Conversational design| AI product design
Enhance Customers Experience in Chatbot
Role: Product designer | Date: Jul - Sep , 2025
Brief
SSE is aiming to develop an omnichannel strategy to reduce 80% of customer traffic from non-digital channels. The chatbot was introduced as a key self-service platform to support the business in achieving this goal.
However, despite being positioned as a primary digital touchpoint, the chatbot has not delivered meaningful outcomes and has instead introduced friction into the customer journey.
Goal
With this project, we aim to transform the chatbot into a reliable, task-oriented digital assistant that improves self-service success and reduces operational costs. The goals are to:
-
Increase engagement and reduce conversation abandonment rates
-
Improve intent recognition accuracy and reduce misrouting
-
Strengthen the underlying infrastructure, including handoff and fallback solutions
Outcome
82%
45%
41%
Reduction in telephony and email traffic for common enquiries, reflecting a significant shift in how customers are choosing to engage.
Increase in chat usage, demonstrating growing customer confidence in the digital channel.
Increase in customers being directed to self-serve for task completion, highlighting the impact of a more intuitive and effective chatbot experience.
Quantitative data analysis: Understand the “AS-IS”
.png)
At each layer, we analysed the data from five different perspectives to answer the key question.
Does this layer reduce friction and get users to the right solution faster?
What can we see is that...
With around 60080 distinct interactions in the past year, we have recognised 62% of the abandonment rate of the conversation with the 56% drop out noticed before the routing layer. In the beginning of guided layer, users who have clicked the button instead of free typing have 30% higher query resolution rate. However, the menu click through rate are only 26% among all users. With free typing users, there are 72% of missed message rate at intent detection layer.
What insights can we draw from this data?
-
Customer drop-off at the point of verification: The bot currently redirects customers to an external site when identity verification is required. This handoff has resulted in a significant drop-off at this stage, with the majority of customers not returning to complete their intended task.
-
High abandonment rates across functional flows: The current flows are not designed to reflect a natural, conversational experience. The number of steps involved, combined with unclear guidance, results in very low completion rates and a high proportion of customers abandoning the journey particularly within the first half of the process.
-
No flexibility once a functional flow begins: Once a customer enters a flow to complete a task, they are unable to ask questions or seek support until the task is fully completed. This rigidity makes it difficult to provide assistance at the moments customers need it most.
-
Intent detection as a significant failure point: The bot frequently struggles to interpret customer intent, particularly when customers use free text. This leads to customers being directed to fallback or generic responses, creating frustration and driving early abandonment.
-
No fallback solution, resulting in high dead-end rates: When the bot is unable to handle a query, there is currently no alternative pathway in place. This forces customers toward non-digital channels, undermining the overall digital containment strategy.
Following the playback of findings to wider stakeholders, the business has aligned on the following priorities:
-
Improve intent recognition to reduce the risk of missed or misunderstood messages.
-
Ensure all actions can be completed within the conversation to minimise drop-off risk.
-
Preserve context throughout the journey, allowing customers to interrupt a flow at any point without losing progress on an incomplete task.
-
Review and redesign functional flows to deliver a more streamlined and cohesive customer experience.
To Be Architecture Overview
.png)
-
Create an interruptible flow with context preservation which can mirror how a human agent would handle it: pause, answer the quick question, then say "right, back to where we were.
-
Verification is handled inline, removing the need to redirect customers to an external website. By keeping authentication within the conversation, all interactions remain in one central place — reducing the risk of drop-off before functional journeys begin.
-
A feedback loop is created by integrating the survey directly into the conversation. Allowing the survey to follow naturally at the end of the interaction streamlines the customer experience and increases the volume of feedback collected.
Streamline functional flows
Example: Submit meter reading
-
Transparency throughout the process: Meter readings must be submitted within a 10-day window. The previous bot accepted submissions at any time, which caused confusion when customers continued to receive estimated bills despite having submitted a reading. The updated bot now informs customers when their next submission window opens and offers a reminder to prompt them at the right time.
-
Active customer engagement: Customers are shown their previous submission record and alerted if they have already submitted a reading within the current window. Without this safeguard, a customer could unknowingly submit duplicate readings, leading to billing errors.
-
Automatic error detection: Readings are validated before submission. If a new reading is lower than the previous one, the bot flags this as a potential issue, displays the last recorded reading for reference, and prompts the customer to retry — avoiding a dead end and ensuring only accurate data is submitted.
Create feedback loop
-
Rating scales disguised as conversation Instead of a 1–5 star widget, the bot asks "Have we provide what you need today?" and presents quick-reply buttons like "Really helpful / Fine / Frustrating" — three options maps well to positive/neutral/negative without feeling clinical.
-
Emoji reactions a row of 3–5 emoji (😊 😐 😞) works well in chat because it's the native language of messaging. It's fast, requires no reading, and maps cleanly to sentiment scores behind the scenes.
-
Free text with NLP analysis let the user type anything — "it was a bit confusing but fine" — and analyse the sentiment on the backend. Richer signal, but lower completion rates and harder to aggregate.
Outcome
Following the redesign of the chatbot experience, a significant increase in chat usage has been observed. Alongside this, completion rates across functional flows have improved, accompanied by a notable reduction in customer drop-off during the first half of those journeys.
Quick button click through rate
24% → 65%
Task completion rate
31% → 89%
Positive customers sentiment score
32 → 68
Verification drop out rate