Aviv?”).
Expectations were very high: users wanted personalization, industry-specific acronyms, and features on par with ChatGPT.
Our integrated search went completely unnoticed. We had to point it out to users that it existed.
Given the results and new pressure from above, we were now asked to launch a chatbot in one month.
Key Design Solutions
Balancing Visual Design & Brand Perception
Initially, the chatbot was designed in light mode. Leadership rejected it as too “cute” and pushed for a Bloomberg-like, professional aesthetic. I pivoted to dark mode minimalism with a monochrome palette, making answers and actions the visual focal points.
Later, as clients requested light mode, I extended the design system to support both, creating a foundation for future platform-wide theming.
Establishing Credibility Through Sources
Sources were the core part of establishing ChatCBI's credibility. As an expert voice in the technology space and a research and data provider, we needed the chatbot to substantiate its claims. In addition, it’s opinions had to be in line with our research.
I first had to take stock of all the sources we were ingesting to start and which we would need to add in the future. From there, I had to create a consistent labeling approach.
That meant both creating a source labeling system and a way to visually communicate credibility at a glance:
Blue for CB Insights’ vetted data/research.
Gray for external sources.
Assigned icons for CB Insights content, internal analyst notes, genAI insights, and external content.
This system helped users immediately see whether an answer reflected trusted CBI research or outside data without needing to read footnotes.
The system also extended into the answer itself. Communicating immediately, which specific parts of the answer was CB Insights sourced. Including a hover state to dig deeper.
Suggested Questions
In order to help users get ideas for what to ask ChatCBI next, we had the LLM generate follow-up questions for them. Testing showed that users appreciated the ideas and it increased their engagement with the app.
It had some downsides though. Users looked to those questions for guidance on how to craft the better prompts, which was misleading. Sometimes suggestions were outside the scope of CB Insights’ data, leading to questions we were less capable of answering. They also weren’t anchored to the best way to prompt an LLM, they simply generated what the LLM thought would be relevant to ask.
We decided to keep the questions though, as they were a net benefit to the user since our sourcing clearly indicated an area that was inside our expertise.
Easing Cognitive Load with Curation
Prompting was one of the heaviest cognitive load on the user. Since chatbots are inherently open-ended, they are problematic for a data company like CB Insights. Users can ask about literally anything and receive an answer. So what should they ask? What does CB Insights cover? How should they phrase that question for the best answer?
An LLM is also trained on the knowledge of the entire internet. While CB Insights has a limited universe of data. We needed the chatbot to embody CB Insights and research in an accurate way. Again, since the chat is open-ended, we needed affordance to let the user know what to to ask.
For initial release, we launch with a set of rotating static questions that the user could pick from. They provided some inspiration, but quickly became stale and were not directly applicable to most users.
Inspired by pi.ai, I designed a two-tier system:
High-level use cases (e.g., company screening, market analysis).
Personalized prompts based on each client’s past interactions.
I'm writing a separate case study on how we approached the project.
Handling Latency with Transparency
Early versions took up to 30 seconds to generate answers. The reason was that, unlike ChatGPT, we had to go through a series of complex tasks to reference our data and insert it into the context window of the LLM.Instead of hiding this, I worked with engineering to design a progressive status system that explained what was happening.
Instead of hiding this, I worked with engineering to design a progressive status system that have an indication of what was happening.
Waiting is never ideal, but it did reduce frustration somewhat. For some users it actually built confidence. One user said:
"
I’d rather wait twice as long for a great answer than not wait and get a mediocre one.
Iterating Toward True Chat
Due to technical constraints, the alpha shipped as a single-answer experience. Testing showed users were frustrated they couldn’t “dig deeper.”
Introducing Ask ChatCBI
Next, leadership wanted to extend ChatCBI into the broader platform. Letting users ask questions directly from reports, insights, and feeds.
Design challenges:
Each content type had different layouts and sourcing conventions.
We needed a consistent pattern that mixed well into the rest of the platform.
I explored multiple UI patterns (expanders, headers, inline embeds) and placements. Ultimately, leadership chose a simple card expander pattern. We also prefilled a suggested question in order to draw the users interest
In addition, we had to design for a new attachment type that would signal the content that was being referenced. Which we also designed to extend to later functionality like file attachments and attaching research content.
Adoption was low, and the few that did use it didn't stay long enough for an answer to finish loading. Most likely, users didn’t want to navigate away mid-task.It's also possible that the option didn't interest them too much in the first place.
Result
By 2025, ChatCBI had become one of the most impactful features on the platform:
40% adoption rate among active users.
Accounts with active ChatCBI users were 2× more likely to renew subscriptions.
Retention increased to ~50% among users who adopted chat.
We also learned that:
The chat experience cannibalizes all other experiences. With the rapid ascent of ChatCBI other features started declining in usage. Users found it easier to navigate our data from one central location even if it was LLM generated.
Clients used ChatCBI primarily to generate reports, screen companies, and explore new technologies.
Even when answers weren’t perfect, users valued the sources as entry points for deeper research.
One enthusiastic client even created a video praising our LLM rollout underscoring the credibility and excitement we had built.
Reflection
Designing ChatCBI was a lesson in balancing hype vs. reality in enterprise AI:
Natural language interaction is simplifies the overall experience. A majority of our users have turned to ChatCBI when exploring the platform. It allows them to bypass complex filtering and have their results at the press of a button.
Trust is everything. Sources mattered more than speed or polish in driving adoption.
Guidance reduces friction. Suggested prompts were essential for lowering the barrier to entry.
Not every request should be built. “Ask ChatCBI” showed that bolting chat everywhere doesn’t create value.