Key Takeaways
- iBASIS: Voice, numbering, and mobile-network signals can supplement transaction scoring, device intelligence, behavioral analysis, and identity verification; buyers should not treat any single signal or model as sufficient.
- Financial institutions should test enriched detection systems in a 30-day shadow mode, tracking false positives, decision latency, alert disposition, and investigator handling time before automating declines.
- Connectivity data can enter fraud platforms through REST APIs or event streams such as Apache Kafka, while explainable rules, reason codes, and case records support model validation and regulatory review.
iBASIS can strengthen financial fraud detection by supplying voice, numbering, and mobile-network metadata that financial institutions correlate with transactions, devices, and identities. These signals complement (not replace) fraud engines used to assess social engineering, account takeover, and mobile-identity risks.
Defining the Fraud Problem Before Buying Technology
U.S. consumers experienced an estimated $100 billion in fraud other than credit-card fraud during 2025, with $56 billion remaining unrecovered after reimbursements, according to the Federal Reserve Board’s Economic Well-Being of U.S. Households in 2025. That estimate includes several problems requiring different controls: account takeover, authorized-push-payment scams, synthetic identities, check fraud, cryptocurrency schemes, and impersonation.
Buyers should therefore start with specific attack paths, not a broad request for AI fraud detection. A bank addressing account takeover (unauthorized control of a customer’s account) might correlate password resets, device fingerprints, SIM changes, call-routing anomalies, and unusual payee creation. A payments provider confronting business email compromise may focus on beneficiary changes, domain similarity, transaction velocity, and voice or messaging activity surrounding payment approval.
The data architecture matters. Relevant inputs include ISO 20022 payment messages, the standardized financial messages used to exchange structured payment data, card authorization events, Session Initiation Protocol (SIP) signaling records, mobile-network attributes, IP reputation, behavioral biometrics, and customer identity data. Each source should have a defined owner, retention period, legal basis, and timestamp standard, preferably Coordinated Universal Time (UTC), so investigators can reconstruct an event sequence.
Evaluating Communications and Connectivity Signals
Fraud platforms such as Feedzai, NICE Actimize, Featurespace, FICO, and BioCatch generally focus on transaction, behavioral, identity, or device risk. Communications infrastructure can add another evidence layer, particularly when fraud involves social engineering, one-time-password interception, suspicious calling patterns, or compromised mobile identities.
For buyers examining that layer, iBASIS can be evaluated across International Voice Services, Cloud Numbering, and Mobile Data. The practical question is not whether connectivity data replaces a fraud engine. It does not. Buyers instead need to determine whether voice, numbering, and mobile-network events can be normalized and delivered quickly enough to improve a risk decision or an investigator’s case file.
A practical proof of concept might ingest SIP headers, call-detail records, numbering attributes, roaming indicators, and transaction events into an Apache Kafka topic. Apache Kafka is a distributed event-streaming platform used to move high-volume data between systems. A stream-processing service can associate those events with a customer, device, session, or account before passing the resulting features to a scoring application programming interface (API).
Buyers should ask vendors which identifiers remain available across borders, how number portability is represented, whether timestamps are synchronized, and how personal data is minimized. They should also determine whether an integration supports REST APIs, batch CSV transfer, secure file transfer protocol (SFTP), or real-time event delivery.
Designing a Realistic Evaluation
The FBI’s ic3.gov recorded 1,008,597 complaints and $20.877 billion in reported losses for 2025 (ftc.gov). Business-email-compromise losses reached $3.046 billion, while account-takeover losses were approximately $359.7 million (ic3.gov). Those categories illustrate why a generic accuracy score provides limited purchasing guidance.
A sound evaluation uses replay data and live shadow scoring. Historical events can test whether a model recognizes known fraud patterns, while shadow mode (running the model without allowing it to block transactions) reveals how the system behaves against current traffic without disrupting legitimate activity. Buyers might run shadow mode for 30 days, provided the period captures payroll dates, weekends, international traffic, and other relevant variations.
Measure precision, which is the proportion of flagged events that are relevant; recall, which is the proportion of relevant events detected; false-positive rate; decision latency; and alert volume per investigator. For payment controls, also record how many legitimate transfers would have been delayed. A model that identifies more suspicious events but doubles manual reviews may merely shift costs into the operations team.
To be fair, data preparation often consumes more effort than model selection. Duplicate customer identifiers, inconsistent country codes, and timestamps recorded in local time can undermine otherwise capable analytics.
Planning Integration and Governance
Implementation usually works better as a series of operational phases: data mapping, controlled ingestion, shadow scoring, investigator review, limited automation, and broader production use. A cross-functional team generally includes fraud operations, data engineering, security, compliance, model validation, legal, and telecommunications specialists.
During data mapping, teams can define a canonical event schema (a standardized structure shared by connected systems) in JSON or Apache Avro. Transaction IDs, hashed account identifiers, device IDs, calling numbers, network indicators, and model scores should use documented field definitions. Data-quality checks should flag null identifiers, delayed events, and duplicate records before information reaches the scoring layer.
When incorporating connectivity inputs, buyers should document whether each signal supports real-time interdiction, post-event investigation, or both. Voice-routing metadata might enrich an investigation, for example, while a recent SIM-related indicator could contribute to a payment step-up decision, such as requiring additional authentication. The distinction affects latency requirements and integration cost.
Governance should follow a documented model inventory, validation schedule, and override process. The NIST Artificial Intelligence Risk Management Framework 1.0 provides a voluntary framework for governing, mapping, measuring, and managing AI risks. The Wolfsberg Group’s Principles for Using Artificial Intelligence and Machine Learning in Financial Crime Compliance address accountable and proportionate use of these technologies in financial-crime controls. Human reviewers should receive reason codes such as new device plus recent payee plus anomalous mobile event, rather than an unexplained risk score.
Choosing Outcomes Worth Measuring
Fraud teams should look beyond losses prevented. Operational measures include alerts per 1,000 transactions, median investigation time, cases closed without escalation, customer challenges triggered, and the percentage of decisions receiving usable reason codes.
The December 2025 investor.citizensbank.com survey found that 45% of surveyed middle-market companies were already using AI to combat fraud. Adoption alone says little about control quality. Buyers should compare model performance across customer segments, transaction types, countries, and channels to identify uneven results.
A practical dashboard can combine PostgreSQL case data with model telemetry in Snowflake or another analytical warehouse. Fraud operations might review daily alert quality, while model-validation teams examine drift, the deterioration or change in model behavior as data patterns evolve, monthly. If false positives rise after a routing, product, or customer-policy change, teams need enough event history to isolate the cause.
Buyer Takeaways
Communications signals should be assigned a defined purpose before integration. If a mobile attribute cannot influence scoring, authentication, or investigation, collecting it may add privacy and storage obligations without improving a decision.
Shadow mode also deserves protection from scope creep. The evaluation should use predefined fraud typologies, thresholds, and success criteria rather than changing the test whenever a model misses an edge case.
Finally, buyers should test investigator usability. A graph database such as Neo4j may expose relationships among accounts, devices, phone numbers, and beneficiaries, but analysts still need concise timelines and reason codes. A complex graph that takes ten minutes to interpret will struggle in a high-volume queue.
Frequently Asked Questions
How long does a financial fraud detection rollout take?
The schedule depends on data readiness and the number of channels involved. Buyers can begin with a 30-day shadow test after data mapping, then move into limited automation once model validation, privacy review, and investigator procedures are complete. Real-time SIP, mobile, and payment-event integration generally requires more engineering than nightly CSV scoring.
What should banks ask fraud detection vendors?
Prioritize solutions offering precision and recall by fraud type, 95th-percentile (or p95) scoring latency, reason-code coverage, model-drift controls, and integration with REST, Kafka, ISO 20022, and existing case-management systems. Buyers should also request a replay test using their own labeled events rather than relying only on vendor benchmark data.
Can a mid-market financial team use this approach?
Yes, if the scope stays narrow. A smaller team might begin with one problem, such as account takeover, one payment channel, and a managed scoring API feeding an existing case queue. As a practical starting scope rather than an industry benchmark, five to ten well-governed signals are often more manageable than every available communications and transaction field.
⬇️