Customer success was drowning in the same repetitive questions (“Where’s the pricing tool?”), delaying real issues. Guides existed—but no one read them. We needed instant answers without overwhelming users or breaking their workflow.

Challenge

Objectives

Design an AI assistant to autonomously handle 50% of repetitive CRM queries within 3 months, cutting user wait times from 48 hours to under 5 minutes while maintaining 80% satisfaction—freeing customer success teams for strategic work.

Learning from AI’s Wild West

I dissected 4 CRM AI agents to steal their best moves:

1. HubSpot Bot

2. Zendesk Answer Bot

3. Intercom

4. Freshsales Freddy

Research

The AI Rules We Defined

1. Role: A smart helper (not a bossy automator).

   • Example: Pulse answers, “Which deals are at risk this quarter?” but won’t auto-send contracts.

2. Control: Users steer every interaction.

   • Like Google Docs’ Smart Compose—suggest, don’t dictate.

3. Failure Plan: Always include an “Escalate to humans” button.

   • Because even AI has bad days.

Users don’t want ‘AI’—they want answers without the wait.

Why Pulse Needed to Exist

We implemented AI to solve two critical problems:

Ideation

Information Architecture: Scoping Pulse’s Role

We defined Pulse as a beginner’s guide + virtual CS teammate, focusing on:

AI should act as a force multiplier—handling routine work, so humans can tackle what truly requires judgment.

Building Trust Through Transparency

To ensure users understood and controlled Pulse:

1. Explainability:

   • Clear reasoning: “Recommended template based on: Client tier (Enterprise), past deal terms.”

2. User Agency:

   • Edit/reject outputs (e.g., modify suggested email templates).

   • Report inaccuracies via “Teach Pulse” button.

3. Graceful Failure:

   • When unsure, Pulse responds: “I can’t find that—try rephrasing or [contact support].”

• ❓ Answering platform questions (e.g., “How do I merge contacts?”)

• 🔗 Directing to tools (e.g., “Take me to the bulk email editor”)

• 📊 Recommending templates (e.g., “Use Enterprise upsell template”)

• 🎯 Suggesting follow-ups (e.g., “This lead hasn’t responded in 14 days”)

• 🛠️ Proposing client questions (e.g., “Ask about budget timelines”)

Pulse should feel like asking a knowledgeable colleague—fast, accurate, but never overbearing.

Trust isn’t given—it’s earned through consistency and transparency.

Prototyping

The Challenge

1. Pulse needed a home, but every option had tradeoffs:

   1. Floating Bottom-Right

      • Problem: Hid critical CRM table data (e.g., client contract dates).

      • Tech Limit: Design system couldn’t make it “dodge” content.

   2. Lateral Menu

      • Risk: Cluttered core navigation (reserved for daily-use tools like Deals or Contacts).

   3. Top Support Menu

      • Pros: Aligned with help/notification buttons (auxiliary actions).

      • Cons: Future button sprawl potential.

   Final Solution:

   • We placed Pulse in the top menu because:

     • Mental Model Fit: Users expect “help” features there (like notifications).

     • Scalability: Future buttons could be grouped under a dropdown.

     • Safety Net: Didn’t compete with primary tools in the lateral menu.

2. We also needed to choose how to introduce users in the chat with Pulse. These were the first options:

   1. Button Overload: Separate buttons for each action created maze-like navigation.

   2. Pre-Guide Workflow: Forcing users to categorize queries (“Is this about tools, data, or templates?”) added friction.

   Final Solution:

   1. Single Input Field with example prompts (“Try: ‘Show me Acme’s overdue invoices’”)

   2. No UI Clutter: Let users ask naturally, like messaging a colleague.

The Problem

• Sales teams wasted hours searching for client data

• Customer support drowned in repetitive platform questions

Our Mission
Create an AI assistant to:
✔ Resolve 50% of routine queries instantly
✔ Cut response times from 48hr → 5min
✔ Maintain 80%+ satisfaction via transparency

Key Breakthroughs

1. Seamless Integration

   • Top-menu placement matching user expectations

2. Frictionless Interaction

   • Single text input (no complex menus)

3. Built-in Trust

   • Clear explanations for every suggestion

   • Easy override options

Next steps

1. Controlled Testing Phase

   • Conduct Wizard of Oz testing with 10 enterprise users to identify edge cases.

   • Refine conversation flows for ambiguous queries (e.g., vague or multi-intent requests).

2. Responsible Scaling Preparation

   • Establish baseline metrics for:

     • Suggestion accuracy

     • User override rates

     • Support ticket deflection

   • Develop quarterly bias audit protocols with the data science team.

3. Phased Production Rollout

   • Pilot with 5 strategic client accounts.

   • Monitor real-world usage patterns before enterprise-wide deployment.

Conclusions

Key Learnings: Designing AI for the Real World

1. AI Design Demands Precision

   Creating effective AI tools requires:

   1. Clear boundaries: Explicitly defining capabilities (e.g., Pulse retrieves data but never auto-sends emails).

   2. Failure protocols: Built-in escalation paths (“Contact Support” buttons) and explanation systems (“Why this suggestion?”).

   3. Continuous validation: Quarterly bias audits and accuracy monitoring to sustain trust.

2. Interfaces Must Endure

   As a designer, you architect for:

   1. Current needs: Solving today’s user pain points (e.g., slow query resolution).

   2. Future scalability: Pulse’s top-menu placement anticipates new auxiliary tools without cluttering core workflows.

   3. Technical constraints: Working within CRM UI limitations (e.g., no floating buttons) forced simpler, more resilient solutions.

What I learned

See more proyects!