How Tala built a credit engine for the world's most invisible borrowers

When Shivani Siroya founded Tala in 2011, she was trying to solve a problem that conventional credit scoring had simply declared unsolvable: how do you underwrite someone who has never had a bank account, a utility contract in their name, or a loan of any kind? The thin-file customer, in a mature market like the US or UK, is an edge case. In Kenya, Tanzania, the Philippines, and India, that description fits the majority of the adult population. Traditional FICO-style models need a minimum history to generate any score at all. Tala's first market, Kenya, had mobile-money penetration above 70 percent but formal credit bureau coverage well below that. The addressable population was enormous; the underwriting data, by conventional definitions, was absent.

What Tala actually built

Tala's approach starts from a simple technical premise: a smartphone is a transaction ledger, a behavioral diary, and a social-network mapmapUsing software to automate repetitive marketing tasks and campaigns, enabling personalisation at scale across channels like email, web, and social.View full definition →, all at once. With explicit user consent obtained at onboarding, the Tala app requests access to a specific set of on-device signals: SMS transaction records from M-Pesa, call logs, app usage patterns, and metadata about contacts (not content, but frequency and reciprocity of communication). The company then runs these raw signals through a proprietary model to generate a credit score where none previously existed.

The feature engineering is where the real work sits. Tala's data scientists found that the ratio of incoming to outgoing M-Pesa transactions was predictive of repayment behavior, as was the diversity of counterparties a borrower transacted with. Someone who receives money from many different people, rather than a single employer, often has a more resilient informal income stream than a narrow transaction history would suggest. The stability of a contact network over time, meaning low churn in who a borrower calls, also correlated with reliability. None of these variables appear in any credit bureau.

This is not just feature selection; it is an entirely different theory of creditworthiness. Where bureau-based models measure a borrower's past behavior with formal financial institutions, Tala's model measures economic participation in the informal economy.Understanding what payment flows and contact patterns actually signal about a borrower's financial life requires building that analytical capability from scratch, because no vendor can sell you a pre-packaged proxy for M-Pesa reciprocity ratios.

Tala also invested heavily in model governance from an early stage, and for good reason. Lending to underserved populations in emerging markets triggers scrutiny from multiple directions simultaneously: consumer protection regulators in each operating country, global investors wanting evidence that loan pricing is defensible, and civil-society groups alert to any pattern that looks like it targets the vulnerable. The company publishes what its models consider and what they do not, and it explicitly excludes variables that could function as proxies for race, religion, or ethnicity under local fair-lending frameworks.Getting this wrong, even unintentionally, creates the kind of adverse action liability that can stop a lending operation entirely.

The results

By 2026, Tala has disbursed more than 9 billion dollars in loans across its markets, according to company-published figures. It claims a customer base of over 8 million borrowers, the majority of whom had no prior formal credit history. Those numbers come from Tala itself, so treat them as indicative rather than independently verified.

What is verifiable from external reporting is the default rate benchmark. Tala has consistently cited default rates in the 7 to 10 percent range for its emerging-market portfolios, which compares favorably with informal moneylender losses in the same markets, though it is higher than the sub-3 percent defaults a well-collateralized prime lender would accept. The comparison that matters is not against a prime portfolio; it is against the previous option available to these borrowers, which was no formal credit at all, or a loan shark charging 200 percent annualized.

The model has also improved over time as the dataset has grown. Repayment behavior from earlier cohorts feeds back into the scoring model, which means Tala's competitive position compounds in a way that a new entrant without that behavioral history cannot easily replicate. This is the genuine moatmoatA lasting edge over competitors: a resource, capability or position they cannot easily replicate, letting a firm earn above-average returns over time.View full definition → in alternative-data underwriting: the data flywheel, not the algorithm.

What transfers, and what does not

For a CDO at a fintech operating in more regulated Western markets, the Tala case teaches several things that do not require a Nairobi office.

First, the principle that transactional behavior reveals creditworthiness generalizes, but the specific signals do not. In a UK or EU context, open banking access under PSD2 gives lenders sight of current-account inflows and outflows with a level of detail that M-Pesa SMS parsing only approximates. The analytical question is the same: what does income regularity, counterparty diversity, and spending pattern stability predict about repayment? The legal pipelinepipelineAll active sales opportunities across the stages of the sales process, together with their combined potential value and probability of closing.View full definition → for getting that data is entirely different.

Second, consent architecture is not a checkbox exercise. Tala's app requests specific permissions and explains them in the local language at a fifth-grade reading level. In a GDPR environment, your lawful basis for processing alternative data signals must be documented, purpose-limited, and defensible to a Data Protection Authority. The consent moment is also where you lose borrowers who are uncomfortable with the scope of access you are asking for, which itself becomes a selection signal worth analyzing.

Third, the model governance requirements in Western markets are materially stricter. The EU AI Act's credit-scoring provisions, the FCA's consumer duty obligations in the UK, and the CFPB's adverse action notice rules in the US all require that a declined applicant receive an explanation they can actually act on. Building explainability into a model trained on 10,000 behavioral features is a genuine technical problem, not a compliance footnote.

Fourth, the data flywheel advantage is real but takes time to build. A lender entering thin-file underwriting in 2026 with no existing repayment data will have worse models than one with four years of labeled outcomes. Buying alternative data from a bureau aggregator (FinScore in Southeast Asia, Aire or Nova Credit in Western markets) lets you start faster, but you are renting someone else's signal rather than building your own.

The concrete lesson is this: Tala's advantage is not the algorithm; it is the willingness to define creditworthiness differently and then collect the data to support that definition. Data leaders who want to replicate the result need to start with the same question Siroya started with, which is not "what data do we have?" but "what behavior actually predicts repayment for this population, and how do we get lawful access to it?" The gap between those two questions is where most thin-file initiatives stall.

The full course on this sector:Data in Fintech.