7 criteria for choosing an anti-fraud tool at customer onboarding

In short: a good anti-fraud tool for customer onboarding is judged on seven criteria: the range of fraud covered (identity, documents, businesses, IBAN), the convergence of signals rather than isolated rules, behavioural analysis, the pooling of known fraud, the false positive rate, the explainability of decisions and the impact on the customer journey.

‍

Fraud at customer onboarding has become industrialised: synthetic identities, fake documents generated by AI, shell companies set up in series, hijacked bank details. Blocking one type of fraud while leaving the others open is like closing a door in a house with no walls. Here is the grid for evaluating a tool, and the questions to ask in a demo.

‍

1. The range of fraud covered

‍

First map your fraud: identity theft, fake documents (identity document, payslip, tax notice, bank statement), shell companies, director fraud, hijacked IBAN, multiple-application abuse. Then check what the tool covers natively. A specialist in a single type of fraud leaves you to assemble the rest yourself, with gaps at every seam.

‍

Question to ask: "Of these six types of fraud, which do you detect natively, and which require another tool?"

‍

2. Signal convergence, not isolated rules

‍

No single signal is enough on its own: a recent email, a prepaid operator or a company at a mass registered address are not proof. It is their convergence that speaks. Steer clear of binary rule engines ("block if X") that generate false positives en masse: demand scoring that weighs hundreds of signals and returns its reasons.

‍

Question to ask: "Show me a file where three weak signals converge: how does the score combine them?"

‍

3. Behavioural analysis and the device

‍

Professional fraudsters give themselves away by how they use the journey: automated typing speed, inconsistent browser language, high-risk time zone, several IP addresses within a few hours. Device fingerprinting and behavioural analysis detect automated patterns that no document check can see, without adding the slightest friction.

‍

Question to ask: "Do you collect device and behavioural signals, and can they feed automated rules?"

‍

4. Pooling: benefiting from fraud already seen elsewhere

‍

A fraudster blocked at one player tries the next one, often in the same month. Two mechanisms change the game: multiple-application detection (the same profile approaching several players within a few weeks) and lists of confirmed fraud shared across clients. It is a network effect: the more players in your sector the tool equips, the earlier it sees.

‍

Question to ask: "If a fraudster identified at another of your clients shows up at my door tomorrow, what happens?"

‍

5. The false positive rate

‍

The hidden cost of anti-fraud is the good customer refused or over-checked: lost revenue, saturated analysts, a degraded journey. Demand the false positive rate observed at comparable clients, and a risk-proportionate control mechanism: light for healthy profiles, reinforced for risky profiles only. To compare the figures vendors put forward, you need to understand anti-fraud tool metrics.

‍

Question to ask: "What is your observed false positive rate, and what share of my files will undergo reinforced checks?"

‍

6. Explainability: decide, prove, contest

‍

An unexplainable block is indefensible: in front of the customer who contests it, in front of the auditor, in front of the AI Act. Every alert must be returned with its reasons, and the analyst must be able to decide with context. The tool guides, the human decides on ambiguous cases.

‍

Question to ask: "For this blocked file, list the exact reasons, exportable for an audit."

‍

7. The impact on the journey and integration

‍

An anti-fraud tool that adds 30 seconds of friction drives away more good customers than it blocks fraudsters. Demand a check that is invisible to the legitimate customer (2 to 5 seconds, in the background) and an integration that fits your use: API for real time, a portal for analysts, batch for back-book reviews.

‍

Question to ask: "What exactly does my legitimate customer see during the check?"

‍

The decision grid at a glance

‍

• Coverage: the minimum, one type of fraud; the differentiator, identity + documents + business + IBAN.
• Detection: the minimum, static rules; the differentiator, convergence of hundreds of signals.
• Behaviour: the minimum, nothing; the differentiator, device fingerprint + behavioural analysis.
• Pooling: the minimum, internal list; the differentiator, cross-player multiple-application detection + confirmed fraud list.
• False positives: the minimum, not measured; the differentiator, published rate, risk-proportionate controls.
• Explainability: the minimum, raw score; the differentiator, exportable reasons, analyst at the centre.
• Journey: the minimum, visible step; the differentiator, invisible, 2-5 s in the background.

‍

How Meelo meets these 7 criteria

‍

Meelo covers identity fraud (more than 400 cross-referenced signals, Mobile ID), document fraud (automated analysis, consistency between documents), business fraud (shell companies, director score), IBAN fraud (SEPAmail + VOP) and behavioural fraud (Tracketio: device and behaviour). Multiple applications are measured across players and the Warning List pools confirmed fraud. Decision in 2 to 5 seconds, reasons returned, explainable AI audited by IEEE. Documented results: 93% of fraud identified at customer onboarding, 2.5% false positives, fraud divided by 12 at a financing player. To see these controls at work on your journeys, you can schedule a demo.

‍

FAQ

‍

What are the main types of fraud at customer onboarding?

‍

Identity theft, synthetic identity, document fraud (identity documents, proof of income, bank statements), shell companies, director fraud, IBAN hijacking and organised multiple applications across several players.

‍

What is the difference between a rule engine and fraud scoring?

‍

A rule engine applies binary conditions ("block if prepaid operator"), a massive source of false positives. Scoring weighs the convergence of hundreds of signals and returns a nuanced risk level, together with its reasons.

‍

How can you reduce false positives in anti-fraud?

‍

By cross-referencing more signals (a single signal never blocks on its own), by proportioning checks to the real risk of the file, and by letting the analyst decide on ambiguous cases with the score's reasons.

‍

Does an anti-fraud tool slow down the customer journey?

‍

Not necessarily: multi-signal checks run in 2 to 5 seconds in the background, invisible to the legitimate customer. Only risky files trigger additional verification.