The phone channel is where fraudsters go when digital channels get harder, because a contact centre agent can reset access, change contact details or move money on the strength of a voice and a few answers. Contact centre fraud prevention tools score the call before and during the conversation: network and carrier signals about the calling number, device and behaviour analytics, authentication that does not rely on what the caller knows, and guidance that keeps agents from being talked round. Voice cloning has changed the category, because a voice is no longer evidence of a person. This vendor-neutral buyer's guide is for fraud, contact centre and identity leaders choosing caller authentication and voice fraud tools for customer and member calls. Detecting synthetic media in general is covered in the Deepfake Detection Buyer's Guide, internal IT help desk resets in the Workforce and Help Desk Identity Verification Buyer's Guide, and AI voice agents in the Voice AI Agent Security and Fraud Guide.
Key takeaways
- Do not buy voice as an authenticator. NIST SP 800-63B-4 states that biometric comparison based on voice shall not be used, and treats no biometric as an authenticator by itself. Voice analytics can add risk signal; it should not open the account.
- Caller ID is a signal, not proof. STIR/SHAKEN attestation helps on US IP networks, but coverage gaps, non-IP legs and international traffic mean a verified number still does not prove who is speaking.
- Move authentication out of the call. The strongest pattern is a push or passkey challenge to the customer's registered app or device, started by the agent and answered outside the voice channel.
- Protect the high-risk actions, not just the greeting. Changes to phone number, email, address, payee or authenticator are what fraudsters call for; those need step-up and delays even after a caller passes.
- The agent is part of the product. Scripts, risk prompts and a no-override rule matter as much as the model score, because social engineering targets the person.
Define requirements
- Fraud types in scope: account takeover by phone, contact detail and authenticator changes, payment and payee changes, SIM swap preparation, first-party fraud, and reconnaissance calls that harvest data for later attacks. CISA's Scattered Spider advisory describes voice calls used to have staff reset passwords and MFA and the use of SIM swaps; see the Scattered Spider Threat Actor Profile.
- Synthetic voice: FinCEN's November 2024 alert (FIN-2024-Alert004) warns of deepfake media used to get around identity verification and authentication, and the FBI's December 2024 public service announcement describes AI audio used to impersonate people and reach bank accounts. Assume a caller can sound like your customer.
- Channels and routes: inbound IVR and agent calls, outbound callbacks, chat that escalates to voice, outsourced and offshore sites, and branch or store staff who take calls.
- Telephony estate: carriers, SIP trunks, contact centre platform, IVR, CRM and case tools the product must integrate with, and whether calls arrive over IP end to end.
- Customer reach: share of callers with your app installed, with a registered device or passkey, and those with neither, who will need a fallback.
- Regulatory constraints: recording and analytics consent, biometric data rules (voiceprints are biometric data in many jurisdictions), retention, and calling rules. The FCC has ruled that AI-generated voices are "artificial" under the US Telephone Consumer Protection Act, which matters if you deploy synthetic voices yourself.
- Success measures: fraud losses and attempts by phone, false positive rate on genuine callers, average handling time impact and containment in IVR.
Capability areas
| Capability | What it does | Strength | Limits to plan for |
|---|---|---|---|
| Network and carrier signals | Checks the calling number: STIR/SHAKEN attestation (A full, B partial, C gateway) where present, line type, recent porting or SIM change, carrier data, known fraud numbers | Scores the call before it is answered; no customer effort | Partial coverage; attestation shows the originating provider vouched for the number, not who is calling. See the STIR/SHAKEN Caller ID Authentication Guide |
| Call device and audio analytics | Analyses signalling and audio characteristics to spot spoofing, virtual phone systems, replay and synthetic speech | Catches spoofed and synthetic calls that pass number checks | Model drift as voice generation improves; needs evidence on current attacks |
| Out-of-band app or device authentication | Agent or IVR sends a push or passkey challenge to the registered app or device; result returns to the agent desktop | Strong; independent of what the caller knows or sounds like | Only for customers with the app or device; useless if the device itself was swapped |
| One-time passcodes by SMS or voice | Code sent to the number on file | Universal reach | NIST treats PSTN delivery as restricted and advises checking device swap, SIM change and number porting first. See the SIM Swap Fraud Guide |
| Voice biometrics | Compares the caller's voice to an enrolled voiceprint, actively or passively | Low friction | Not acceptable as an authenticator under NIST SP 800-63B-4; cloning risk; biometric data obligations. See the Voice biometrics and voiceprint authentication in the age of voice cloning |
| Behaviour and account risk | Links the call to recent digital activity: new device logins, password resets, contact changes, repeat calls across accounts | Spots campaigns and account takeover sequences | Needs data feeds from digital channels and the fraud platform |
| Agent guidance and workflow | Risk banner, scripted step-up, locked fields for high-risk changes, case creation, cooling-off periods | Makes the score actionable and consistent | Fails if agents can override without a second person |
Evaluation criteria
| Criterion | What to test | Evidence to request | Suggested weight |
|---|---|---|---|
| Detection on your traffic | Fraud caught on a historic labelled call set and live shadow traffic | Detection rate and false positive rate on your data, not a reference set | 25% |
| Synthetic and spoofed call resistance | Cloned voice, replayed audio, spoofed number, virtual phone system, call forwarding | Red team results; how often models are updated against new generators; any ISO/IEC 30107-3 presentation attack testing | 20% |
| Strong out-of-band authentication | App push or passkey challenge from the agent desktop with clear result codes | Live demonstration in your app and contact centre platform | 15% |
| High-risk action controls | Step-up and holds on contact detail, authenticator and payee changes | Configuration walkthrough; audit log samples | 15% |
| Integration and latency | Score available before answer; CRM, IVR, case management and fraud platform links | Measured latency; integration list for your estate | 10% |
| Agent experience | Clarity of risk prompts; effect on handling time | Pilot handling time and agent feedback | 5% |
| Data protection | Recording, voiceprint and metadata retention, location, model training use | Data processing agreement; data protection impact assessment support | 5% |
| Product security | Admin access, API keys, separation of customer data | Security attestation; admin role model and logs | 5% |
Weights are a starting point; set your own before demonstrations so scoring is not shaped by the best presentation.
Questions to ask vendors
- What exactly does your score use: network, carrier, audio, device, behaviour or account data? Which of these work for calls that arrive without STIR/SHAKEN attestation?
- Show detection of a cloned voice of a real staff volunteer, generated with current tools, played over a mobile call. What does the agent see?
- How often are synthetic speech models retrained, and how do you learn about new generation techniques?
- If you offer voice biometrics, how do you position it given NIST SP 800-63B-4 does not permit voice comparison for authentication? Can we use your product with voice matching switched off?
- How do you detect recent SIM swap, number porting or call forwarding on the customer's number before a passcode is sent?
- Which high-risk actions can your workflow lock until out-of-band authentication succeeds, and can an agent override that?
- How do you link a call to recent digital activity on the same account and to calls about other accounts from the same source?
- Where are recordings, voiceprints and call metadata stored, for how long, and are they used to train models for other customers?
- How do outsourced contact centre sites use the product, and do their agents get the same controls and logging?
Red flags
- Voice biometrics presented as sufficient to authenticate a caller on its own.
- Caller ID or attestation treated as proof of the caller's identity.
- Detection figures only from the vendor's reference data, with no offer to test on your calls.
- No answer on synthetic voice, or claims of complete detection.
- Agent can bypass step-up for contact detail or payee changes without a second approver or an alert.
- Knowledge-based questions (date of birth, address, last transaction) still the main fallback.
- Recordings or voiceprints kept indefinitely by default, or used to train shared models without a contractual opt-out.
Proof of concept
- Agree success criteria first: detection rate on fraud, false positive ceiling on genuine callers, maximum handling time increase, and latency before answer.
- Run the product in shadow mode on live traffic for four to six weeks, scoring calls without changing agent behaviour, and compare scores with confirmed fraud cases.
- Replay a labelled historic set of fraud and genuine calls, where your recording consent allows it.
- Run controlled attacks with consent: a cloned voice of a volunteer, spoofed numbers, a call after a simulated SIM swap, and a social engineering script asking the agent to skip step-up.
- Pilot out-of-band app or passkey authentication with one queue and measure completion, handling time and complaints.
- Switch high-risk actions to enforced step-up for the pilot queue, with a documented manual fallback for customers without the app that needs two people and a delay.
- Confirm events reach the fraud platform and SIEM, and that a high-risk change without successful authentication raises an alert.
- Complete a data protection impact assessment with the vendor's actual data flows before rollout.
Fraud and phone channel figures for the business case are in the Deepfake Fraud Statistics page; agent training material is in the Vishing and Smishing Guide.
How NHI Mgmt Group can help
We provide independent requirements, RFP and evaluation support for fraud and identity verification tooling. Browse vendors in our products directory or contact us.
Related NHI Mgmt Group resources: Voice AI Agent Security and Fraud Guide · Deepfake Detection Buyer's Guide · Voice biometrics and voiceprint authentication in the age of voice cloning · STIR/SHAKEN Caller ID Authentication Guide · SIM Swap Fraud Guide