Executive Summary
Freshworks was accelerating its AI strategy through Freddy AI, investing in machine learning that could help support teams identify frustrated customers, prioritize conversations, and improve response quality at scale.
The Problem Beneath The Problem
The technology already existed. The challenge was that agents didn't trust it. Without trust, AI generated predictions but no business value.
I led the design of a trust framework that helped agents understand, validate, and act on sentiment predictions, while creating feedback loops for model improvement and analytics experiences for support leaders. The result was a scalable AI interaction pattern that expanded across both Freshchat and Freshdesk, connecting machine intelligence, human decision-making, and organizational insight.
Project Snapshot
Company
Freshworks
Products
Freshchat
Freshdesk
Team
Product Manager
ML Engineers
Frontend Engineers
Support Operations
Ownership
01 · The Business Context
AI capability existed. Adoption didn't.
Freshworks serves organizations ranging from small businesses to Fortune 500 companies. As customer conversations increased across chat and ticketing channels, support teams faced a growing challenge: how do agents quickly identify which conversations need attention first?
The company had already invested in sentiment analysis models capable of detecting frustration, urgency, and satisfaction from customer conversations. The expectation was straightforward: Use AI to help agents prioritize work more effectively.
However, adoption remained low. Agents continued relying on manual review rather than AI recommendations.
This wasn't a machine learning problem. It was a trust problem.
02 · Reframing The Problem
From "prioritize faster" to "trust enough to act"
The original request focused on prioritization. Research revealed a different challenge.
After interviewing agents and observing support workflows, a recurring pattern emerged. Agents were not struggling to work quickly, they were struggling to trust AI-generated recommendations.
"I'd rather scan conversations myself than act on something I don't understand."
Support agent · Research interview
This insight fundamentally changed the project.
Original goal
Help agents prioritize faster
Reframed goal
Help agents trust AI enough to incorporate it into their workflow
03 · What was breaking trust?
Four barriers, consistently surfaced
The challenge wasn't prediction accuracy. The challenge was confidence calibration between humans and AI.
Interpretation
Agents couldn't immediately understand what a sentiment prediction meant.
Verification
Agents wanted a way to confirm whether AI was correct.
Control
Agents feared losing decision-making authority.
Flexibility
Different organizations interpreted sentiment differently.
04 · Designing for trust under uncertainty
Designing around uncertainty, not certainty
One of the most interesting constraints was that the machine learning model was still evolving. Accuracy continued improving throughout development. This meant I couldn't design around certainty. I had to design around uncertainty.
Wrong question
How do we expose sentiment predictions?
Right question
How do we help people confidently use imperfect intelligence?
Principles for trust in AI, through design
Instant comprehension
AI should be understood within seconds.
Human control
AI should assist judgment, not replace it.
Transparent feedback
Users should be able to challenge predictions.
Organizational flexibility
Teams should adapt the system to their own workflows.
05 · Evaluating trust mechanisms
How should sentiment be represented?
A critical decision involved determining how sentiment should be represented. Several approaches were explored.
Why emoji won
Agents already think emotionally. When reading a conversation, they don't think "Sentiment score: 0.84." They think "This customer sounds frustrated."
Emoji indicators aligned with existing mental models and required zero translation effort, reducing cognitive load while increasing confidence.
06 · Designing the agent experience
Launching inside Freshchat
The first implementation focused on helping agents quickly identify customer sentiment while managing multiple conversations simultaneously.
Real-time sentiment visibility
Conversation sentiment displayed directly within inbox workflows.
Sentiment-based prioritization
Agents could sort conversations by sentiment and focus on customers requiring immediate attention.
Explainability via hover states
Hover interactions provided additional context without overwhelming the interface — balancing simplicity with transparency.
Beginning vs ending sentiment
Teams should adapt the system to their own workflows.
07 · Designing for AI failure
Acknowledging uncertainty instead of hiding it
One of the most important design decisions involved handling incorrect predictions. Many AI systems attempt to hide failure. I took the opposite approach: when sentiment predictions were incorrect, agents could provide feedback directly within the workflow.
Preserved user trust
Agents retained authority over decisions
Created learning signals
Feedback became input for future refinement
Rather than positioning AI as infallible, the experience acknowledged uncertainty and encouraged collaboration between humans and machines.
08 · Scaling the framework across products
From feature to platform capability
After validating the interaction model within Freshchat, the framework expanded into Freshdesk. Though both products served support teams, their workflows differed significantly.
Shared patterns
Sentiment visibility
Feedback mechanisms
Prioritization logic
Product-specific adaptation
Information hierarchy
Placement of sentiment indicator
Workflow integration
This transformed the solution from a feature into a reusable platform capability.
09 · Beyond agents: Supporting administrators
Trust wasn't only an agent problem
Organizations also needed flexibility. A healthcare provider and a software company often interpret customer frustration differently. I designed administrative controls that allowed organizations to configure sentiment thresholds while maintaining a simple experience for agents — balancing standardization with operational flexibility.
10 · Turning sentiment into organizational intelligence
From "how does this conversation feel?" to organizational insight
I also designed the automation experience so that the conversations gets auto assigned to agent groups from the bots specifically for agents who are specialised for certain operations like escalations
Automation
Agent question
How is this conversation feeling?
Leadership question
What patterns are emerging across thousands of conversations?
I designed a sentiment dashboard and analytics experience that surfaced:
Total conversations analyzed
Sentiment distribution over time
Beginning vs. ending sentiment comparisons
Resolution performance by bots
Positive and negative sentiment averages
Sentiment trends in resolved conversations
Resolution performance by agents
Resolution performance by conversation topics
Dashboard
Business questions supported
Which topics consistently create frustration?
Are conversations ending in a better state than they begin?
How effective are automation workflows?
Where should support leaders focus operational improvements?
Analytics
This transformed sentiment from a workflow feature into a decision-making capability.
11 · Creating a Closed Feedback Ecosystem
The full loop
What began as a sentiment feature evolved into a broader ecosystem connecting users, AI systems, operational insights, and business outcomes.
12 · Impact
From manual triage to trusted AI workflow
85%
First week AI usage adoption by agents
Adoption
40%
Reduction in manual prioritization effort
Operation
3x
Increase in AI-assisted workflows
Trust
5x
Faster identification of high-risk conversations
Efficiency
30%
Feedback loop through overrides
Model improvement
2
Cross product rollout
Platform scale
User impact
Reduced effort to identify high-priority conversations
Increased confidence in AI-assisted prioritization
Improved visibility into customer sentiment
Product impact
Expanded across Freshchat and Freshdesk
Established reusable AI interaction patterns
Introduced feedback mechanisms for model improvement
Organizational impact
Enabled sentiment visibility at operational scale
Helped leaders identify recurring pain points
Connected frontline activity with strategic decisions
12 · Key decisions & tradeoffs
What I chose not to do, and why
Why not automate prioritization completely?
Agents needed visibility and trust before automation could be introduced responsibly.
Why not expose model confidence scores?
Confidence scores increased cognitive effort without increasing decision quality.
Why allow users to challenge AI predictions?
Trust grows when users remain in control. Feedback also created valuable training signals.
Why design analytics alongside agent workflows?
Adoption requires value at multiple organizational levels, not only for frontline users.
14 · What I learned
Designing the relationship, not the indicator
AI adoption is a design challenge first
Even highly accurate systems fail when users don't trust them
Human control increases adoption
Allowing users to validate and challenge predictions strengthened trust rather than weakening automation.
The goal isn't prediction, it's action
The most successful AI experiences aren't the ones with the smartest models. They're the ones people trust enough to use.
The challenge wasn't designing a sentiment indicator. It was designing a relationship between human expertise and machine intelligence, and that relationship ultimately determined whether AI created value at all.
15 · Reflection
This project changed how I think about AI products, moving from interface decisions to designing the relationship between people and predictive systems.