Ask Alfa™ — Driving Adoption by +34%

After the initial rollout, Ask Alfa engagement metrics were below expectations. Users were hesitant to type prompts, unsure what the AI could do, and rarely returned after the first use.

The core tension: Ask Alfa was a powerful tool hidden behind a blank text field. Without knowing what to ask, users couldn't discover what it could do. And without discovering value, they never returned.

Before exploring solutions, I needed to understand why adoption was low despite strong AI capabilities. I combined qualitative and quantitative methods to identify where users struggled.

FullStory Analysis

Session recordings and heatmaps revealed friction and drop-offs. Many users paused before typing, rewrote prompts several times, then left after one query. The issue wasn't lack of interest. It was uncertainty.

The Interviews

The interviews revealed that users weren't intimidated by AI itself. Many were already experimenting with ChatGPT and similar tools. The issue was that users didn't know what datasets the system has access to, and how to prompt effectively

Journey Mapping

Visualized a typical first-time interaction to capture the emotional curve from curiosity to confusion to disengagement.

"Curiosity → Confusion → Frustration → Drop-off"

Key Findings

Research Synthesis

After analyzing interviews, session recordings, and the user journey map, I organized all findings into a synthesis document in Notion and distilled them into three core insight clusters.

[ customer success prompt mind map ]

Around the same time, our Customer Success team built a detailed mind map of effective prompts for every major use case — turning cross-functional insights and research into clear, actionable design ideas.

This cross-functional insight helped shift the framing of the problem: users didn't need more AI capability — they needed a clearer interface for discovering it.

The research surfaced two distinct failure modes, each requiring a different solution.

• Users who never typed a query had a cold-start problem: no mental model of what Ask Alfa could do, and low confidence to begin.
• Users who did type had a recovery problem: vague inputs led to weak outputs, with no visibility into why or how to fix them.

These are structurally different issues. A single solution would have solved one and degraded the other. I designed two complementary systems, each targeting a specific failure point.

SOLUTIONS I EXPLORED

Solution 1: Prompt library

Our Customer Success team had already mapped dozens of effective prompts across different investor workflows. The design challenge was turning that internal knowledge into something usable without overwhelming the interface.

Organizing by user role and workflow rather than topics was a deliberate choice based on interview insights: portfolio managers and analysts approached the tool with different goals and starting points, so a generic list felt irrelevant to both.

Key Design Decisions

1. What to show and how much

The first version displayed three large prompt templates as full-width cards. Testing revealed two problems: the widgets felt too big and pushed the input field out of focus, and the "More suggestions" CTA redirected users to the Discover page — breaking their flow entirely by taking them out of the chat context.

We reduced the size of prompt chips and kept all suggestions inline. Users needed just enough to spark an idea, with an easy path to explore more without leaving the conversation.

Large templates redirected users out of context. Smaller inline chips kept the flow intact.

2. How to organize the prompts

An early approach grouped prompts by topic (e.g., "Earnings", "Macro", "Risk"). In testing, neither portfolio managers nor analysts found the categories immediately useful — they were scanning for their workflow, not a subject area.

We reorganized around user roles and common workflows. Portfolio managers and analysts now see examples that match how they actually think about their work, not how the data is structured internally.

Solution 2: Prompt Improver

The goal was to reduce failed or low-quality AI outputs without making users feel at fault. Instead of treating issues as errors, the system detects ambiguity in real time and suggests a stronger prompt before the user hits a dead end. This required close collaboration with ML to understand what "ambiguity" looks like at the model level and when it can be reliably detected. The feature is proactive, not corrective.

Key Interaction Decisions

1. Timing of the suggestion

Early testing showed a clear problem: when suggestions appeared instantly while typing, users who already knew what they wanted found them disruptive. The suggestion interrupted their thinking before they'd finished forming a query.

We introduced a ~5-second delay after typing stops. A pause is a stronger signal of friction than active input, indicating the user may be stuck. Suggestions now appear when support is more likely to be needed, not while the user is still typing.

Before

Suggestion fires mid-thought — disrupts users who already know what they want

After

Suggestion appears only when the user pauses — a signal they may be stuck

2. Explicit "Accept suggestion" CTA

Auto-applying the improved prompt tested poorly. It felt intrusive and reduced confidence, especially for professional investors making high-stakes queries who need to feel in control of what gets submitted.

We replaced it with an explicit CTA. This made the system's intent clear and gives users a moment to review before committing. The decision prioritizes control and trust over automation.

After finalizing the designs, I presented the Prompt Library and Prompt Improver to all stakeholders — Product, Engineering, and Customer Success. I framed the presentation around the user journey map rather than the features themselves, which helped non-design stakeholders understand why the solutions worked, not just what they looked like.

Once approved, I handed off Figma files with annotated interactions, edge cases, and prompt logic documented directly in the file — not in a separate spec doc. Edge cases mattered especially here: what happens when the Prompt Library has no relevant suggestions? What does the Improver show for very short queries? Every scenario was accounted for before handoff.

Beyond the numbers: the Prompt Library was adopted internally by the Customer Success team for onboarding new clients — something unplanned, but a strong signal that it clarified Ask Alfa's value even outside the product.

This project reinforced something specific to AI product design: the interface has to do the job the AI can't do itself — which is explain its own capabilities in context.

The biggest lesson wasn't about UI patterns. It was about trust as a design material. Users weren't evaluating features — they were deciding whether to believe the product would understand them. Every decision, from how we surfaced prompts to how we framed the Improver CTA, was really a decision about how much trust to ask for, and when.

Lessons learned

• Embedded guidance beats tutorials — contextual help at the moment of action is more effective than upfront explanation
• Showing beats telling — demonstrating capability through examples was more persuasive than any onboarding copy
• Agency is a trust mechanism — in AI interfaces, giving users explicit control over suggestions reduces friction more than automation does
• What I'd do differently: Push earlier for a lightweight response feedback loop ("Was this helpful?"). We discussed it but it didn't make the sprint. That loop would have made the Prompt Library smarter over time.

Next steps

• Introduce analytics-driven, personalized prompt suggestions based on user role and past queries
• A/B test more contextual prompt hints triggered by session behavior (e.g., long pause = show suggestion)
• Expand Prompt Library templates based on usage patterns and CS onboarding feedback

After updating the design system, we simplified the UI and removed the emojis. Here is how the initial "new chat" screen looks now.

New chat screen — updated UI