Between November 2024 and the end of August 2026, Indonesia's Financial Services Authority (OJK) recorded 659,419 bank accounts blocked over suspected financial fraud. A newer regulation, POJK No. 12 of 2024 on Anti-Fraud Strategy for Financial Services Institutions, now requires every bank and fintech to run an early-detection system built on internal audits, data analytics, and machine learning, catching suspicious transactions before the money actually moves. OJK has also stood up the Indonesia Anti-Scam Centre (IASC), linking banks, payment providers, marketplaces, and crypto platforms into one coordinated detection and account-blocking pipeline.
The signal is clear. AI-based fraud detection is no longer a future consideration for Indonesia's financial sector, it is a live obligation today. The problem is how many institutions respond to that pressure, reaching straight for an AI vendor or an anomaly-detection model, without first checking whether the underlying data is actually ready to support one.
The Mistake Institutions Keep Making, Jumping Straight to the Algorithm
The pull toward buying or building an AI model first is understandable. Regulatory pressure is real, implementation deadlines are already ticking, and every technology vendor is happy to sell the latest fraud-detection solution. But one assumption rarely gets tested first, whether the transaction data that model will learn from is actually accurate, complete, and recorded on time.
In practice, machine learning models for anomaly detection work by learning what normal looks like from historical data, then flagging anything that deviates. If the data feeding that model is itself messy, unreconciled transactions, unexplained discrepancies, records that land in the system late, the model is not learning real fraud patterns. It is learning operational chaos, and flagging it as if it were fraud.
POJK 12/2024 itself already points in this direction, even if the detail rarely gets read closely. The regulation does not just require financial institutions to have a detection system, it specifies that detection has to run through a combination of periodic internal audits, real-time transaction monitoring, and pattern analysis using approaches like Benford's Law to surface statistical anomalies. All three methods depend on the same precondition, transaction data that is complete, accurate, and recorded on time. Without that, internal audits slow down, real-time monitoring stops being real-time, and statistical pattern analysis is biased from the start.
Why an AI Model Is Only as Accurate as the Data It Eats
The consequence is concrete and immediate for the operations team. With a weak reconciliation foundation, an anomaly-detection model produces too many false positives, legitimate transactions flagged as suspicious simply because the data arrived late or never got matched correctly. Investigation teams end up flooded with alerts that are mostly not real fraud, while the genuinely risky transaction can get buried among thousands of false ones.
The reverse holds too. When reconciliation runs accurately and in real time, with an audit trail logging every match, every exception, and every resolution complete with a timestamp and who handled it, the AI layer sitting on top gets a much cleaner signal to learn from. It is not only about detection accuracy either, a complete audit trail also doubles as compliance evidence an OJK auditor can check directly, with no manual reconstruction required.
Three Layers That Have to Be Right First, Before AI Fraud Detection Actually Works
From our own experience building reconciliation and settlement systems for banking institutions, the order is always the same, three layers that need to be correct one at a time, not attempted all at once from the top.
The first layer is accurate transaction matching and reconciliation. Every transaction from every payment channel genuinely matched against its source data, and every discrepancy given a clear, recorded reason rather than left hanging.
The second layer is a full audit trail and the Four-Eyes principle at every critical point. Every data change, every matching decision, every exception passing through two-party verification and getting logged permanently, traceable at any time.
Only once both of those layers hold does the third layer, AI anomaly detection, get a data foundation clean enough to actually learn the difference between a normal pattern and a suspicious one. Jumping straight to the third layer without the first two in place is like building a third floor without the first two, it stands for a while, but buckles the moment real transaction volume tests it.
A Checklist Before Your Institution Invests in AI Fraud Detection
Before signing a contract with any AI fraud-detection vendor, a few basic questions deserve an honest answer first.
First, is every payment channel your institution uses reconciled automatically and in real time, or is there still a manual process prone to delay and human error.
Second, does every transaction discrepancy and exception have a clear, logged resolution path, or does it still tend to get set aside to check later.
Third, can your current audit trail actually answer who changed what and when, for any given transaction, whenever compliance or an external auditor asks.
Fourth, does your team have enough clean historical data to train or evaluate an AI model, before that model ever touches production.
Fifth, who is accountable for keeping this data foundation clean once the AI system is live, because a model that looks strong on day one can start producing noise as the data quality underneath it quietly degrades.
If most of these answers come back not yet or not sure, that is not a reason to delay the AI investment, it is a signal to reverse the build order, fix the foundation first, then add the intelligence layer on top.
Our team at XETUP builds reconciliation and settlement systems for banking institutions on the same principles, Four-Eyes verification at every critical point, a complete audit trail, and a matching engine built to handle high transaction volume without trading away accuracy. We position ourselves as the infrastructure that makes whatever intelligence layer sits on top of it, whether built in-house or bought from a vendor, worth trusting in the first place.
If your institution is evaluating its reconciliation and audit-trail readiness before stepping into an AI fraud-detection investment, our team is open to an initial conversation, no commitment required.