All work

Case study 04 · AI · Telco and contact centres · SaaS

Sidekick AI agent assist and AI sentiment analysis

Product strategy and ownership for an agent assist panel. Know the customer, get a suggested answer, find the right document and read the mood of the conversation, without leaving the chat.

Role
Product strategist and owner
Built for
Contact centre agents and team leads
Sector
Telco and contact centres
Platform
SaaS, cloud-based system
Year
2023
01

Project Overview

Agents in a contact centre carry a lot in their heads. They read a customer, judge their mood, remember the account, and hunt for the right policy, all while a chat or call is running. This project puts that help beside the conversation. A right hand panel shows who the customer is and how the interaction is going, and an assist widget suggests answers, searches documents and tracks sentiment.

The goal was simple to say and hard to design. Give agents more information without giving them more to manage.

02

Platform

  • Desktop
  • Web app
  • Voice, chat and email

The panel lives inside the agent workspace of the Console Next suite. Agents work in it for a full shift, usually on one wide monitor with the chat in the middle. Anything I added had to sit at the edge and never push the chat aside.

03

Challenge

Before this work an agent had to leave the chat to find a customer detail, open another tool to read a policy, and rely on instinct for how the customer felt. Every one of those trips cost seconds and broke focus. New agents made the most trips.

Show the customer and the interaction at a glance, without opening anything.

Offer an answer while the customer is still typing the question.

Let agents search company documents without leaving the chat.

Show mood as a trend, so a bad moment is visible before it becomes a complaint.

04

Strategy & Approach

I treated the right panel as a quiet companion to the chat. Three decisions shaped it.

  • Help at the edge

Everything lives in the right column. The chat keeps its width, and the agent never loses the thread of the conversation.

  • Suggest, never send

The widget offers answers with copy and send controls. The agent reads, edits and decides. Nothing reaches the customer without a human choice.

  • Mood as a line, not a label

One score hides what happened. A small time line shows the dip and the recovery, so agents learn what moved the customer.

05

My Contribution

I owned the product from research to prototype: the structure of the right panel, the assist widget with its Suggestions and Sentiments tabs, the document search, the sentiment visuals, and the way each state looks while it loads, finds nothing or finds three results.

  • Right panel structure
  • Assist widget and search
  • Sentiment donut and trend line
  • Interactive Figma prototype
06

Scope

The prototype covers one end to end flow with AI assist, starting from an open conversation.

  • Know the customer
  • Ask or let AI listen
  • Answer in one click
  • Read the mood
07

User Research/Journey

  • User Interviews

I ran 12 one on one interviews: 7 contact centre agents, 3 team leads and 2 quality analysts. Each session was about 40 minutes. I asked people to talk me through their last difficult conversation and show me every screen they touched.

Agents opened three to five tabs per interaction, and most of the time went on finding one fact.

Customer mood was judged by gut feeling, and agents often noticed a problem only when the customer raised their voice or wrote in capitals.

Team leads wanted a quick way to see how a customer had been treated earlier in the day.

08

User Personas

Three personas came out of the interviews. The first two shaped most of the panel.

  • Sana Malik, frontline agent

Handles voice and chat at the same time, 60 to 80 interactions a day. Needs answers fast and a quiet screen. “If I have to open another tab, the customer is already waiting.”

  • Bilal Ahmed, team lead

Leads a team of 18 and steps in on escalations. Needs the story of a call before he joins it. “Tell me who they spoke to and how they felt, then I can help.”

  • Mehwish Raza, new agent

Three months on the floor and still learning policies. Needs reliable documents and wording she can trust. “I know the answer exists. I just do not know where.”

09

Research Workflow

  • Interviews and shadowing
  • Task and flow mapping
  • Wireframes
  • Prototype and walkthrough

I shadowed agents for two shifts, then mapped their work into four moments: pick up, listen, look up and hand over. Wireframes compared a tabbed side panel against a floating widget. The floating widget won because it could be opened over any panel and closed without losing place. Each prototype walkthrough with agents fed the next round of changes.

10

User Roles & Priorities

Frontline agent: speed. This shows up as ready suggestions, quick chips and one tap copy and send.

Team lead: context. This shows up as the interaction panel with transfers, conference participants and sentiment.

New agent: trust. This shows up as search that opens the full document, so wording can be checked against the source.

11

Defining The Architecture

The structure is deliberately small. One workspace, two panels on the right, and a widget that floats above them.

12

Agent Journey Map

The journey follows one interaction from pick up to hand over. Each moment has a need and a design answer.

13

Core flow: Know, ask and read the mood with AI

The four panels I first designed separately, Customer Info, Interaction Info, Agent assist and Sentiment analysis, serve one moment: an agent in a live conversation who needs to know who they are talking to, find the right answer, and sense how it is going. I merged them into one flow in a single right panel, and made the AI do the searching and the mood reading.

  • Know
  • Ask
  • Answer
  • Read the mood
  • 1. Know the customer and the interaction

The panel opens with the customer’s name, number and a Premium badge, then a sentiment donut that shows how this customer usually feels and who spoke to them in the last 24 hours. The interaction view adds type, queue, wait time and tags, plus who transferred the call and who joined it, so a handover never starts from zero.

  • 2. Ask, or let AI listen

Assist floats over the panel with a search field and a Keywords toggle. With the toggle on, AI listens to the conversation and offers answers as the customer writes, such as a price card for the product they asked about. With it off, the agent searches by hand. The choice is theirs and stays visible.

  • 3. Answer in one click

Each suggestion comes ready to copy or send, with a loading state while new ones arrive. For longer policy questions the agent gets linked documents that open in a large panel over the chat, then close with the conversation exactly where it was left.

  • 4. AI-based sentiment analysis, live

An AI model scores every customer message as it arrives. It returns a label (positive, neutral or negative), a confidence level and the cue words that drove the score, so the agent can see why it reads the way it does. The Sentiments tab then shows three faces and a line of the mood across this interaction. Green, amber and red tell the story without a legend. In the example the customer starts happy, dips when they ask about the credit card, and recovers once the agent answers. The same model raises an alert when the score crosses a threshold, so it flags the dip while it happens, so the agent can slow down and acknowledge the customer. A team lead can read the same line after the call and see where it turned.

The widget answers how this call is going now. The donut answers how this customer usually feels. I kept them apart on purpose, because they serve different decisions.

▶

Interactive prototype

A working feel of the agent assist panel. Press play for a guided walkthrough, or press Customer writes to move the conversation along yourself. Send AI suggestions with one click and watch the sentiment line react.

Agent workspace · Chat · Hamza R. (Premium)Live
14

Usability Testing Based On Heuristic Evaluation

I reviewed the flows against standard usability heuristics. This is where each one shows up in the design.

  • Visibility of system status
  • Match with the real world
  • User control and freedom
  • Consistency
  • Recognition over recall
  • Flexibility and efficiency
  • Aesthetic and minimal design

Visibility of system status comes from the loading states and the live sentiment line. Match with the real world comes from faces and colours that agents already read as good, neutral and bad. Control and freedom comes from the Keywords toggle and the close button on every panel. Recognition over recall comes from quick chips and the document list.

  • A/B Five Minute Testing

I tested with 8 of the interviewees across two rounds. Round one was a five second test of the right panel. Round two was a task test, such as “find out what the customer said earlier today” and “find the registration policy and read the first paragraph”.

  • Conclusion & Learnings

Testers read the donut quickly but wanted dates on the contact history, so I added them. The widget was found by everyone, though two people missed the Keywords toggle at first, so I gave it a label. The sentiment line was the favourite. Team leads said it saved them a listen to the recording.

15

Some key takeaways from this project are

Put help at the edge. The chat kept its width and agents kept their focus.

Suggest, do not send. Agents trusted the widget because it never acted for them.

Show trends. A line explains more than a score, and it starts a better coaching conversation.

Design the empty and loading states. Half of an assistant’s life is waiting for an answer.

Next time I would test with real chat volume earlier, because a busy agent handles more messages than a prototype can show.

Thank you for your time!

Work together

Have a complex problem of your own?

Start a conversationSee all work