~25%

More Collections

~30%

Cost Reduction

In this case study

Fintech

Industry

Overview

Client overview

Our client is a fintech company, with over 25 million MSMEs as their customers across several products. One of their major product offerings includes providing short-term credit to MSMEs. The re-payment for such credit is to be made on a daily basis (EDI-Equated Daily Installments).

For missed EDI payments, the client makes reminder calls to the customers. As their customer base was growing rapidly, the scalability of making reminder calls was becoming infeasible due to cost considerations.

Challenge

The customer support challenge

Collections calls are necessary but repetitive. In most cases, they just involve reminders and follow ups to the customers to make them pay back on time. However, our client faced several challenges:

No centralized system to access critical data real-time

  • Calling data scattered across multiple agents (eg # calls made, % calls connected)
  • Inconsistent and delayed data entry by agents leading to reporting challenges
  • No single dashboard to provide live status of collections (eg # active agents, % total defaulting customers called, % collections made)

Scalability challenges

  • Increasing costs due to the need to increase the number of calling agents
  • Redialing logic was not followed in >30% cases (eg if a customer promised to pay by 3pm, and if they had not paid by then, they had to be called again)

Because of these challenges, efficiently solving for collections was a major priority for our client. 

Solution

Robylon's solution

Robylon’s AI voice agents are designed to effectively perform high-value, repetitive calls at a fraction of the costs of human agents. Hence, our client was keen on exploring how Robylon could solve their problems. 

We did a pilot for around 3 weeks where our AI voice agents dialled up some low risk customers reminding them about their missed payments. Our pilot was a success and we were able to increase collections by 10% already within 3 weeks.

Hence, we are now onboarded to handle 90% of all collection calls using our AI agents. Only escalated and priority cases were to be handled by human agents.  Here is a brief of how we approached the problem:

Step 1: Specialized AI agents to refine the calling SOP

  1. A concise 2 page SOP was provided to set the context for the AI agent
  2. Transcript of 1000+ minutes of high quality call recordings fed into the model to train and identify patterns
  3. SOP was refined using the historical transcripts to handle edge cases

Step 2: Building, testing and deploying voice AI agents

We developed specialized voice agents for collections with the following features: 

Voice setup

  • Mix of English, Hindi and regional languages (Tamil, Kannada, etc.)
  • 10+ variations in voices across male and female options
  • Slight variations in tonality based on the number of missed payments (eg milder tone for upto 2 missed payments)

Human-like conversational measures

  • Ability to not just speak and listen, but also to engage in 2-way conversations 
  • Natural speed and pauses while speaking along with acknowledgment of customer responses

Defining the guardrails

  • Strict compliance with applicable rules around collections calls
  • Always remain polite, concise and to the point

We tested and iterated upon the voice agents with 250+ scenarios covering most of the edge cases we could think of (e.g. how to handle rude customers, what to respond in case a customer is busy, etc.)

Step 3: Setting up the calling operations

The next step was to set the calling process to make these calls using our AI agent:

  1. Every day, defaulting customers’ data was given to our specialized risk-analysis AI agent. It created a priority order list of customers to call based on risk analysis parameters. 
  2. The voice agent made calls to the customers based on the defined priority. Post the call, the captured data was updated in the system real-time (eg by when the customer promised to pay)
  3. Certain calls were flagged for human intervention (eg in cases of rude behaviour)
  4. Redialing criteria was defined and automated (eg redialing in an hour if a call remains answered, redialing to remind customers to pay if they hadn’t paid by the promised time)

Step 4: Structured reporting and post-call analysis

Since data capturing was made real-time and streamlined by automation, it was possible to do real-time analysis of calls and track critical parameters. A single view dashboard was created covering the # calls made, % calls answered, % amount collected, etc.

Result

Robylon’s voice AI agents solved the problems of scalability while additionally saving on costs too. Using our voice agents, the client was able to make 3 times more calls per day with only a fraction of additional cost they otherwise would have had to spend. 

What is more, the net collections increased by 25% which offset the extra calling costs by more than 4 times.

Other case studies