Gabriel Delman | Projects

Context

Customer support agents often operate under intense time pressure while handling complex and emotionally charged conversations. During live calls, agents are expected to listen carefully, understand the customer’s issue, search for historical data, interpret policies, and provide accurate solutions, often while switching between multiple systems.

This fragmented workflow creates several challenges:

High cognitive load during live interactions
Time wasted searching across tools for customer, order, and payment data
Inconsistent resolutions due to unclear or hard-to-access policies
Longer handling times and higher re-contact rates

As support volumes grow, these inefficiencies directly impact both customer satisfaction and operational costs.

Competitor's Analysis

To ground the solution in real-world patterns, I reviewed how modern support tools and AI-first assistants handle live interactions, context retrieval, and agent guidance.

The goal wasn’t to copy interfaces, but to identify interaction standards, failure modes, and trust patterns that repeatedly show up in high-stakes support environments.

The calculation process

The procedure's registration main feature of the AnestPlan app was first designed following the specifiations required to automatically calculate the proceedure's final price, such as time and proceedure's duration, age, gender and patient's weight, numbers and quantity of anesthesias’ codes and many others.

To calculate the price of an anesthesia procedure, it is necessary to follow these steps:

Product Hypothesis

If we develop a mobile-first support app that enhances live calls with AI-powered tools, including pre-call customer insights, real-time transcription, contextual smart suggestions, and automated follow-up options, then we can reduce average call duration by 20 - 30%, improve follow-up consistency by 40%, and increase customer satisfaction by 15 - 25%.

experiment

expected outcomes

1

Code consulting
‍They may vary according to the anesthetic classification, from 0 to 8 – the higher the number, the more expensive

2

Consider 100% of the value of the anesthesia with the highest classification
‍The values of the other anesthesias will be added: - 50% if performed at the same incision site; - 70% if performed at a different incision site

3

Double the total amount
In the case of a child weighing less than 2.5kg or if in a private room

4

Add 30% to the total amount
If the anesthesia classification is 7 or 8;
In the case of extracorporeal circulation;
In neonatal surgery;
In the case of gastroplasty for morbid obesity;
For surgeries lasting more than 6 hours;
For emergency surgeries between 7 PM and 7 AM from Monday to Saturday or at any time on Sundays and holidays.

Product Goals

This project explores how an AI-assisted support agent could improve the quality and speed of customer service while preserving human judgment and accountability.

Primary goals:

Reduce time to resolution (TTR) and average handle time (AHT)
Improve first-contact resolution
Lower cognitive load for support agents
Increase consistency and policy compliance across interactions

🔎 Focus
No-code chatbot builder

‍💪🏼 Strengths 
Fast to deploy, solid NLP

‍🎯 Opportunities 
Weak in telephony & real-time calls

🔎 Focus
Voice AI for contact centers

‍💪🏼 Strengths 
Human-like voice, handles basic issues

‍🎯 Opportunities 
Limited customization for small and medium businesses

🔎 Focus
Generative AI + automation rules

‍💪🏼 Strengths 
Powerful hybrid AI approach

‍🎯 Opportunities 
High technical onboarding curve

✅ Identified gap: few solutions provide real-time, AI-assisted phone call management with actionable insights, follow-up automation, and seamless integration with task management tools.

Solution Overview

The proposed solution is a real-time customer support workspace that consolidates conversation, context, and decision-making into a single interface.

The application listens to live calls, transcribes the conversation in real time, and uses AI to suggest relevant actions, responses, and procedures, all while keeping the agent firmly in control.

Key feautres:

Live transcription of customer conversations
Automatic retrieval of customer history, orders, payments, and refunds
AI-generated suggestions for responses and next-best actions
Clear confidence cues and references to previous customer's orders
Human-in-the-loop decision making at all times

The goal is not to replace the agent, but to augment their decision-making and reduce unnecessary friction during live support sessions.

Claude: problem framing & ideation

I used Claude during the early exploration phase to help structure the problem space and challenge my initial assumptions.

Because of its strength in long-form reasoning and textual depth, Claude supported higher-level thinking rather than UI generation.

How Claude was used:

Structuring and clarifying an open-ended problem space
Challenging pre-conceived ideas and early solution bias
Generating alternative product approaches and mental models
Organizing complex thoughts into clearer design hypotheses

Claude functioned as a thinking partner, helping expand the solution space before converging on a final direction.

Lovable: rapid exploration & wireframing

Lovable was used as a practical tool to accelerate visual exploration once the problem space was defined.

It enabled quick iteration across multiple layout and structural directions without committing too early to a single solution.

How Lovable was used:

Generating early wireframe concepts and layout variations
Exploring different information hierarchies for a support console
Accelerating low-fidelity exploration to save design time

Lovable allowed me to move faster through early-stage exploration, freeing up time to focus on interaction details, edge cases, and product-level decisions.

👨🏻‍💻Using IA in my Design Workflow: In addition to designing an AI-powered product, I intentionally integrated AI into my own design process, not as a replacement for design thinking, but as an accelerator for exploration and iteration.

💡 While AI supported exploration, all critical decisions and refinements were done manually, such as:

Final information architecture and interaction model
UX writing optimized for live, high-pressure conversations
Visual hierarchy, consistency, and system coherence

By combining AI-assisted exploration with hands-on refinement, I was able to focus more deeply on usability, responsibility, user feedback and real-world constraints.

Key Screens & UX Decisions

Pre-Call

Upcoming interaction preview: agent sees the next call with relevant details like customer name, issue category, and ticket/order number.
Live KPIs: real-time performance indicators such as average call time, resolution rate, and customer wait time.
Interaction history: access to the customer’s previous support interactions with the option to expand details.

Pre-Call

Real-time transcription: live transcription of the conversation, updated line-by-line with auto-scroll.
Smart suggestions: AI detects key topics mentioned and proactively offers suggestions, such as tracking links, return policies, or escalation paths.

POst Call

Call summary: automatically generated summary outlining the main topics discussed.
Quick action shortcuts: AI suggests next steps like scheduling a callback, sending an email with key info, or assigning a follow-up task to another team.
Assistant evaluation: agent is invited to rate the usefulness of the assistant to support ongoing AI improvement through a contextual survey.

Pre-Call & In-Call User Flow

When agents open the application, they are presented with the next customer in the queue, along with key service metrics and a brief summary of recent interactions. This provides immediate context before the call begins.

During the conversation, the agent receives AI-powered suggestions and recommendations generated from the live call transcription combined with the customer’s historical data, such as purchase history and order status. These suggestions are surfaced through bottom sheets, ensuring that guidance appears at the right moment without interrupting the flow of the conversation.

The interaction model is designed to feel fluid and intuitive, allowing agents to stay focused on the customer while accessing relevant insights in real time.

The calculation process

The procedure's registration main feature of the AnestPlan app was first designed following the specifiations required to automatically calculate the proceedure's final price, such as time and proceedure's duration, age, gender and patient's weight, numbers and quantity of anesthesias’ codes and many others.

To calculate the price of an anesthesia procedure, it is necessary to follow these steps:

1

Code consulting
‍They may vary according to the anesthetic classification, from 0 to 8 – the higher the number, the more expensive

2

Consider 100% of the value of the anesthesia with the highest classification
‍The values of the other anesthesias will be added: - 50% if performed at the same incision site; - 70% if performed at a different incision site

3

Double the total amount
In the case of a child weighing less than 2.5kg or if in a private room

4

Add 30% to the total amount
If the anesthesia classification is 7 or 8;
In the case of extracorporeal circulation;
In neonatal surgery;
In the case of gastroplasty for morbid obesity;
For surgeries lasting more than 6 hours;
For emergency surgeries between 7 PM and 7 AM from Monday to Saturday or at any time on Sundays and holidays.

Post-Call User Flow

After the agent ends the call, the application automatically generates an AI-powered summary highlighting the main topics discussed during the interaction. The interface then presents quick action shortcuts, such as scheduling a follow-up or forwarding a task to a specific team.

Upon returning to the home screen, the system displays a contextual survey with a small set of questions designed to capture the agent’s satisfaction and perceived usefulness of the tool, helping inform future improvements.

The calculation process

The procedure's registration main feature of the AnestPlan app was first designed following the specifiations required to automatically calculate the proceedure's final price, such as time and proceedure's duration, age, gender and patient's weight, numbers and quantity of anesthesias’ codes and many others.

To calculate the price of an anesthesia procedure, it is necessary to follow these steps:

1

Code consulting
‍They may vary according to the anesthetic classification, from 0 to 8 – the higher the number, the more expensive

2

Consider 100% of the value of the anesthesia with the highest classification
‍The values of the other anesthesias will be added: - 50% if performed at the same incision site; - 70% if performed at a different incision site

3

Double the total amount
In the case of a child weighing less than 2.5kg or if in a private room

4

Add 30% to the total amount
If the anesthesia classification is 7 or 8;
In the case of extracorporeal circulation;
In neonatal surgery;
In the case of gastroplasty for morbid obesity;
For surgeries lasting more than 6 hours;
For emergency surgeries between 7 PM and 7 AM from Monday to Saturday or at any time on Sundays and holidays.

Expected Impact

While this project does not include production metrics, the proposed solution is expected to:

Reduce handling time by minimizing manual searches
Improve resolution quality through contextualized suggestions
Lower cognitive load for agents during live calls
Increase consistency across support interactions

These outcomes would be validated through usability testing, pilot deployments, and analysis of operational KPIs.

Next Steps

To guide the next steps of the project and validate the impact of the solution after launch, a set of quantitative KPIs would be monitored.

✅ Average call duration
✅ Follow-up action logged rate
✅ CSAT (Customer Satisfcation)
✅ Resolution time per ticket
✅ AI suggestions adoption rate
✅ Satisfaction rate with AI assistant

These indicators are designed to assess performance improvements in call handling, follow-up consistency, AI effectiveness, and overall user satisfaction.

Additionally, behavioral analytics tools such as Firebase or GA4 would be implemented to track feature adoption, engagement funnels, and usability drop-off points.

The calculation process

The procedure's registration main feature of the AnestPlan app was first designed following the specifiations required to automatically calculate the proceedure's final price, such as time and proceedure's duration, age, gender and patient's weight, numbers and quantity of anesthesias’ codes and many others.

To calculate the price of an anesthesia procedure, it is necessary to follow these steps:

1

Code consulting
‍They may vary according to the anesthetic classification, from 0 to 8 – the higher the number, the more expensive

2

Consider 100% of the value of the anesthesia with the highest classification
‍The values of the other anesthesias will be added: - 50% if performed at the same incision site; - 70% if performed at a different incision site

3

Double the total amount
In the case of a child weighing less than 2.5kg or if in a private room

4

Add 30% to the total amount
If the anesthesia classification is 7 or 8;
In the case of extracorporeal circulation;
In neonatal surgery;
In the case of gastroplasty for morbid obesity;
For surgeries lasting more than 6 hours;
For emergency surgeries between 7 PM and 7 AM from Monday to Saturday or at any time on Sundays and holidays.

From live conversations to confident decisions: designing an AI-assisted custommer support app