Is false positive reduction reshaping transaction monitoring?

In the rapidly evolving financial security world, transaction monitoring stands as a critical defence against fraud and illicit activity. However, the persistent challenge of false positives—legitimate transactions flagged as suspicious—has long strained resources and efficiency. Now, innovative approaches are transforming this landscape, promising sharper accuracy and streamlined operations.

In the view of Laurence Hamilton, CCO at Consilient, false positive reduction has been a hot topic for many years. While some larger organisations have made in-roads in improving false positive alerts, many still struggle.

He exclaimed, “In smaller organizations without technical expertise and, in many cases, the required data, the challenge of false positive reduction remains significant. With regulatory oversight remaining at high levels and scrutiny strong, ensuring no risks are missed, this plays against the ability to reduce false positives. Alongside this challenge is the ability to change existing technology in terms of cost, risk, and will. “

Hamilton detailed that it was ‘hard to think’ of another industry that is still operating with such a high level of false rates – often referenced as high as 98%. “Even some of the world’s largest banks struggle with legacy systems and with transaction monitoring systems being blunt tools and with data silos,” he said.

All this, Hamilton stated, adds up to making false positive reduction utilising traditional transaction monitoring extremely difficult. “Certainly, it is fair to say that organizations are extremely conscious of the operational effort and costs associated with the false positives driven by transaction monitoring, and reducing these, while not an overriding priority, is ongoing. But the good news is that AI- machine learning is beginning to not only reshape transaction monitoring but also have a positive impact of false positives,” he remarked.

Are AI models reliable enough to slash false alerts? On this point, Hamilton believes that ML models have an essential role to play in false positive reduction.

He said, “Machine learning models are far better than traditional transaction monitoring rules. These rules are static — they catch only the specific patterns they’re told to look for (e.g., “flag transactions over $10,000”). At the same time, machine learning models can detect subtle, complex, or evolving patterns of suspicious activity. Machine learning models can analyse behaviour over time and look for specific patterns that are unusual.”

However, In-house machine learning models have limitations in this regard. The most potent models learn from broad and deep data sets. The challenge here is that financial crime detection in each institution is a rare event, meaning there is limited data on which to train a model. There is more data in larger institutions, but this remains limited to that organization and ignores behaviours occurring elsewhere in the ecosphere.

The other challenge with machine learning is that they are only as good as what you teach it and the amount of data available to teach it. If you only ever teach the model in a closed loop process, then the model doesn’t evolve.

To significantly impact false positive rates and reliably reduce false positives across the entire financial ecosystem, machine learning models must be trained and validated on diverse, representative data.

Hamilton said, “Training models only within an organization have limitations. Large banks may see high transaction volumes, international activity, and complex client profiles. Whereas smaller banks may have fewer transactions but different patterns (e.g. local/regional behaviour). A model trained only on large-bank data might flag small-bank activity as “anomalous”.

“Federated machine learning solves many of the limitations of just using machine learning built at one institution. It transforms transaction monitoring and reduces false positives by letting multiple banks collaborate on model training without ever sharing raw data,” he added.

The result of this, the Consilient CCO stressed, is a smarter, more well-rounded model that sees more patterns and reduces false positives. “Think collective intelligence without compromising privacy and more diverse data equals better pattern recognition. Federated machine learning captures risk signals from multiple environments, bank sizes, and customer types. That diversity improves model accuracy and helps filter out false positives that only look suspicious in one narrow context,” he detailed.

There are still key trade-offs to consider. “In AML and banking in general, an industry with tight regulations and huge reputational stakes, banks must carefully tune their systems. Too many false positives can overload teams and frustrate customers. But too few, and the risk of non-compliance skyrockets.

“From a regulatory perspective, banks seek to limit risks to their business, and therefore, the higher volume of alerts that are investigated can demonstrate strong controls. But as we all know, and a particular case in point here, quantity doesn’t always equal quality,” said Hamilton.

For Hamilton, the trade-offs are fairly straightforward when reducing false positives. “A pro is less time wasted by compliance teams on non-issues. A con is real suspicious activity  might be missed.”

Another pro for Hamilton is better customer experience as well as lower operational costs. The respective cons for these are increased risk of non-compliance with regulators and possibility of regulatory fines, reputation damage and aiding criminal activity. Another pro is avoiding alert fatigue in analysts, however a con is that the model becomes too lenient, leading to under-reporting SARs.

These trade-offs can be mitigated for Hamilton through introducing machine learning and moving beyond traditional TM systems and through a well-structured and defined Risk-Based Approach (RBA)

He explained “For example, organizations can focus more scrutiny on high-risk customers and transactions and apply lighter screening to low-risk ones. This reduces FPs without compromising overall detection quality.”

The potential for transforming the day-to-day for compliance teams is not misunderstood on Hamilton – who sees that reducing false positives doesn’t just change what compliance teams look at, it transforms how they work, where they focus, and what skills they need.

“Instead of clearing hundreds of routine alerts, analysts focus on fewer, higher-risk cases. Their work becomes more investigative and strategic, not just mechanical,” Hamilton said. “With federated learning models in place, compliance professionals increasingly interact with dashboards, trend data, and ML outputs. Analysts become analysts in the truest sense — not just alert processors. Compliance teams are evolving from manual reviews to being more data-literate so that they can interpret, challenge, and optimize these models.”

For Hamilton, false positive reduction should be the driver of an organisation’s approach to transaction monitoring – identifying existing and new risks and unusual activity should be the aim of all organizations.

“Federated machine learning, which allows organizations to collaborate to identify risks, emerging patterns and other financial crime activity can transform both transaction monitoring and materially reduce false positives.”

He continued, “By incorporating federated learning, models can be trained on a richer, more representative dataset, improving their ability to distinguish between normal and suspicious behavior more accurately. This leads to more relevant and higher-quality alerts, reducing the operational burden on compliance teams and allowing them to focus on actual risks. Ultimately, federated learning offers a path toward more innovative, more proactive AML strategies while respecting the data sovereignty and confidentiality required in the financial sector.”

Transformed thinking 

Addressing false positives in 2025 – when the amount being generated are potentially higher than ever before – is a crucial issue for many compliance teams.

For David Caruso, VP of financial crime compliance at Workfusion, the approach to addressing false positives is outdated, and it’s time for the financial crime compliance industry to change its thinking on this issue.

“AI solves the operational and compliance risk problems caused by false positives, but it does so differently than the methods used for the past 15 years,” said Caruso.

The so-labelled problem of false positives is that they create too much low-value work that ends up bogging down AML and fraud teams. This, Caruso claims, is costly – requiring the hiring of too many analysts or the outsourcing of work, and risky – backlogs of work are ‘radioactive’ with regulators, and the distraction of focusing on low-value matters can mean high-value matters can be missed.

Caruso explained, “The approach to reducing this low-value work was to hire consultants or internal teams to “tune” systems and reduce false positives. Then, software companies began touting applications and “data science” to reduce false positives. Much of this work helped marginally, eliminating a small percentage of false positives, but it did not solve the problem.

“AI has transformed this approach,” detailed Caruso.  “Instead of merely reducing false positives, AI agents can now review and resolve millions of alerts generated by financial institutions every day. One AI agent can perform the tasks of entire alert review teams, accomplish them in seconds, and do so consistently. Speed and consistency in this area are always challenging.”

For Caruso, he strongly believes that AI agents are more reliable than humans in reviewing and resolving false alerts. He exclaimed that alert review work is repetitive and monotonous, with employee turnover in these positions high.

“AI agents, or models, do not tire, do not deviate from procedures, and do not quit when the drudgery becomes overwhelming. The models mimic the work of qualified, experienced human analysts, and all the work can be reviewed by human quality control teams if desired. There is no barrier to their use when models are transparent and easy to explain,” said Caruso.

The old approach of tuning systems was always prone to the risk of ‘screening out’ the ‘true positives’, which is matches that required review but were never presented. “This fear led tuning teams to devote considerable time and money to “below the line testing,” meaning looking for that one alert that will not generate if thresholds were changed. This conundrum presented AML and Fraud executives with the unpleasant choice of enduring more low-value alerts, hoping to catch everything necessary, or reducing alert volumes and costs but risking compliance failure,” said Caruso.

“The new AI Agent approach eliminates this risk,” he added. “Now, AML and Fraud teams can actually “open the aperture” wider and produce more alerts if they’d like. Adding more alerts to the AI Agent’s workload is easy and ensures greater coverage.”

The day-to-day impact this has on compliance teams can be massive. For Caruso, a leadership team now no longer needs to worry about rising costs, replacing employees who have resigned, training new staff or hiring contractors.

“Programs are now immediately strengthened. Quality improves, remains high, and never declines. Work backlogs are no longer a concern. Attention and resources can be directed elsewhere to more pressing matters. With the additional time, leaders can implement new processes they have desired but could not find a budget for,” remarked Caruso.

The AI Agent approach for Caruso establishes what he believes is an entirely new standard. “With less distraction from high volumes of low-value matters, the expectation will be that people are now focused on identifying, investigating, and mitigating actual risk. AI Agents will enhance financial crime compliance in numerous ways. Leaders should anticipate that the standards to which they and their program are held will increase,” he concluded.

Improved detection

Meanwhile, Harsh Pandya of Saifr said, “AI can analyze complex transaction patterns and contextual data to find patterns and anomalies. Neither simple nor complex rules-based systems are flexible or comprehensive enough to compete with AI. But performance needs to be measured in terms of false positive reduction and true positive discovery. Indeed, decreasing obvious false positives is essential to alleviate the burden on compliance teams.

“However, it is arguably more important to improve detection of true positives. Singular focus on reducing false positives can lead to selection of simplistic models that minimize alerts but overlook complex patterns of high-risk activities. Compliance analysts should then focus their expertise to scrutinize those complex cases featuring subtle, suspicious behaviors warranting human investigation. And alerts should possess sufficient indicators of potential risk, justifying the allocation of skills human resources for further analysis.”

He concluded, “Consider a compliance team that usually receives 100 alerts per day, on average 90 of which are false positives. If we believe false positive reduction is the goal, we’d want a precise AI system to bring that 90 down to 50, 30 or maybe 10. Is that success? An effective AI system may discover 20, 30, or maybe 40 true positives while still producing 90 false positives that are difficult to resolve. Compliance executives need to ask themselves what the cost is of reviewing an additional false positive versus the cost of missing an additional true positive and make choices based on that tradeoff.”

Achieving the balance

From the standpoint of South Africa RegTech RelyComply, given the sheer influx of transactions faced by each varying arm of a financial institution, consistent monitoring is paramount to identify laundering activity, all with an accuracy rate that could lead to diminishing returns for criminals.

The firm explained, “In traditional rules-based systems, these rates can be obscured by a proliferation of false positive alerts—misidentified as ‘suspicious’—that waste valuable hours better spent investigating true cases of malicious activity from bad actors infiltrating the system.”

The trouble for compliance teams, the company said, is that minimising these false positives can come at the cost of lowering false negative rates—the opposing error measurement sees suspicious cases wrongfully denoted as safe and, therefore, being missed by analysts.

Often, new-look transaction monitoring instead sees compliance teams adopting machine learning models to predict outcomes and lessen the impact of this trade-off – an intelligent supervised approach, RelyComply remarked, requires training models on historical data and significantly reduces false positives while maintaining accurate detection.

“For instance, creating various machine learning models can determine success per specific AML use cases, such as activity occurring in dormant accounts or excessive cash withdrawals. By utilising one model per AML rule, each is tailored to reduce false alerts caused by their corresponding rule. Model explainability tools, such as SHAP (Shapley Additive exPlanations), further aid compliance teams by clarifying the predictive impact of specific features, enhancing interpretability and confidence in outcomes,” said RelyComply.

However, performance success can vary from team to team, influenced by the accuracy of labelled data, the quality of AML data used, and the clarity of model predictions. RelyComply remarked that robust cross-validation protocols and statistical tests ensure that improvements in reducing false positives are statistically significant, providing robust validation for the efficacy of these ML approaches.

The company added, “Robust real-time transaction monitoring is crucial to maintaining AML compliance across banks and fintechs and highlights the interoperable roles of human analysts and automated systems. While machine learning-based models are not a panacea, informed, explainable, and well-validated usage can significantly reduce false positives. Automated capabilities can save hundreds or thousands of analyst hours a month and, over time, are set to become standard practice for eliminating false positives across many integrated financial systems, keeping compliance teams and measures rigorously on track.”

Lack of confidence

For Lucie Novotna, AML & fraud lead solution engineer at Resistant AI, as to whether AI models are reliable enough to slash false alerts – this is an area she is in agreement with.

She said, “Definitely. The question that holds back AI models is not their performance. Some of our own models have false positives of just 4% while increasing coverage of risks. But it is an odd question to ask when the alternative is rules-based transaction monitoring systems with consistent 95% false positive rates.

“This suggests reliability isn’t really an issue—rather the real challenge for AI models is a general lack of confidence in defending their use to regulators,” said Novotna. “And it’s an understandable reaction given the personal liability involved in deploying them. But many regulators are now pushing for them because even they know just how much more efficient they are. That’s why an approach that allows you to slowly replace subsets of rules targeting specific typologies with tailored AI models the way we do at Resistant AI is such a powerful way for compliance leaders to grow that confidence.”

There is undoubtedly a trade-off between fewer false positives and missed risks. However, as to what this trade-off is, Novotna believes there’s no magic answer.

She detailed, “It’s a risk appetite question of precision (how often is the alert correct) vs recall (how much risk are you catching). There will always be residual risk, but the right detection mechanism should allow you to cover all the known angles and a bit more, while being right more often than it is wrong.”

According to the Resistant AI solution engineer,  the core of the issue is an insistence to pretend rules-based systems with 95% false positive rates and overall lower detections are considered the “safe” alternative to AI. “For one of our customers, 70% of their rules-based alerts were generated from events and behaviors their analysts had already reviewed and discounted. That’s not a false positive problem—that’s cruelty,” she remarked.

The impact of such a properly deployed AI solution for compliance teams – surrounding areas such as analyst productivity, motivation, well-being and attrition – is ‘massively positive’ believes Novotna.

She explained, “One of our customers has seen their analyst productivity 5x: the alerts they deal with are more complex, regroup several events their old system would have treated separately with added context to boot, and triage them all so that they are routed to analysts of appropriate seniority. They are burning through their backlog while more efficiently reporting their SARS—and their teams feel like crime fighters again, not false positive detectors.”

Can false positive reduction set a new standard for monitoring? Novotna sees false positive reduction as a ‘surface-level metric’ that the transaction monitoring industry has run into the ground.

She concluded, “The real measure of success for next-generation monitoring engines ought to be increased analyst productivity & efficiency (how much more can you accomplish with the same team), and reduced analyst attrition (how well are you retaining that team).”

Timewaster

In the viewpoint of Joseph Ibitola, growth manager at Flagright, false positives have long been the biggest timewaster for compliance teams. “AI has gotten better at cutting the noise. The goal is smart reduction, without sacrificing real risk detection,” he said.

He continued, “Is there a trade-off? Always. If you cut too deep, you miss real threats, but if you cut too shallow – your team drowns in alerts. If it is done right, fewer false positives means faster case resolution, less compliance fatigue, and stronger team focus on what really matters.”

Ibitola remarked that he believed this shift is ‘quietly setting a new gold standard’ in transaction monitoring.

He concluded, “We’re already seeing it with Flagright customers – leaner compliance teams, sharper insights, and smoother audits.”

Keep up with all the latest RegTech news here