This article is written by Pritesh Gupta, a serial entrepreneur, advisor, and investor.
Why Alternative Data and Behavioural AI Are Unlocking Financial Access for Emerging Markets
How voice, video, and digital behaviour are revealing creditworthiness in populations that traditional finance cannot see — and what this means for banks, NBFCs, and lenders across Asia.
There’s a problem at the heart of financial services that doesn’t get talked about enough, and it’s this: the people most likely to be denied credit are not always the people least likely to repay it.
I’ve spent a significant part of my career looking at the gap between what data systems can observe about a person and what is actually true about them. In credit, that gap is enormous — and the consequences fall most heavily on the populations that can least afford it.
Here’s the situation. According to the World Bank’s Global Findex Database, approximately 1.4 billion adults worldwide have no access to formal financial services. No bank account, no credit history, no way to participate in the organised financial system. And within that 1.4 billion, a large and identifiable segment of people are fully capable of repaying loans responsibly. They have income. They have discipline. They have genuine intent. What they don’t have is the specific type of documented financial track record that traditional credit scoring systems require to make an assessment.
So the system declines them. Not because they cannot repay — but because the measurement instrument has no signal.
That’s the problem InsightGenie was built to solve. And the solution requires rethinking, from first principles, what creditworthiness actually is.
The unbanked and underbanked are not a monolithic group of high-risk borrowers. That conflation — the assumption that no credit history means high credit risk — is one of the most expensive errors in financial services.
Consider who actually makes up this population. A motorcycle taxi driver in Jakarta who has paid his phone bill without fail for six years. A market trader in Dhaka who has run a profitable micro-business for a decade. A young professional in Manila who has only recently entered formal employment and never needed a loan before. Traditional credit scoring looks at all three and sees the same thing: insufficient data. Application declined.
But these people are not equivalent risks. Their personality traits, behavioural patterns, and habitual responses to financial pressure differ enormously — and those differences predict, with substantial accuracy, how they would behave if extended credit.
The challenge is that traditional credit scoring has no way to access those differences. It can only read financial history. And for people who have no financial history, it is blind.
This is not a niche problem. In markets like Bangladesh, Vietnam, Indonesia, Cambodia, and the Philippines — all markets where InsightGenie operates — the majority of adults have either no credit history or insufficient credit history to access mainstream financial products. The scale of excluded creditworthy borrowers represents an opportunity of extraordinary magnitude for financial institutions willing to look beyond traditional signals.
To understand why alternative data is necessary, it helps to be precise about what traditional credit scoring actually measures and why those measures fail.
The core limitation can be stated simply:
Traditional credit scoring: measures what a borrower has done financially in the past InsightGenie: measures who a borrower actually is — the stable behavioural traits that determine how they will act under financial pressure in the future
Credit bureau data is retrospective. It records what a person has done with credit in the past. For someone who has never had access to credit, the bureau returns a null result — not a negative result, just nothing. The system declines the application not because the person is high-risk, but because the measurement instrument has nothing to say. The absence of data is being treated as evidence of risk, which is a logical error with enormous financial consequences.
Income verification fails in informal economies. For the self-employed, the informally employed, or those with seasonal or irregular income — which describes the majority of adults in many emerging markets — income documentation is incomplete, unavailable, or easily fabricated. The verification infrastructure that works in Singapore collapses in Cambodia.
Self-report psychometric tools can be gamed. Some alternative lenders have tried deploying questionnaire-based psychometric assessments as a substitute for credit history. The problem is fundamental: when a borrower knows that appearing conscientious and responsible improves their access to credit, they present themselves that way. The questionnaire ends up measuring the applicant’s understanding of what the lender wants to hear — not their actual behavioural disposition. In a credit context, where the financial incentive to misrepresent is high and the stakes are real, this limitation is critical.
The competitive gap in scoring accuracy is stark. A comparison of credit scoring solutions across Southeast Asian markets shows that the most sophisticated providers using traditional alternative data achieve predictive accuracy in the range of 40–50%. Standard bureau-based approaches cluster at 20–40%. InsightGenie’s multi-modal approach — combining voice, digital footprint, and video — achieves 82–93% predictive accuracy for loan default probability. That’s not an incremental improvement. It’s a different class of capability.
The core claim I’m making is this: a person’s willingness and ability to repay a loan is encoded in their behavioural patterns — patterns that are measurable through voice, digital footprint, and video, independently of whether they have ever had a loan before.
That claim requires a scientific foundation. Here it is.
The personality trait most consistently linked to responsible financial behaviour is Conscientiousness. Conscientious individuals are characterised by self-discipline, planning orientation, commitment to obligations, and persistence under adversity. These are precisely the psychological traits that predict whether someone will prioritise debt repayment when money is tight. Research by Demin et al. (2019) specifically examined the psychological components of credit scoring and confirmed the link between self-control, attitudes toward debt, and loan repayment outcomes.
Conversely, high Neuroticism — emotional instability, poor stress management, impulsive responses to adversity — is associated with a higher probability of payment disruption when the borrower faces financial pressure. This is not a moral judgement; it’s a probabilistic statement about how different personality profiles respond to stress.
These traits are stable over time. They don’t fluctuate month to month. And — critically — they are not self-reported. They are encoded in physiological outputs that can be read from voice and video data, without any deliberate cooperation from the applicant. That’s what makes them valuable in a credit assessment context.
The same prosodic features of speech that reveal personality in a hiring context also reveal it in a credit assessment context — because they’re measuring the same underlying traits.
InsightGenie’s voice analytics engine analyses the rhythm, intonation, stress patterns, pitch variability, and temporal structure of a short voice sample. These features are generated by the larynx under autonomic nervous system influence — they’re physiological signals, not linguistic ones. As documented in InsightGenie’s Science Behind Predictive AI Technology whitepaper, the platform identifies over 40 prosodic biomarkers that map to credit-relevant psychometric dimensions:
These biomarkers feed machine learning models trained on hundreds of thousands of labelled samples to produce a credit default probability score. The result across InsightGenie’s client base: 82–93% accuracy in predicting loan repayment behaviour, a Gini coefficient of 65–77% — a standard measure of a model’s discriminative power — and 30–50% reduction in non-performing loans.
A Gini of 65–77% is genuinely strong performance for a credit model. Most bureau-based models in mature markets target a Gini in the 50–70% range. Achieving it on a population with no credit history, using behavioural signals alone, is something traditional approaches simply cannot do.
Alongside voice analytics, InsightGenie analyses the applicant’s digital footprint — not as surveillance, but as a structured reading of behavioural patterns that reveal habitual traits without self-report.
The digital footprint model operates on a layered principle. Core identity signals — phone authentication, email authentication, eKYC verification — establish identity and flag fraud risk. The behavioural layer above is where the predictive signal lives: the patterns of digital engagement that reveal personality and financial disposition.
Platform category is among the most powerful features. Whether an applicant’s digital presence clusters around professional networks, consumer platforms, or other categories reveals systematic information about their financial discipline and orientation. In InsightGenie’s work with clients across Indonesia, this platform category signal showed particularly high feature importance scores in predicting credit outcomes.
I want to be transparent about what the digital footprint model can and cannot do. InsightGenie’s own documentation is candid on this point: authentication signals alone provide verification, not prediction; personality inference from behavioural data is probabilistic rather than certain; and the stability of platform-category signals requires ongoing validation across different cultural contexts. That transparency is the right posture. Any model that claims more certainty about human behaviour than this should be treated with scepticism.
InsightGenie’s third signal layer is the one I find most technically remarkable. Using remote photoplethysmography (rPPG) — a computer vision technique that extracts the blood volume pulse from subtle variations in skin colour captured by a standard camera — the platform can measure cardiovascular signals, including heart rate variability (HRV), from a brief video clip. No contact. No wearable device. No medical equipment.
The scientific basis for rPPG was established definitively by de Haan and Jeanne in a 2013 paper in IEEE Transactions on Biomedical Engineering, which demonstrated that chrominance-based rPPG methods achieve 92% agreement with contact pulse oximetry across a population of 117 subjects spanning the full range of skin phototypes, with RMSE and standard deviation both a factor of two better than prior methods. Crucially, the method was validated specifically across diverse skin types — not just lighter-skinned populations — which matters enormously for the Southeast Asian and South Asian markets InsightGenie serves.
HRV, extracted from the video signal, is a validated measure of autonomic nervous system regulation. As documented in the research literature (Shaffer and Ginsberg, 2017; Baevsky and Chernikova, 2017), HRV metrics including RMSSD and spectral power in the HF and LF bands index the individual’s stress state, emotional regulation capacity, and autonomic resilience. In credit terms: a borrower whose physiological stress response is poorly regulated is at higher risk of payment disruption when financial pressure mounts. That’s measurable from a 30-second video clip taken during an application process.
Because I’ve heard the sceptical questions in every room I’ve presented this in, let me address them directly.
It is not a lie detector. Polygraphs attempt to detect deception through physiological arousal — a model that is scientifically disputed and legally inadmissible in most jurisdictions. InsightGenie measures stable personality traits encoded in the prosodic structure of speech. These are dispositional characteristics, not arousal states. The science is fundamentally different, and so is the purpose.
It is not surveillance. All data collection is with the applicant’s full knowledge and consent. The assessment is an explicit step in the application process, not a passive background check. InsightGenie does not access data the applicant has not consented to share.
It does not penalise people for how they sound. This is an important distinction. The biomarkers InsightGenie analyses are not about accent, dialect, language, or speaking style in any cultural sense. They are physiological properties of speech — generated by the autonomic nervous system — that correlate with stable personality traits across languages and cultures. The model has been validated across diverse linguistic populations.
It does not replace the credit officer or underwriter. InsightGenie’s behavioural score is an input to a credit decision, not the decision itself. Human oversight remains central to the process. What changes is the quality and breadth of information that human oversight is working with.
The application process requires minimal friction. A loan applicant provides a short voice recording and a brief video during the application — steps that integrate into any digital or in-person lending workflow. The platform generates a behavioural risk profile and default probability score within minutes.
For applicants with credit histories, the behavioural score provides an additional signal that can confirm or challenge the bureau-based assessment. For applicants without credit histories, it becomes the primary signal — enabling credit decisions that would otherwise be impossible.
The ROI case is compelling. To illustrate with modelled estimates based on conservative assumptions: for a lender offering loans at standard emerging market margins, a conservative 20% increase in the number of loans that can be safely extended — enabled by better risk discrimination across a previously excluded population — generates returns that significantly exceed platform cost. The reduction in non-performing loans delivers further value: a 30–50% NPL reduction directly translates to preserved capital, reduced provisioning costs, and improved regulatory ratios.
These are modelled estimates, not audited financial results. But the platform-level metrics — Gini improvement of 65–77%, NPL reduction of 30–50% — are documented across real deployments. The directional logic is sound.
→ See how InsightGenie’s credit risk platform works — book a discovery call
Commercial banks use InsightGenie to expand their addressable market without expanding their risk exposure. BMW Singapore and EON Capital Malaysia both engaged InsightGenie to complement their existing credit assessment for auto loan applicants with behavioural signals that traditional methods couldn’t provide. Canadia Bank Cambodia deployed the platform to assess borrowers for whom bureau data was insufficient or absent.
Fintech lenders and NBFCs are often the first formal financial institution a new borrower encounters. They have the most to gain from accurate assessment of thin-file and no-file applicants, and InsightGenie’s platform is designed for their scale and speed requirements. HubTel in Ghana and CardX in Thailand have both deployed InsightGenie’s digital footprint assessment as a complement to their existing credit models.
Insurance companies are beginning to apply behavioural profiling to underwriting risk classification — identifying applicants whose personality profiles predict lower claims behaviour. Conscientiousness and emotional stability correlate with better health management, lower accident rates, and more responsible risk behaviour generally.
Microfinance institutions extending small loans to rural or urban micro-entrepreneurs — the deepest end of the financial inclusion challenge — find that InsightGenie’s voice-based assessment, accessible through any smartphone in any language, provides a scalable pathway to creditworthiness assessment for populations that no other tool can reach.
If you’re evaluating alternative credit scoring solutions — whether InsightGenie or any other provider — these are the questions that separate genuinely differentiated science from repackaged data products:
1. What is the tool’s predictive accuracy, and how was it validated? Ask for Gini coefficient, KS statistic, or AUC figures measured against a labelled out-of-sample dataset. “Our clients see lower default rates” is not a validation. A Gini figure with a methodology note is.
2. Can it generate a reliable score for applicants with no credit history? Many alternative data products still depend on some existing financial signal. Ask specifically: what is the accuracy of the model on zero-file or thin-file applicants? This is the core use case and the answer will be revealing.
3. How has the model been tested for demographic bias? A scoring model that systematically under-scores applicants from particular demographic groups based on proxies is both ethically unacceptable and a regulatory liability. Ask for the disparate impact analysis.
4. Is the data collection GDPR-compliant and locally regulated? Data privacy requirements vary by jurisdiction. Ask specifically how consent is obtained, how data is stored, and which regulatory frameworks the tool is compliant with in each market you operate in.
5. Does the scoring signal degrade when applicants know they’re being assessed? For self-report tools, the answer is yes — and the vendor knows it. For biometric and behavioural tools that operate on physiological signals, the answer should be no. Push on this distinction directly.
The World Bank estimates that 1.4 billion adults worldwide are excluded from formal financial services. That’s not a fixed number — it’s a failure of measurement. The people in that figure are not all high-risk. Many of them are among the most creditworthy people in their communities, with decades of demonstrated financial responsibility in informal systems that formal finance doesn’t recognise.
InsightGenie’s platform doesn’t change who those people are. It changes what lenders can see about them.
Across 10 countries, with over 5 million credit assessments completed, the results are consistent: behavioural AI narrows the measurement gap substantially. It finds the creditworthy within the excluded. It reduces NPLs for lenders who adopt it. And it does so while maintaining the ethical standards — bias mitigation, data privacy, human oversight — that responsible financial services require.
The technology exists. The science is validated. The business case is documented. The remaining question is how quickly financial institutions decide they want to see the customers they’re currently turning away.
Ready to explore what InsightGenie can do for your credit assessment process?
Book a discovery call — we’ll walk through your specific use case and show you what our assessment generates in practice. No pitch deck. Just the data.
InsightGenie is a Behavioural AI platform headquartered in Singapore, backed by HSBC, SOSV, Artesian, and Vynn Capital. Active across 10 countries with over 5 million assessments completed.
What is alternative credit scoring? Alternative credit scoring uses non-traditional data sources — such as voice behaviour, digital footprint patterns, and physiological signals extracted from video — to assess a borrower’s creditworthiness when traditional credit bureau data is insufficient or absent. It is particularly valuable for assessing thin-file and no-file applicants in emerging markets where the majority of the adult population lacks a formal credit history.
How accurate is voice-based credit assessment? InsightGenie’s voice-based models achieve 82–93% predictive accuracy for loan default probability, with a Gini coefficient improvement of 65–77% over traditional methods. These figures are documented across real deployments in multiple markets across Asia and Africa.
What is a Gini coefficient and why does it matter? The Gini coefficient (also called the Gini index) measures how well a credit model distinguishes between borrowers who repay and those who default. A Gini of 0% means the model is no better than random; 100% means perfect discrimination. Most mature bureau-based models target 50–70%. InsightGenie achieves 65–77% on populations with no prior credit history — a result that traditional approaches cannot replicate.
How does the assessment work for applicants who don’t speak the lender’s language? InsightGenie’s platform is language-agnostic. The prosodic biomarkers it analyses — rhythm, pitch variability, stress patterns, temporal structure — are physiological properties of speech that exist consistently across all languages. The model has been validated across multiple Asian and European language populations.
Is this compliant with financial regulations in emerging markets? Yes. InsightGenie operates within the regulatory framework of each country it serves, including compliance with data protection laws (GDPR, PDPA, and local equivalents), consent requirements, and applicable guidelines on model risk and algorithmic decision-making. All data is processed on AWS private cloud infrastructure with AES-256 encryption.
Does this replace credit bureau data? No — it complements it. For applicants with credit histories, InsightGenie provides an additional behavioural signal. For applicants without credit histories, it provides the primary assessment signal, enabling credit decisions that would otherwise be impossible.
How is demographic bias prevented? InsightGenie builds de-biasing methodology into model development from the outset. The rPPG component has been validated across the full Fitzpatrick scale of skin phototypes, demonstrating consistent accuracy regardless of melanin content — a deliberate design requirement for the markets InsightGenie serves. Regular bias audits are conducted across demographic groups to monitor for disparate impact.
How quickly can a lender deploy InsightGenie? InsightGenie’s pre-built models are operational within days of integration. The platform is API-ready and designed to integrate with existing lending workflows and core banking systems without significant technical overhead.
