Adviser Hub CloudEnterprise Wealth Tech
Compliance

Nine Transaction Monitoring Buying Mistakes That Bury Analysts in False Positives

Most false-positive problems get locked in when the contract is signed. Here are nine selection mistakes, how to fix each one, and the evidence supervisors in London, Frankfurt, Singapore, Dubai and other FATF-aligned markets expect to see.

By the Adviser Hub Cloud editorial team · November 2026 · 9 min read

Buying on scenario count instead of typology coverage

The most common mistake in a transaction monitoring software selection happens in the first vendor meeting. Someone asks how many scenarios ship out of the box, and the biggest number wins. A library of several hundred rules tells you nothing about whether those rules fit your products, customers and channels. A trade finance bank in Dubai and a payments institution in Lithuania face different typologies. Switch on generic rules all at once and analysts drown in alerts, while the risks that matter still get missed.

What to do instead: start from your business-wide risk assessment. List the typologies it identifies, such as funnel accounts, trade-based laundering, mule networks and rapid pass-through. Then ask each vendor to map its scenarios to each typology and to name the data fields each scenario needs. The UK FCA, the Monetary Authority of Singapore (MAS), the Central Bank of the UAE (CBUAE), EU national authorities and FATF-aligned supervisors from India to South Africa all expect a documented link between your risk assessment and your monitoring coverage. They won't ask about the vendor's catalog. At an inspection, ten well-calibrated scenarios with a clear rationale do better than two hundred that nobody can explain.

Advertisement

Accepting false positive claims tested on someone else's data

Vendors routinely quote false positive reduction figures taken from other clients or from synthetic datasets. Those numbers depend on the other firm's customer mix, data quality and starting point. A bank that already tunes reasonably well won't get the gain a bank running untuned defaults would.

What to do instead: insist on a proof of concept that uses six to twelve months of your own pseudonymized transaction history. Seed it with cases you already know were suspicious, including past suspicious activity reports, so you measure detection as well as noise. Track three numbers: alerts per thousand customers, alert-to-case conversion and the share of known-positive cases caught. If a vendor halves your alerts but misses a third of your known cases, it has cut detection, not false positives.

Check data protection before you share anything. GDPR in the EU, the UK GDPR, Singapore's PDPA, India's DPDP Act and the UAE's federal and free-zone regimes all limit how test data can be moved and processed. Your legal team should sign off on the transfer.

Treating machine learning as exempt from model validation

Behavioral analytics and machine-learning risk scoring can improve alert quality. The problem is that buyers often treat them as black boxes and take the vendor's word that "the model learns" without asking how anyone will validate it. The firm then can't explain to a supervisor why one customer triggered an alert and another didn't.

What to do instead: require model documentation that covers features, training data, performance metrics, drift monitoring and retraining triggers. Plan for independent validation, by internal model risk or a third party, before go-live and at least once a year after that. Ask for explainability at the alert level, meaning the factors behind each score shown in plain language on the analyst's screen.

Reference points differ by market. UK firms often benchmark against the PRA's model risk principles in SS1/23. Singapore firms look to MAS's FEAT principles and its newer work on AI risk management. EU firms should check how the AI Act and the AMLA-led regime affect their governance duties. In every market, treat a vendor that resists giving validators access as a red flag.

Ignoring data lineage from core systems to the monitoring engine

Many published enforcement actions trace monitoring failures back to data, not detection logic. Transactions never reached the system, a product was never mapped, or a field arrived empty. The scenarios worked; they just never saw the activity. Even so, buyers tend to scrutinize the analytics during selection and barely look at ingestion.

What to do instead: ask each vendor how it proves completeness. Strong platforms reconcile record counts and values against source systems daily, flag missing or malformed fields, and keep lineage from the source system through every transformation to the alert. Build a data mapping inventory covering every product, channel and currency before you sign, not during implementation. Supervisors expect firms to show that monitoring covers all in-scope activity, and "we believe it does" is not an answer an MLRO wants to give at an inspection.

Running AML scenario tuning without documented governance

When alert volumes start to hurt, teams raise thresholds. They often do it informally, and sometimes only to clear a backlog. Examiners look specifically for threshold changes made to manage workload rather than risk.

What to do instead: pick software that supports disciplined AML scenario tuning, and check these capabilities in the demo:

  • Above-the-line and below-the-line testing: sampling alerts just under a proposed threshold to show you aren't dropping productive alerts.
  • Segmentation, so retail, SME, corporate and correspondent customers can have different thresholds.
  • A sandbox that runs proposed changes against historical data before they reach production.
  • A full change log recording who proposed each change, the analysis behind it and who approved it, usually the MLRO or a compliance committee.

A common baseline is to tune at least once a year and to run extra reviews when products or risk ratings change. Expectations differ in detail between the UK, EU, Singapore, the UAE and elsewhere, so confirm the cadence your own supervisor expects.

Bolting sanctions screening onto monitoring as an afterthought

Sanctions screening and transaction monitoring solve different problems. Monitoring finds patterns after the fact. Screening has to stop a prohibited payment before it leaves, and in many jurisdictions a sanctions breach carries strict or near-strict liability. Firms often buy a monitoring suite and assume the bundled screening module is good enough. That is a mistake.

What to do instead: evaluate screening as a separate product. Test fuzzy matching on transliterated names from Arabic, Cyrillic and Chinese scripts. Check latency on instant payment rails. Confirm list coverage for the UN, EU, UK OFSI and US OFAC lists, plus local lists such as the UAE's Local Terrorist List. Measure match rates against a seeded file of true matches and near-miss names. Then check integration: screening hits and monitoring alerts on the same customer should land in one case, so analysts don't investigate the same person twice.

Scoring the detection engine and ignoring the analyst's screen

False positives cost money because a person has to clear every one, and buyers often leave case management to the end of the evaluation. Take a hypothetical: a system produces slightly better alerts but stretches each review from 10 minutes to 25. At 2,000 alerts a month, that adds about 500 analyst hours every month, which wipes out any gain from the better alerts.

What to do instead: during the proof of concept, have your own analysts run timed review exercises. Look for:

  • entity resolution that groups linked accounts
  • consolidation of multiple hits on one customer into a single case
  • network views and prefilled KYC context
  • one-click escalation to a suspicious transaction report in your local regulator's format

Every disposition needs a reason code and a timestamp in an immutable audit trail. Better workflow design is often the cheapest way to cut the cost of false positives.

Comparing license fees instead of five-year cost

The license fee is usually the smallest cost you can control. Pricing models include per customer, per transaction, per alert and tiers based on assets. If your volumes grow fast, transaction-based pricing can push costs up steeply, so model year-three and year-five volumes before you compare quotes.

Cost driverTypical 2026 range (indicative, USD)What to check
Annual SaaS license, mid-sized bank or fintech$75,000 to $400,000Volume bands, overage pricing, uplift caps
Annual license, large or tier-1 institution$500,000 to several millionModule bundling, entity-wide terms
Implementation and data integrationRoughly 0.5x to 1.5x first-year licenseFixed price versus time and materials
Independent model validationRoughly $40,000 to $150,000 per cycleVendor cooperation, documentation access
Analyst headcountOften the largest line over five yearsModeled alert volumes and review time

These are typical market ranges. They vary widely by contract, region and scope, and local pricing in markets such as India, Southeast Asia or Latin America can look quite different, so confirm figures with vendors. Established suites such as NICE Actimize, SAS and Oracle usually cost more to implement. Cloud-native vendors such as Napier, Hawk, ComplyAdvantage and Feedzai often deploy faster. Neither label matters as much as how well the platform fits your data.

Signing without outsourcing evidence or an exit plan

Most modern AML compliance software runs as SaaS, so most jurisdictions will treat it as a material outsourcing or third-party arrangement. EU firms have to meet DORA's ICT third-party requirements and the EBA outsourcing guidelines. UK firms work under PRA SS2/21 and FCA SYSC 8. Singapore firms apply the MAS outsourcing and technology risk management guidelines, and UAE banks follow the CBUAE's outsourcing rules. Each regime expects due diligence, audit rights and a credible exit strategy.

What to do instead: settle these terms before you sign:

  • audit and supervisor access rights
  • data export in usable formats, including historical alerts and case notes
  • transition support
  • clear disclosure of sub-processors

FATF-aligned rules usually require monitoring records to be kept for at least five years, and some jurisdictions require longer. You'll need that history long after you leave a vendor.

Decision rule: shortlist only vendors that will run a proof of concept on your data, give validators access to their models and accept supervisor audit rights in the contract. If a vendor refuses any one of the three, you'll struggle to defend choosing it at an inspection.

Frequently asked questions

How much false positive reduction is realistic when switching transaction monitoring software?

It depends on where you start. A firm running untuned default scenarios usually has far more room to improve than one that already segments its customers and tunes regularly. The only reliable estimate comes from a proof of concept on your own historical data that measures both alert volumes and detection of known suspicious cases.

Do regulators accept machine-learning models for AML monitoring?

Supervisors in the UK, EU, Singapore, the UAE and most other FATF-aligned markets generally don't prohibit them. They do expect the firm to understand, validate and explain the models. In practice that means documentation, independent validation, drift monitoring and alert-level explanations an analyst can act on. Accountability stays with the firm, not the vendor.

How often should AML scenarios be tuned?

A common baseline is to tune at least once a year and to run extra reviews when something changes, such as a new product, entry into a new market or a revised risk assessment. Every change needs documented analysis, including below-the-line testing, and formal sign-off. Expectations differ in detail by jurisdiction, so check your supervisor's guidance and get advice from qualified compliance counsel.

Should sanctions screening and transaction monitoring come from the same vendor?

Not necessarily. Buying both from one vendor can make integration and case management simpler, but screening quality should be evaluated on its own, especially name matching across scripts and real-time performance. Many firms pair a specialist screening provider with a different monitoring platform and connect them through shared case management.

What data should we provide for a vendor proof of concept?

Usually six to twelve months of transactions, customer reference data, account data and risk ratings, plus a set of known suspicious cases to measure detection against. Pseudonymize where you can, and confirm the transfer complies with the data protection laws that apply to you, such as GDPR, UK GDPR, Singapore's PDPA, India's DPDP Act or UAE rules. Your legal and data protection teams should approve the arrangement before any data moves.

How long does implementation usually take?

Cloud-native platforms at mid-sized firms often go live in roughly four to nine months. Large multi-entity rollouts of established suites can take 12 to 24 months or longer. Data mapping and integration almost always set the timeline, so the best way to shorten it is to finish your data inventory before you sign.

Written by the Adviser Hub Cloud editorial team. Prices, features and rules change often - confirm details with the vendor or a licensed professional before you decide.