Financial inclusion and artificial intelligence: what algorithms actually unlock — and what they risk reproducing.
There is one statistic the banking sector likes to quote, and another it mentions rather less often.
The first: 79% of adults worldwide now hold an account — at a bank, another financial institution, or a mobile money provider — up from 51% in 2011. In low- and middle-income economies the figure reaches 75%. It is a historic advance, driven above all by the mobile phone.
The second: 1.3 billion adults remain entirely outside the formal financial system. More than half of them live in just eight countries — Bangladesh, China, Egypt, India, Indonesia, Mexico, Nigeria, Pakistan. Women account for 55% of that population. Add the accounts that have been opened but lie dormant, and the number of people excluded or poorly served is closer to 1.6 billion.
Put differently: the wave of digitalisation did the easy part. It banked the most accessible, most urban, most documented populations. What remains is structurally harder — and that is precisely where artificial intelligence is now being presented as the next breakthrough. Rightly so, but not unconditionally.
1. Why the classic banking model stalls at the “last billion”
Three constraints explain most of the exclusion, and none of them is ideological. They are economic.
The data constraint. A bank does not lend to someone it cannot assess. Traditional scoring rests on a credit history, payslips, a central credit register. The informal trader, the market vendor, the ride-hailing driver, the young graduate with no banking past: all are statistically invisible. They are not judged risky — they are unscoreable. That is a different thing, and a far worse one.
The unit-cost constraint. Opening an account, verifying an identity, processing an application, handling an incident: the cost is roughly identical for a €200 loan and a €200,000 one. On small amounts, the margin does not cover the process. The poor customer is not excluded out of malice; they are excluded by arithmetic.
The distance and language constraint. Physical distance from the branch, but above all cultural distance: forms in the official rather than the vernacular language, contractual jargon, low financial literacy. The 2025 Findex is a reminder that distrust is a major brake — 22% of unbanked adults in low- and middle-income countries say they do not trust financial institutions, and close to one in three in Latin America.
AI matters to inclusion only insofar as it attacks these three constraints. So let us look at where it actually does.
2. Where AI genuinely changes the equation
Alternative scoring: rating those with no history
This is the most transformative use. Machine learning models can extract a creditworthiness signal from data never designed for the purpose: the regularity of rent and utility payments, telecom top-up history, mobile money flows, the seasonality of a trader’s receipts, aggregated transaction data.
The effect is twofold. We move beyond the binary of “complete file / no file”, and we replace a static snapshot with a dynamic assessment. A micro-entrepreneur whose revenue is irregular but growing becomes legible — where a classic score rejected them for want of a permanent employment contract.
In East Africa, India and the Philippines, this mechanism has already opened very small credit lines to populations formal credit had never reached. And here is the point that matters: every repayment builds a history. The first micro-loan is a ticket into the system.
Compressing the cost to serve
Automating onboarding and KYC (document recognition, biometric verification, liveness checks) drives customer acquisition cost down from tens of euros to a few cents. That is what makes a low-balance account viable.
The same logic applies to fraud and money-laundering detection: finer behavioural models reduce false positives, and therefore unwarranted account freezes — a badly underestimated cause of attrition among lower-income customers, for whom an account frozen for ten days is a cash-flow disaster.
Lifting the language barrier
This is the most recent contribution, and probably the most underrated. Language models now make it possible to interact in Wolof, Hausa, Swahili or Creole, by voice, with someone who cannot read. A voice assistant that can explain an APR, read out a statement, or warn before an overdraft solves a problem thirty years of educational leaflets did not.
This is the shift from access to use — the real frontier of financial inclusion. Opening an account achieves nothing if nobody uses it.
3. The flip side: AI can automate exclusion just as well
Let us be clear: nothing in the technology guarantees an inclusive outcome. Four risks deserve to be named.
Inherited bias. A model trained on a bank’s history learns that bank’s past decisions — including its implicit preference for stable, urban, salaried, male profiles. It does not “correct” discrimination: it learns it, systematises it and scales it, with the reassuring authority of a number. The risk is not a racist or sexist algorithm; it is the proxy. Postcode, handset model and mobile operator are excellent statistical substitutes for origin or income level.
The hidden price of data. Alternative scoring works because it observes a great deal. Where does legitimate observation stop? Location data, contacts and browsing habits carry real predictive value — and an equally real democratic cost. Inclusion paid for by permanent surveillance of the poorest is not inclusion.
Opacity and the absence of redress. A customer refused credit is entitled to an intelligible explanation and to a right of challenge. An unexplainable model turns refusal into a verdict without appeal, with no way to correct an erroneous data point. This is a question of dignity as much as of compliance.
The digital divide. Around 900 million unbanked adults own a phone — but not all have a smartphone, a stable connection or a digital identity. Any fully digital journey, however elegant, mechanically leaves people out. The assisted human channel remains indispensable.
4. The regulatory framework: a reprieve, not a dispensation
In Europe, creditworthiness scoring of natural persons is classified as “high risk” by the AI Act (Annex III). The associated obligations — risk management, governance and bias testing on training data, technical documentation, logging, transparency, effective human oversight — were due to apply from 2 August 2026.
The timetable shifted this summer. The “Digital Omnibus”, in force since late July 2026, pushes application of the high-risk regime for Annex III systems back to 2 December 2027 (and to 2 August 2028 for AI embedded in regulated products). The Article 50 transparency obligations, by contrast, have applied since 2 August 2026.
It would be unwise to read this delay as an invitation to wait. Bringing a decision model into compliance — inventory, traceability, fairness testing, a redress mechanism — is measured in quarters, not weeks. And above all, sectoral banking regulation, the GDPR and supervisory expectations (ACPR, EBA) continue to apply with no delay at all.
5. Five principles for genuinely inclusive AI
For institutions that want AI to be a lever for inclusion rather than an accelerator of social sorting:
- Measure inclusion the way you measure risk. Track acceptance and pricing rates by gender, age, geography and employment status. What is not measured does not improve.
- Test the proxies, not just the variables. Removing gender from the model is not enough if ten variables still reconstruct it.
- Guarantee a right to human redress. Every adverse decision must be explainable in one comprehensible sentence and challengeable before a person.
- Design for the weakest channel. If the journey does not work by voice, in a local language, on an entry-level handset, it is not inclusive.
- Use alternative data sparingly. The question is not “what can I collect?” but “what would I accept being collected about me?”.
Conclusion
Financial inclusion has never been a problem of goodwill. It was a problem of cost and information — and AI attacks both dimensions head-on. For the first time, serving a customer with an average deposit of €50 can be profitable, and assessing someone with no history can be rigorous.
But the same technology that makes it possible to say “yes” to millions of people also makes it possible to say “no” faster, on a larger scale and more quietly than any credit committee in history. What will separate the two trajectories is not model performance. It is the governance choices made upstream: which data, which tests, what transparency, what redress.
Technology has removed the excuse of economic impossibility. What remains is responsibility.
Sources
- World Bank, Global Findex Database 2025 (July 2025) — survey of ~145,000 adults across 141 economies.
- CGAP, Findex 2025: Why Financial Inclusion Needs to be More Responsible (2025).
- Accion, analysis of the Global Findex 2025 (July 2025).
- Regulation (EU) 2024/1689 (AI Act), Annex III, point 5(b) — credit scoring.
- “Digital Omnibus on AI” — in force 27 July 2026, deferring Annex III high-risk obligations to 2 December 2027.
- ACPR / AEFR, work on machine learning and new data sources for credit scoring.
Leave a Reply