CSR: Using AI to improve Financial Inclusion

About 1.7 billion people worldwide are “unbanked,” that is, they do not have an account at a financial institution or through a mobile money provider. Companies in India can also have difficulties in accessing finance. Artificial Intelligence can help address core difficulties holding back financial inclusion: difficulties in verifying identities and in a lack of traditional data for underwriting services to vulnerable populations, says a new report from McKinsey Global Institute on applying AI for social good.

Artificial Intelligence-based financial inclusion products, which are already being deployed, bypass the need for a traditional credit score through analyzing digital footprints. Many fintech startups are entering the alternative credit scoring space with AI-enhanced solutions, particularly in countries such as Bangladesh and Pakistan where the populations are large and significant portions are unbanked.

Successful solutions

Companies such as CreditVidya, ZestFinance, and Lenddo capture alternative data by device, browser, and social media fingerprinting to generate a predictive model of creditworthiness.

Behind its SMS and internet-based interface, predictive algorithms leverage several Artificial Intelligence capabilities to analyze social and telecom data and assess creditworthiness. The information is then processed in minutes and produces a credit score, which determines the size of the loan allowed.

Various possibilities

A range of capabilities can be leveraged for these products, including natural language processing, structured deep learning, and person identification of social and telecom data. Long short-term memory recurrent neural networks can be trained to recognize an individual’s credit risk. Image-processing capabilities can be used as an additional layer of verification to confirm an individual’s identity.

Structured and unstructured data from sources including social media, browsing history, telecom, and know-your-customer data can be used to train Artificial Intelligence models. Solutions are likely to start with external data such as longevity as a telecom customer, and the model is then augmented against a client’s actual product borrowing performance.

Key challenges

Integrating multiple data sources, given different methods of storing information, is one challenge. The outcomes must be tested rigorously and explainable where necessary, given that an Artificial Intelligence model is analyzing personal data to sort people, assess the credit risk of customers, and potentially reject some.

Representative positive and negative data need to be collected to help reduce unwanted biases. Identifying the levers considered most strongly by the model when determining a credit score could help, as could providing information on the model decision-making process, particularly for rejected users.

“Last mile” implementation is also a potential bottleneck, because many adults in emerging economies who are unbanked also do not have mobile phones or internet access. This prevents the creation of a digital footprint, including telecommunications and online social history, which are the data needed for this form of alternative credit scoring. Overcoming such bottlenecks may require building partnerships with NGOs to provide funding for basic technology access.

This article is part of a series on ‘Artificial Intelligence for social good and CSR’

Subscribe