Understanding Intent from Conversations AI

In this blog series, I will share some of my lessons from my experiences driving AI strategy and implementing path-breaking solutions.

On this page

The banking and financial services industry runs on trust. Whether inside a branch or on a digital channel, people feel valued when they get good customer service from their financial institution. In an effort to keep customers happy while controlling rising costs, banks and credit unions have deployed interactive voice response (IVR) systems and chatbots to automate routine requests through self-service. This has resulted in reduced traffic into physical branches and a significant increase in the volume and complexity of digital interactions.

AI technology that can understand the intention of the customer — to accurately predict the flow of next-best actions or follow-on questions to gather relevant data — is critical to providing the best digital experience.

What happens when you don’t know the intent?

Consider a typical virtual agent that mimics a human in the contact center handling conversations by phone, chat and text channels. If the AI understands the intent poorly, it can result in bad responses that negatively affect customer satisfaction and can eventually hurt brand quality and loyalty. 

When the first widespread digital AI assistant, Apple’s Siri, was introduced, there was a problem. A person used a voice-activated personal assistant in his car to make a phone call while driving. He said, “Call my girlfriend.”

The personal assistant replied, “Which one?”

Traditionally it has been difficult, to say the least, to communicate one’s intent by voice. Even today, people prefer to use buttons and links to communicate their intentions with a computer to get what they need. With recent advances in AI technology, however, capturing voice-based intention is now possible with a much higher degree of accuracy. Especially when you are looking to do simple things like “Check my balance,” “Pay my rent,” “Speak to an agent,” and so on.

How does it work and what are the challenges?

Let me explain the intent recognition flow in simple terms and outline how we have overcome the challenges in how users interact.

The AI system processes the intent of an utterance or a question by taking the input, pre-processing it by converting it to plain text, analyzing it using an intent classification model, extracting sentiment, and determining the intent behind the question based on what it has learned. The goal is to accurately determine the intent behind the utterance or a question so that the AI can respond in an appropriate manner.

Here are some challenges and their solutions:

ChallengeSolution
It is not scalable to have an intent engine for each channel and the type of interaction.Have one intent recognition model that can convert a range of utterances into a single intent to ensure consistency in responses.
Anticipating typographic errors, grammatical errors, poorly formed sentences like mumbling words, repeating words, filler words etc.Using an augmentation algorithm, an AI model can be trained to generate human-like examples of errors and mistakes. This data is used to further refine the intent engine.
There are many ways to say one thing, such as “What’s up?”“How are you doing?” or “How’s it going?” etc.Using a paraphraser algorithm, generate several samples for every intent with variation of utterances. This data is used to further refine the intent engine.
Ambiguous questions can have more than one meaningThis happens when the utterance is ambiguous. For example, the word “credit” can point to multiple intentions. Here, we use a disambiguation process to eliminate misfits.

Credit unions are jumping on the AI bandwagon

Credit unions are evaluating their AI and digital strategies to reimagine the member experience.

In the context of virtual assistants or chatbots, credit unions are leveraging AI to deflect calls to the AI virtual agent to automate 40 percent of the requests and reduce the call wait time in the call center. For example, if a customer asks “What’s my account balance?” the AI system would understand the customer’s intent as a request for account information and respond with the current balance.

If the same request comes on the website’s live chat channel, the same AI would understand the intent and respond with the current balance.

From a member perspective, the experience is consistent across channels, and they can begin disciplining themselves to think digital-first in resolving their problems — instead of driving to the nearest branch or picking up the phone on impulse to talk to a human agent for simple queries.

This is a new focus and a priority for credit unions to be relevant in the financial services industry. It’s not going to be solved overnight, but progressive-thinking credit unions are picking up the pace, going through technology refreshes in the contact center and automating communication channels with virtual agents.

Best practices in implementing intent automation

Credit unions need AI solutions that offer natural language understanding (NLU) models, pre-trained on large datasets of conversational data pertaining to the financial and banking industry. The intent engines created from these models should support custom utterances that can be added to the existing ontology for intent determination — using a zero-shot learning approach. As a good practice, a weekly report of all missed intents that the virtual agents are not trained to recognize is essential to make continuous improvement to intent automation.

To accelerate the deployment of virtual agents, it’s helpful to have an automatic intent discovery tool that can capture publicly available data on credit union websites and make it instantly available for members to get responses.

An advanced level of intent automation involves multi-intent understanding models — where the virtual agents or chatbots can process multiple diverse intents and execute the flows in sequence without dropping the member requests.

Finally, a multi-tenant federated model architecture allows credit unions to leverage AI learning from each other as the model is continuously enhanced.


Read the next post in this series: No code? No Problem. Designing conversational experiences made easy

Scroll to Top

Let's Talk