Integrating artificial intelligence into small business digital onboarding: Wins and challenges
This post is by Lucinda Revell, co-founder of Boost Capital, an organization that enables financial service providers to serve customers via popular chat platforms such as Facebook Messenger, Telegram, and WhatsApp, without downloading another app. Boost Capital has partnered with Mastercard Strive to support the digitalization of Filipino and Cambodian small businesses.
Don’t worry, I’m not going to start this blog by telling you that ChatGPT wrote this first sentence.
We use plenty of AI at Boost Capital, but the ideas and instructions that direct that AI are still traditionally generated by us, either in strokes of genius, in fits of desperation, or in the course of regular tech roadmap achievements. At Boost, we’ve integrated AI into our chat-based financial service onboarding in multiple ways:
- To power unstructured chatbots¹ that make financial services more accessible for small businesses (Apply anytime, from anywhere, just like chatting with a friend.).
- Natural language processing (NLP)² to allow entrepreneurs to chat in everyday language.
- Facial recognition/liveness detection³ to protect applicants and financial service providers (FSP) from fraudulent applications.
- ID/document interpretation using optical character recognition (OCR)⁴ to save applicants time (Just snap a pic! Don’t waste time transcribing info!).
- ID/document fraud detection⁵ to provide verified info to FSPs, allowing FSPs to confidently serve online applications.
In partnership with the Mastercard Center for Inclusive Growth, in 2024 Boost set out to increase the digitalization of Filipino small businesses by providing access to digital financial services, digital sales and transaction channels, and financial literacy training to improve their financial health and accelerate the growth of their businesses. The integration of AI into our onboarding processes could bring transformative benefits to small businesses by:
- Simplifying processes: AI-driven onboarding simplifies applications, enabling small businesses to access loans, grants, and other financial products, fostering inclusivity.
- Enhancing efficiency: Automating data extraction and verification reduces the time and effort required for onboarding, allowing small businesses to focus on growth and operations rather than filling in forms.
- Improving accuracy: AI minimizes human errors in data processing, ensuring reliable financial and operational decisions by FSPs and improving the provision of financial services to small businesses.
Intelligently deciding when and how to integrate AI into financial service provision for maximum efficacy includes nimbly adjusting tech to account for the difficulties particular to emerging markets and small business–specific use cases. Given the potential impact for small business growth when doing AI right, Boost is excited to be at the forefront of these AI learnings.
“Boring” AI is moving the needle for small business onboarding
Right now, unstructured chatbots and NLP are the flashy front end that gets lots of attention in the news. Unstructured chatbots and NLP allow for an alternative UI/UX for digital onboarding rather than structured chat conversations. Imagine the difference between a rambling conversation with a friend (unstructured) and ordering lunch from a busy waiter at a restaurant (fairly structured). Both serve a purpose, but the experience is very different.
However, we’ve found that “boring AI” is what really moves the needle for our FSP partners. Boring back-end AI doing know-your-customer (KYC), document validation, and fraud detection is powering the most bank growth right now.
Successfully deploying these AI-powered back-end features can have a huge impact on your customers, giving greater access to finance because digital channels allow convenient and easy applications (that can happen anytime, anywhere). The chat apps that Boost utilizes as interaction channels with applicants have particularly low barriers to entry. They are easy to understand and popular on any smartphone: Facebook Messenger, WhatsApp, Viber, and Telegram.
What’s been easy, and what’s been challenging in launching AI?
AI integration comes with varying degrees of complexity. Some features are straightforward to implement, while others face significant technical and contextual challenges. Our FSP partners have told us they’ve found challenges and a lack of results from international vendors providing AI features. To achieve better results, we found that we needed to train our AI on local documents, as there are many AI features that require localization to produce quality results. We also needed to build a variety of features to account for lower-quality cameras on low-cost smartphones that are prevalent in emerging markets.
For fraud detection, which is crucial to e-KYC (think “Is this ID fake or not?”), we use a minimum of nine AI-driven features. Some have been easy to implement, and others more challenging, as some are impacted by lack of standardization and poor image quality.
Similarly, we’ve found challenges in using optical character recognition (OCR) for local document interpretation, which is required to render the information on those documents (provided as part of an application) into a format that can easily be imported into the back-end system of the bank partner to approve the application. For instance, we rely on OCR to read the name and birth date from IDs rather than asking the client to type that information, making the client experience more efficient. Beyond the difficulty created by having a wide diversity of document types, especially when it comes to documentation provided by small business owners, we encounter challenges in:
- Poor document quality: Low-quality scans or photographs result in blurry or skewed images, making it difficult for OCR to interpret unclear text. As a solution, we’ve deployed image pre-processing techniques, such as resolution enhancement, to improve OCR accuracy.
- Complex layouts and formats: Documents with intricate designs, multiple columns, or embedded images are difficult for standard OCR systems to read. We use a human-in-the-loop approach to fine-tune the parameters necessary to handle different layouts and train our AI on local formats.
- Handwritten text recognition: Many small businesses still use handwritten invoices and accounting. Even some government documents are handwritten. Deciphering handwritten text remains challenging due to diverse handwriting styles, resulting in data errors. We’ve deployed Intelligent Character Recognition (ICR) technology to improve accuracy by leveraging machine learning.
- Language and character set limitations: In countries like the Philippines, where there are a plethora of dialects, documents with multiple languages or special characters reduce OCR accuracy because of misinterpretation of text. We’ve trained our OCR systems on diverse fonts and character sets to improve recognition.
For FSPs who deploy our white-labeled onboarding tech (commercial banks, microfinance institutions, and payment platforms), the impact is significant — an instantaneous expansion of their reach to access new client bases. We also see these banks utilize the greatly expanded data pool provided by AI-enabled digital onboarding to improve their products and underwriting to serve the underserved.
The benefits of great customer experience that is always “on and efficient” are worth figuring out the technical challenges, but some require bigger investments than others. At Boost Capital, we definitely think the “boring AI” is where there is scalable high-impact value to customers right now, with more tech upgrades coming soon.
Notes
¹ “An unstructured chatbot is an AI-powered conversational system that can process and understand data without a predefined, rigid format.” (OpenAI. (2023). ChatGPT (GPT-4 version) [Large language model] https://www.perplexity.ai/search/what-is-an-unstructured-chatbo-uEDDXuUMSkeNeCsMR4fKTQ#0)
² “Natural Language Processing (NLP) is a subfield of artificial intelligence and computer science that focuses on enabling computers to understand, interpret, and generate human language” (OpenAI. (2023). ChatGPT (GPT-4 version) [Large language model] https://www.perplexity.ai/search/what-is-natural-language-proce-_9TvUKgNREe24XaI2b7uIg#0)
³ “Facial recognition is a technology that identifies or verifies a person’s identity by analyzing and comparing their facial features to a database of known faces. It essentially answers the question “Who are you?” and “Liveness detection, also known as anti-spoofing, is a crucial component of modern facial recognition systems. Its primary purpose is to determine whether the biometric data being presented comes from a live, present person rather than a fake representation. Liveness detection addresses the question “Are you really there?”” (OpenAI. (2023). ChatGPT (GPT-4 version) [Large language model] https://www.perplexity.ai/search/what-is-facial-recognition-liv-87fyo7UzSmO3PqvGB2Gzgg#0)
⁴ “Optical Character Recognition (OCR) is a technology that converts images of text — whether typed, handwritten, or printed — into machine-readable text format.“ (OpenAI. (2023). ChatGPT (GPT-4 version) [Large language model] https://www.perplexity.ai/search/what-is-ocr-8ds0KwWzRwiV0zBccnRvhg#0)
⁵ It’s pretty simple to explain: this is the digital version of showing a fake ID to security personnel.