Can New Credit Scoring Be Fairer?

Imagine two small companies with very similar economics.

Both have regular customers.

Both generate cash.

Both pay suppliers.

But only one has a long formal credit history, substantial collateral and conventional financial records.

Traditional assessment may find the second company easier to understand.

That does not necessarily mean the first one is riskier.

It may simply be less visible.

That is where alternative credit scoring promises something important.

Indonesia’s Financial Services Authority is increasingly encouraging richer credit information and scoring to help financial institutions understand borrower profiles and repayment capacity more comprehensively.[1]

But a harder question follows:

Does using more data automatically make lending fairer?

No.

Indonesia already has an alternative-scoring ecosystem

OJK regulates Alternative Credit Rating providers, known as PKA, through Regulation No. 29/2024.[2]

The framework allows relevant alternative data to support creditworthiness assessment, including for people and businesses with limited traditional credit histories.[2]

Its regulatory explanation cites sources such as telecommunications, utilities and e-commerce data.[2]

That potentially changes how thin-file SMEs are assessed.

A business with little bank borrowing history may still generate meaningful evidence through transactions and operating activity.

The industry is already operating at scale

By July 2026, OJK reported eight PKA providers generating roughly 184 million scoring hits, with total assets of around Rp564 billion.[3]

A “hit” is not a borrower or approved loan.

It should not be interpreted as 184 million customers receiving credit.

But the figure indicates that alternative scoring is no longer a niche experiment.

Indonesia’s SME financing rules also explicitly allow PKA to help accelerate assessment, while maintaining governance and risk-management requirements.[4]

The policy objective is therefore clear:

more SMEs should become assessable.

That does not mean every SME should become automatically approvable.

Alternative data can reduce blind spots

The strongest use case involves borrowers with thin conventional files.

A viable small business may lack audited statements, major collateral or years of borrowing history.

Traditional systems can struggle to distinguish between:

“not enough data”

and

“high risk”.

Alternative data can add visibility.

The World Bank notes that technology-enabled screening can help reach previously underserved consumers while potentially improving risk assessment.[5]

BIS research based on a Chinese fintech dataset also found machine-learning and non-traditional-data models outperforming traditional approaches under some conditions examined in that study.[6]

But that evidence comes from a specific market and dataset.

It does not prove every alternative-scoring model is superior.

Better prediction is not the same as fairness

A model can be statistically accurate while creating problematic outcomes.

Data reflect history.

And history can contain structural exclusion.

BIS warned in 2026 that AI systems can reproduce inequalities embedded in past data, potentially scaling unfair outcomes more efficiently.[7]

World Bank guidance similarly highlights biased datasets, poor model design and proxy discrimination among the risks of algorithmic lending.[5]

“Data-driven” therefore does not mean “neutral”.

Proxy discrimination can be hard to see

A model may not explicitly use gender, religion or another sensitive characteristic.

But it may use variables highly correlated with them.

Location.

Device patterns.

Employment type.

Social or commercial network.

Consumption behaviour.

Removing one sensitive column does not necessarily remove its statistical shadow.

The governance question therefore becomes:

Do the model’s outputs create unjustifiable exclusion even when sensitive variables are absent?

Bad data can become bad decisions

Richer data create value only when the data are reliable.

Duplicate transactions.

Outdated phone numbers.

Mixed personal and business accounts.

Identity errors.

Old marketplace profiles.

Incorrect merchant records.

Small errors can become systematic when automated at scale.

NIST’s AI Risk Management Framework highlights reliability, accountability, transparency, explainability, privacy and fairness as important dimensions of trustworthy AI.[8]

Data quality is therefore part of risk management.

A score is an estimate, not an identity

A numerical score can look objective and final.

But it is the product of a model.

It depends on inputs, training data, weights, model architecture and policy assumptions.

The correct interpretation is:

estimated credit risk based on available information.

Not:

“this business is inherently bad”.

That distinction matters.

Explainability allows correction

If an SME is rejected, the natural question is:

Why?

Completely opaque models create problems for borrowers, lenders and regulators.

NIST’s explainable-AI principles emphasise the value of meaningful reasons or evidence accompanying system outputs.[9]

Explainability does not require exposing proprietary algorithms in full.

But the system should support answers such as:

Which factors materially affected the assessment?

Was the input data correct?

Can the result be reviewed?

Is the model operating within the population for which it was designed?

Humans still matter

AI is particularly good at processing hard, codifiable information.

Relationship lending can provide something different: contextual or soft information.

BIS research suggests these approaches can coexist rather than simply replace each other.[10]

A borrower may have temporary weak cash because it bought machinery.

A major order may not yet appear in historical data.

Seasonality may look like instability.

A contract may fundamentally change future cash flow.

Context can matter.

Humans can also be biased

Human judgement is not automatically fair.

Credit officers can carry stereotypes, inconsistent assumptions or overconfidence.

The better model is therefore not:

algorithm versus human.

It is:

auditable models + accountable human judgement + clear governance.

Each layer should be capable of challenging the other.

False negatives matter too

Lenders naturally worry about approving a borrower that later defaults.

That is a false positive from a risk-classification perspective.

But there is another error:

a viable borrower rejected by the model.

The cost is harder to see.

The SME does not expand.

Inventory is not financed.

A machine is not purchased.

An order cannot be fulfilled.

A system can become extremely conservative and appear “safe” while reducing inclusion.

Fair lending requires attention to both sides.

Borrowers need correction channels

If alternative data influence credit access, SMEs need reasonable ways to correct inaccurate information.

Identity mismatches.

Duplicate records.

Old information.

Transactions belonging to someone else.

Incorrect business classifications.

Without correction mechanisms, a data error can become a persistent digital reputation problem.

Consent cannot be treated as decoration

Richer scoring means more data.

More data means greater privacy responsibility.

OJK’s framework includes governance, data-security and consumer-protection requirements.[2]

The relevant question is not simply:

Can we technically access this information?

It is:

Does the borrower understand what is being used and why?

Do not make SMEs optimise for the algorithm

When borrowers learn which signals a model rewards, gaming can follow.

Artificial transactions.

Behaviour designed for scores instead of economics.

Activity that looks healthy digitally while weakening the real business.

Models therefore need monitoring for gaming and unintended incentives.

Credit scoring should interpret the business.

It should not teach the business to perform for the model.

Seven questions lenders should ask

Is the data relevant?

Is the model valid for this borrower segment?

Are outcomes tested across groups?

Can incorrect data be corrected?

Can material decisions be explained?

When is human review required?

Who is accountable when the system fails?

The last question is particularly important.

If nobody owns the decision, governance is incomplete.

What SMEs can do

Small businesses do not need to reverse-engineer scoring models.

A better strategy is to make the business easier to understand.

Separate business and personal accounts.

Reconcile sales.

Maintain clear invoices.

Keep repayment records.

Record obligations.

Document suppliers.

Use consistent business identity.

The cleaner the operating trail, the less assessment depends on assumptions.

More data create opportunity—not certainty

Alternative scoring can help lenders see businesses traditional systems struggle to evaluate.

It can improve speed.

Reduce information gaps.

Support better risk differentiation.

Potentially widen financial inclusion.

But technology does not automatically create fairness.

A poorly governed model can scale exclusion just as efficiently as it scales inclusion.

The right question is therefore not:

“Is this scoring system smarter?”

It is:

“Is it accurate, explainable, correctable and accountable?”

When those conditions are present, alternative scoring can become a bridge.

Without them, finance may simply replace an old gatekeeper with a digital one that is harder to see.