Inside banks of the future, run by bots | Mint – Mint

17 minutes, 25 seconds Read

“I need to open a savings account,” I typed. “Thank you for your interest in opening a savings account with ICICI Bank,” it responded, before providing a link to the application. Next, I was prompted to respond with either a thumbs up or thumbs down emoji. My question about wanting to open a demat account met with a similar response, and a link to more details.


World’s Leading High-rise Marketplace

The going was good. But I decided to have some fun.

“How should I manage my money when dating?” I asked. “Sorry, I am not sure if I got that. Could you please try rephrasing your question?” I tried again, adding a few more words. “Where should I park my money after getting married? Should I choose fixed deposits or SIPs (systematic investment plans)?”

The bot responded with some suggestions, but it was not the advice a human financial adviser would offer. One of them was an advisory on ‘how to open a fixed deposit’. By now, I had realized that I had exceeded my brief.

To be fair, the bot has been asked stuff which is outside its traditional banking syllabus. iPal, nonetheless, is one of the newer generation bots—it is an artificial intelligence (AI)-based system. All answers are automated and the chatbot is trained to understand the queries and improve the responses. As more people interact with the bot, it learns. Over a period of time, the bot would get sharper and more intelligent.

It is quite likely that in a year, iPal would be able to understand and answer questions beyond its syllabus.

For banks like ICICI that had already invested in classical AI, the next logical frontier is generative AI (GenAI). Simply put, classical AI makes predictions based on historical data. GenAI goes further. It takes data as inputs, learns, and then generates new content such as text, images, videos, music, or code. With a much better understanding of language and context, GenAI’s accuracy is higher. Because it is capable of handling complex questions in a way humans can, it is a great fit for contact centres.

Its adoption, therefore, has huge implications for a bank—both on revenue and profitability. While chatbots powered by GenAI can rapidly shrink the cost of running customer contact centres, the technology can also help reduce risks such as fraud. New use cases that can grow revenue, like loan generation and underwriting, is just about emerging.

“We keep investing in new technologies due to necessity. Our investment in GenAI is one such example,” V.V. Balaji, chief technology officer at ICICI Bank, told me. He added that the bank is in the process of transitioning to an LLM-based bot—one that can adapt to various user inputs, understand nuances, and provide relevant responses—from a rule-based one (which can handle basic conversations).

LLMs, or large language models, are algorithms that use large data sets and are capable of recognizing, summarizing and generating content.

The initial use cases of GenAI in the bank are showing promise, Balaji said. “We will know its complete efficacy after a few months, when the technology matures.”

The bottom line

View Full Image

The continuous training of chatbots can result in a 94-95% accuracy, says Deepak Sharma, the former chief digital officer at Kotak Mahindra Bank.

Globally, banks have typically been quick adopters of technology—from automated teller machines (ATMs) in the 1960s and electronic card-based payments in the 1970s to the broad adoption of anytime online banking in the 2000s, followed by mobile banking in the 2010s, traditional AI, advanced AI and analytics, and now GenAI.

In March last year, Morgan Stanley Wealth Management said it will use OpenAI’s LLM-powered chatbot, ChatGPT, to help its team of financial advisers ask questions. ABN Amro is using GenAI tools to produce summaries of notes that its agents take when customers reach out to its call centre.

Apart from ICICI Bank, other progressive banks too have AI-powered chatbots. Yes Bank, for instance, has a Yes Robot, a digital assistant that offers more than 60 banking services to its customers including features such as booking fixed and recurring deposits, fund transfers, paying bills, managing credit card and debit cards, and raising service requests to update personal details.

Yes Robot, according to Ajay Rajan, country head of digital and transaction banking at Yes Bank, currently has a 33% share of retail deposits booked through the bank’s digital channels and more than 30% of the service requests raised digitally by the bank’s customers. “This has helped enhance customer experience besides ensuring reduced branch dependency for repetitive tasks. As a way forward, we are constantly working towards making our AI solutions more intelligent and intuitive for our customers,” Rajan told me over a call.

Axis Bank, too, has a conversational AI chatbot called Uttar (Hindi word for answer) for bank employees to get a quick response to their queries. The bank is experimenting with an LLM that is being trained on the bank’s data, while keeping the running cost low, mitigating security concerns, and overcoming ‘hallucination’ issues with GPT (generative pre-trained transformer) models), Subrat Mohanty, the bank’s executive director, told me.

When GenAI models produce inaccurate information confidently, it is called hallucination.

What are the cost implications? According to Gartner, a technology advisory company, by 2026, conversational AI deployments within contact centres may lower agent labour costs by $80 billion globally since conversational AI can use chatbots or voice bots to automate all or part of a contact centre customer interaction.

According to tech company IBM, AI-infused virtual agents can cut labour costs by reducing the reliance on human intervention, leading to as much as 30% decline in customer support service fees. Chatbots can also handle 80% of routine tasks and customer questions. Furthermore, chatbots do not take sick leave or go on vacations. They do not have a 9am-5pm job and work across time zones, during public holidays, and in rough weather and floods, too.

Interestingly, when robotic firms pressed for higher adoption of robots on the manufacturing floor two decades ago, these were the same reasons forwarded. In India, we have heard how more automation on the factory floor is a good riddance to trouble rising from labour disputes.

Many bankers don’t fully agree with theories that say humans won’t be needed at all, going ahead. When I checked with Deepak Sharma, who was the chief digital officer at Kotak Mahindra Bank till November 2023, he told me that instead of replacing humans, GenAI models will supplement human intelligence. He explained that banks are using training data to improve outcomes. A lot of call centre test scripts (scripts kept ready to engage with customers) are getting automated. “For training of contact centres where the whole call script gets fed into the LLM models, a human agent can decide whether to use or ignore the output. The continuous training of chatbots can result in a 94-95% accuracy,” he said.

Nonetheless, Daniel O’Connell, an analyst at Gartner, put out a note in August 2022 warning of costs that don’t seem obvious in the mounting noise around GenAI. Implementing conversational AI requires expensive professionals—professionals who have expertise in areas such as data analytics, knowledge graphs and natural language understanding. They are a difficult breed to find. Further, complex, large-scale conversational AI deployments can take multiple years. Gartner estimates integration pricing at $1,000 to $1,500 per conversational AI agent, though some organizations cite costs of up to $2,000 per agent.

Catch the fraud

Customers are at the centre of any business strategy, says Shiv Bhasin, chief transformation officer, IndusInd Bank.

View Full Image

Customers are at the centre of any business strategy, says Shiv Bhasin, chief transformation officer, IndusInd Bank.

Shiv Bhasin is chief transformation officer at IndusInd Bank. A veteran technologist, he has seen many changes in the tech landscape over the last three decades. When I asked him how and why IndusInd uses exponential technologies, he responded by saying that “customers are at the centre of any business strategy”. The bank already uses data analytics to improve the banking experience of customers. “Through data analytics, we are able to understand customers’ usage data, analyse complaints and feedback, enable campaign management and regulatory reporting. It also helps in underwriting and fraud detections,” Bhasin told me.

Banking frauds are all pervasive—on 5 February, Mint reported that employees at some Bank of Baroda branches disbursed fake gold loans last year to meet stiff targets. While employees perpetuated the fraud here, outsiders can do severe damage, too, particularly with the tools that are now available. LLM chatbots and dark web tools like FraudGPT are now helping fraudsters draft more contextually-relevant emails without typos and grammar mistakes. Further, a banking customer’s voice can be cloned using deepfake technology if an attacker can obtain voice samples using spam phone calls that fool the call recipient into responding by voice.

So, how can GenAI help here?

Using GenAI, banks can more easily flag anomalies or behaviours that do not fit expected patterns, following which a human banker can review the suspicious behaviour. Banks can also use GenAI to mandate additional user verification when accessing accounts. As an example, an AI chatbot could ask users to answer a security question or perform a multi-factor authentication.

Bankers who handle fraud reviews manually can now be assisted with LLM-based assistants running what is called retrieval augmented generation, or RAG. This system ensures that an AI model can access the most current data at the backend, tap information from necessary documents to expedite decision-making on whether cases are fraudulent.

According to a blog by Kevin Levitt, who leads global business development for the financial services industry at Nvidia, whose processors power GenAI applications, LLMs are being adopted to predict the next transaction of a customer, which can help banks and payments firms preemptively assess risks and block fraudulent transactions.

Building a score

LLM models can assist in the automated evaluation of creditworthiness by analyzing a customer’s financial history and credit score.

View Full Image

LLM models can assist in the automated evaluation of creditworthiness by analyzing a customer’s financial history and credit score.

An emerging use case is in lending. And this impacts the topline of a bank.

By analysing user data and preferences, LLM chatbots can recommend suitable loan products, interest rates, and repayment plans. LLM models can also assist in the automated evaluation of creditworthiness by analysing a customer’s financial history and credit score. After the loan is disbursed, bots can send personalized payment reminders, reducing the chances of late payments and associated penalties.

Well, AI can help even when credit scores aren’t available, helping banks reach out to underbanked segments of the Indian population.

“Given that 80% of the Indian population does not have a credit score, AI can help both the lender and the borrower by building a credit score for them. Rather than using credit score and credit history, fintech companies are now using something called a ‘social loan quotient’ to assess a loan applicant and determine his/her credit worthiness,” Balaji of ICICI Bank told me.

That social quotient includes digital footprints of a customer—online shopping habits, utility and telephone bill payment history and social media profiles can be analysed by AI systems to determine the creditworthiness of a loan applicant.

The task ahead

While GenAI can save costs and boost revenue, it isn’t without its challenges. What must Indian banks guard against?

To begin with, data security and compliance with privacy regulations are paramount in the financial services segment. GenAI models ingest humongous amounts of training data to learn from it. According to Deepak Sharma, though, in-house LLMs do not pose much of a regulatory problem “since no public data is used, and the data used for training resides on private clouds of banks”. But, then, like we mentioned earlier, GenAI models can also ‘hallucinate’. As is the case with classical AI, GenAI, too, can perpetuate bias. Mastercard has pointed out that if a data pool reflects that a certain demographic has historically received fewer loans, the AI application could consider that fact as prescriptive and discriminate against that group.

Mastercard has also said that criminals can exploit GenAI tools to produce deepfakes or “churn out iterations of deceptive email copy in phishing expeditions”. This implies that security experts must strive to stay a step ahead of cybercriminals.

There’s yet another hurdle I learnt of—perhaps the biggest challenge in the days to come.

Since most banks are running GenAI pilots at the moment, scaling up these projects would require a comprehensive change management plan. According to a recent report by McKinsey Global Institute, a GenAI “scale-up is like nothing most leaders have ever seen”. It goes on to urge business leaders at banks to “interact more deeply with analytics colleagues and synchronize often-differing priorities”; add GenAI-skill related personnel such as prompt engineers and database curators to their existing teams of quants, modelers, translators, besides people with AI expertise such as in cloud engineering and data engineering.

Banks, concluded McKinsey researchers, will be able to tap the enormous promise of GenAI “only by following a plan that engages all of the relevant hurdles, complications, and opportunities”.

Indian banks clearly have an interesting but tough task, going ahead.

This post was originally published on 3rd party site mentioned in the title of this site

Similar Posts