Skip to main content
BSS/OSS Academy
💰
Section 6.10

Fraud Management

Last updated 12 Oct 2026

01

At a glance

Fraud Management detects, scores, decides on and investigates suspicious activity across customers, accounts, devices, SIMs, network traffic and payments. Revenue Assurance looks for our own failures after the fact. Fraud management looks for someone else’s abuse, and must often decide within minutes whether to let activity continue, because the loss grows while it runs.

02

The picture

Fraud management simulation

Fraud asks whether someone is deliberately abusing the service, account, network or payment system. Revenue assurance asks a different question — whether our own systems lost legitimate revenue.

An attacker moves the victim’s number to a SIM they control, then uses it to pass one-time-code checks elsewhere.

10 stagesFictional customers · illustrative figures
  1. Detect

  2. Decide

  3. Respond

Narration

Twenty minutes after a SIM replacement, the same account sees a password reset, a login from a new device in an unfamiliar city and a high-value handset order on account credit. Each event alone is ordinary; together they are the classic takeover sequence. Press Raise the alert to follow it through detection to a decision — yours.

Alert not yet raised.

Signals vs baseline

Signals appear when the detection engine evaluates the alert.

Investigate the alert

The alert pauses at Decision management. You take the decision.

Try it: pick a scenario, follow the alert from activity to scoring, investigate the evidence, decide allow, challenge, restrict or block, and see the consequence
03

How it works

A different question from Revenue Assurance

Revenue Assurance asks
“Did our systems or processes lose legitimate revenue?” The cause is internal: a lost record, a missed event, a wrong price.
Fraud Management asks
“Is someone deliberately or suspiciously abusing the service, account, network or payment system?” The cause is external intent, sometimes with inside help.

The two share data and are often sold together as “RAFM”, but they work differently. RA reconciles and recovers. Fraud management decides, often in seconds, because much of a fraud loss cannot be recovered afterwards. Its measure of success is fraud loss prevented at an acceptable cost in false positives and customer friction.

From activity to decision

1
Activity
Network / BSS / channels

Calls, sessions, SMS, SIM changes, logins, orders, top-ups and payments.

2
Ingestion
Fraud platform

Usage from mediation and the OCS, NRTRDE roaming summaries, CRM and order events, payment and SIM events. Real-time feeds for high-velocity fraud, batch for pattern analysis.

3
Enrichment
Fraud platform / reference data

Account age, credit class, products held, device history, home location, known-bad number ranges and IMEIs.

4
Detection
Rules + ML

Rules (thresholds, velocity, blocklists) catch known patterns quickly and explainably. Machine-learning models find behaviour unusual against the customer’s own baseline and peer groups. Most operators run both.

5
Scoring
Fraud platform

Signals are weighted into a score per account, SIM or transaction. An uncalibrated score that everyone ignores is worse than none.

6
Decision
Decision rules

Policy maps score and context to an action. High-confidence, high-value patterns can be automated; the rest go to a person.

Four actions, applied elsewhere

ActionWhat it doesApplied in
AllowLet the activity continue, and monitorFraud platform
ChallengeStep-up verification or call-backCRM / channels
RestrictBar international or premium destinations, cap spendOCS / network policy
BlockSuspend the SIM, reject the order or paymentOCS / CRM / payment gateway

The fraud platform decides; it rarely enforces. Every action needs an integration into a system that does.

From alert to learning

  1. Case management: the alert becomes a case with evidence, owner and status.
  2. Investigation: an analyst confirms or rejects fraud from the account, device, location, usage, transactions and recent changes.
  3. Recovery: withhold payments to partners where the agreement allows, dispute with carriers, reverse charges, refer to law enforcement, reinstate genuine customers.
  4. Improvement: confirmed and rejected cases feed back into rules and models. Without this loop, false positives stay high and new patterns go unseen.

The tensions every control sits between

TensionOne wayThe other way
False positives vs lossesLower thresholds catch more fraud, soonerEvery false alert costs analyst time; every wrongly blocked customer costs revenue and trust
Friction vs protectionStep-up checks stop takeover and payment fraudThe same checks slow every genuine customer
Real-time vs batchIRSF and SIM-box losses grow by the minute; only real-time feeds (OCS, signalling, NRTRDE) limit themReal-time is costlier and sees less context; batch finds patterns across weeks
Automation vs investigationAutomated blocks act in secondsAn automated block on the wrong customer is an outage the operator caused

Standards and data

There is no TM Forum Open API for fraud management. Fraud platforms consume usage (TMF635), customer and account (TMF629, TMF666), product inventory (TMF637) and payment (TMF676) data alongside network feeds. Operators share intelligence (high-risk number ranges, new patterns, stolen-device lists) through the GSMA Fraud and Security Group and the Communications Fraud Control Association (CFCA).

04

Who owns what

EntitySystem of RecordSystem of EngagementSystem of ReferenceNotes
Fraud alert, risk score and caseFraud management platformFraud analyst case workbenchRevenue Assurance, Finance—
Detection rules, models and decision policyFraud management platform——The policy (who may block what, automatically) must be agreed with the business.
Barring or spend restrictionOCS / network policyFraud platform (requests it)CRMThe fraud platform decides; the network enforces.
Usage and roaming recordsMediation / OCS—Fraud management platformTMF635; NRTRDE for roaming.
Customer, account and identity dataCRM / account management—Fraud management platformTMF629, TMF666. Used for enrichment.
05

In practice

Legacy reality
Where a fraud platform is fed only in batch from mediation, detection runs hours behind activity. That suits pattern analysis but not IRSF or SIM-box fraud, where the loss accumulates overnight and the carrier invoice arrives weeks later.

RA and fraud are often sold together as one “RAFM” suite. The shared data helps. The risk is running fraud at RA’s rhythm: periodic reconciliation where minutes matter.

Automated actions depend on integrations that are easy to leave out of scope: a real-time path to the OCS or network policy to bar destinations, and to CRM to flag accounts. Without them, a high score becomes an email to an analyst. The other gap is the feedback loop. Cases closed without a confirmed or rejected outcome teach the models nothing.

06

The trade-off

What it solves

A single place to detect, decide on and investigate abuse across network, account and payment data, with actions that stop loss while the activity is running.

What it does not solve

  • Our own capture and billing failures, which are Revenue Assurance’s job.
  • Losses already committed to the carrier chain before detection.
  • Weak identity checks in shops and contact centres, which only process change can fix.

What it adds

  • Real-time feeds from the OCS, signalling and payments, with their cost.
  • A decision policy the business must own, including who may block a customer automatically.
  • False positives that become customer outages and complaints.
  • Case outcomes that must be recorded consistently to train rules and models.

Architecture decision

An enterprise account scores high for IRSF at 02:00: international calls jumped from about 10 a day to hundreds in an hour. What should happen automatically?

  • Block the account

    Choose when the destinations are on a high-risk range list and the pattern matches confirmed IRSF, so the loss per hour outweighs a false block.

  • Restrict international calls only

    Choose when the signal is strong but the account carries legitimate domestic traffic you cannot cut overnight.

  • Challenge and monitor

    Choose when the destinations are unusual but not high-cost, and a named contact can confirm quickly.

Breaks if the policy was never agreed with the business. The analyst on nights has no authority to act, and the carrier invoice for the weekend arrives weeks later.

07

Takeaways

Key Takeaways

  • RA asks whether our systems lost revenue; fraud management asks whether someone is abusing us
  • The pipeline runs activity, ingestion, enrichment, detection, scoring and decision, then case, investigation, recovery and improvement
  • Rules catch known patterns explainably; ML finds unusual behaviour; most operators run both
  • The fraud platform decides, but the OCS, network, CRM and payment gateway enforce
  • Every control trades fraud loss against false positives and customer friction

Check yourself

  • Why is a batch feed from mediation not enough for IRSF detection?Show answer

    IRSF losses grow with every minute the calls run, and the termination charges are owed regardless. Detection hours later means the loss is already committed. Real-time feeds from the OCS or signalling are needed to restrict while it is happening.

  • The fraud team’s false-positive rate has not improved in a year despite new rules. What is the likely gap?Show answer

    The feedback loop. If cases are closed without a confirmed or rejected outcome, rules and models never learn which alerts were genuine.

  • A high-risk score triggers “restrict international calls”. Which system actually applies it?Show answer

    The OCS or network policy. The fraud platform decides and requests the action; it needs an integration to the enforcing system, or the decision is only a recommendation.