Learn how Robylon’s AI agents automated 80% ticket resolution for an online trading platform

Trading is a complex game. Naturally, the support queries are equally complex too. And what is more, millions of dollars are traded every day. So, it is critical to resolve customer queries accurately and promptly. 

Some of the customer support challenges our client faced were the following:

‍

Complexities of resolving queries

• 600+ use cases (query types), each with a unique way of handling them
• Absence of a structured, comprehensive SOP for the humans agents
• Siloed and limited access to data / information for the agents

Operational challenges

• Higher than optimal average first response time of 3-4 hours
• Inconsistencies in responses to similar queries by different agents
• Need for frequent quality checks to prevent human errors

Increasing costs

• Large and frequent training costs to ensure compliance with strict regulatory rules
• Need to increase the number of customer support agents with rise in queries

‍

Our client is an industry leader and is known for its customer-first approach. However, due to the above challenges, their customer support was missing the WOW factor.

A quick diagnosis of historical tickets revealed that 80-85% tickets were repetitive in nature. Our previous experience has shown that AI agents are very effective in automatically resolving such queries. 

Hence, we set out to build and deploy AI agents by taking the following steps: 

‍

Step 1: Using AI agents to make SOP based on historical tickets

1. 500k+ historical tickets were fed to our specialized AI agents to generate SOPs 
2. The AI agents created well-structured SOPs by logically merging the 600+ use cases into 15 major issue types
3. Manual quality check of SOPs was done to ensure that the suggested procedures had a customer-first approach and were also regulatorily compliant

After final sign off over the SOPs from the client, we started building AI agents. 

‍

Step 2: Building and deploying AI agents

1. 15+ AI agents were made to address the specified use cases with each agent covering 30-40 use cases
2. All new tickets were then fed into the AI agents and their responses were continuously monitored by our team
3. Based on robust, real-time feedback, the AI agents were able to train and re-train themselves over multiple iterations

The monitoring and iteration process was run for ~15 days over 3k+ tickets. At the end of 15 days, the AI agents resolved 83% tickets automatically with 93% accuracy, but it was not enough. The WOW factor was still missing!

‍

Step 3: Handling edge cases by Human-in-the-loop model

We realized that there were always going to be certain edge cases over which the AI agents will have low confidence scores in resolving them. Such cases needed human intervention in order to increase accuracy. Hence, 

1. AI agents were instructed to escalate tickets to humans whenever it had a low confidence score on its action.
2. AI’s suggested responses were then vetted and improved (if required) by an in-house team of 2+ human agents

This model improved the accuracy of responses from 93% to almost 100%.