A Deepfake Fraud Defense Analyst Fixes the Callback, Not the Detector

The Voice Said It Was the CFO, and the Wire Was Already Drafted

It is 4:40 on a Friday. A payments clerk takes a call from a number that matches the CFO's mobile, hears the CFO's voice, is told a supplier switch has to clear before close, and drafts the wire. Nothing in the procedure was skipped. The procedure said confirm verbally with an authorized approver, and the clerk did. That is the fact worth sitting with: no control failed, because no control existed for a voice that lies.

The person you need reads that incident as a procedure defect rather than a technology gap. Roughly half of businesses report having encountered deepfake fraud produced with AI ², and in one security firm's own job posting test about twelve percent of applicants used fake identities ³. That is a volume problem, and volume problems get solved by rules, not by heroics.

The first trait to screen for is the reflex to ask what the control was actually verifying. Describe the Friday call to a candidate and listen. A performed answer goes straight to detection: liveness checks, audio artifacts, a vendor. A real one asks who is authorized to change payment instructions, whether the callback number came from the request or from the vendor record, and how many people had to agree before funds moved. The good answer treats the voice as unverifiable input and redesigns around that.

The second trait is comfort in the attacker's toolchain without theatrics about it. Ask how much audio a usable clone takes today, and what changes when the call is a live video rather than a recording. You want someone who has actually made a passable clone of their own voice in a controlled test, can say what it cost them in time, and is calm about it. Someone who cannot describe the production side will keep telling you a detector's confidence number means something it does not.

The third is investigative discipline that survives contact with generated evidence. Every artifact can be fabricated, so the case has to be built from things that are expensive to fake at scale: account histories, device and network records, the timing of the request against a real business calendar, whether the supplier ever asked for this before. Building a modern fraud team here means bridging classic fraud investigation with a working understanding of how AI-generated attacks are produced ¹. The tell is whether a candidate reaches for corroboration under their own control, or for a score returned by a model they cannot inspect.

The fourth trait is knowing where their remit ends. Model supply chain and inference security belong to an MLSecOps engineer. This analyst owns the human-facing attack surface: calls, video, documents, and applicants.

Which Backgrounds Produce a Deepfake Fraud Defense Analyst Who Can Work the Case?

The strongest feeder is ordinary fraud investigation done at volume: card and ACH fraud analysts, AML investigators, bank BSA teams, insurance SIU. They already reason about corroboration, already write case files that survive a subpoena, and already know that a confident story is the most common feature of a fraudulent one. The synthetic media part is weeks of study. The investigative temperament is years.

KYC and onboarding review is the second pool, and it converts fastest of all, because those teams were the first to see identity documents that were rendered rather than photographed. The judgment involved is close enough that the same screening logic applies as for a judgment-based KYC and claims investigator.

The unexpected backgrounds are usually the better ones. Broadcast and post-production people know how synthetic audio and video are assembled because assembling them is their craft, and they hear a splice or a formant artifact before any tool flags it. Contact center quality reviewers have spent years listening to calls for whether the caller was who they claimed. Call center fraud desks at telcos have handled SIM swap and social engineering since long before any of this was generated. Voice actors and dialect coaches show up occasionally and are startlingly good at the audio half.

One background reads well and often disappoints: pure media forensics research. The skill is real, but a researcher who wants to publish detection accuracy will produce quarterly confidence scores while the wire still leaves on Friday. Hire that person as a consultant or as the second hire. Hire the case worker first.

What none of them arrive with is your money movement map, and that is the part you have to teach deliberately. Who can change vendor bank details, what a payment approval path actually looks like when someone is in a hurry, which exceptions are routinely granted. Budget two weeks of shadowing accounts payable and treasury before the analyst writes a single new rule, because a control written without that knowledge gets bypassed within a month and nobody tells you.

Ask How This Analyst Learned to Distrust a Confident Answer

Ask how they got good, and listen for practice rather than certification. This field has more course completion badges than working practitioners, and the badges say almost nothing. The answers worth hearing are specific: a case they got wrong, what the fabricated evidence looked like, and the check they now run every time because of it.

Good answers share a shape. Someone built a clone of a colleague's voice with consent, called their own service desk, and documented exactly which question the agent failed to ask. Someone else ran a suspected synthetic image through three detectors, got three different verdicts, and stopped treating any single confidence number as evidence. A third keeps a running file of attempted attacks against their employer, with what the attacker knew and where they learned it, which is the closest thing this discipline has to a lab notebook.

On their own use of AI, the useful question is what they ask a model to do and what they refuse to delegate. Strong candidates use assistants for the volume work: summarizing months of call logs, drafting the first version of a procedure, translating a vendor's technical claim into plain language, generating adversarial test cases against a verification script. They do not delegate the finding. Ask what happened the last time an assistant gave them a confident wrong answer during a case, and whether they caught it before or after it reached a report.

There is a specific tell here worth waiting for. The candidate who is good at this will, unprompted, distinguish between a claim they verified against a source outside the conversation and a claim the model asserted. That habit is the whole job, transposed. An analyst who cannot separate those two in their own work will not separate them in a case file either.

One warning about interview format, and it is sharper in this field than in most. Everything above can be performed. Deepfake vocabulary is cheap, and a live interview is now itself an attack surface: fraud teams are building synthetic-applicant defense stacks precisely because remote interviews are being run by people who are not the applicant ³. So do not decide from a conversation. Give the candidate a real anonymized incident and ninety minutes, and read what they produce. The difference between people who can do this work and people who can discuss it appears in the first page.

Recruit From Fraud Operations, Not From Detection Vendors

Look inside first. Your existing fraud, AML or onboarding review team almost certainly contains someone who has already flagged a synthetic document or an implausible verification call and got no traction. That person knows your controls, your exceptions and your approvers, and being asked to own this is usually the best offer they have had. Ask your fraud operations lead who keeps raising this; the name comes back immediately.

Outside, go where cases are discussed rather than where products are marketed. ACFE chapters and their regional conferences are real and long-standing, and full of investigators. The regional ACAMS chapters serve the AML and KYC side. Local fraud and payments meetups convened by banks and processors are where practitioners actually trade attack patterns. Adjacent titles worth approaching directly: fraud investigator, AML analyst, KYC review specialist, trust and safety investigator, telco fraud analyst, SIU investigator.

What kills the offer is almost never money. It is discovering that the role is advisory. Ask an experienced fraud investigator about their last job and you will hear about a memo that changed nothing, a control recommendation that finance overruled because it added a day to close, or a detection tool bought over their objection. Describe this role as monitoring and reporting and the strong candidates will decline politely.

Three things close the hire. Name the authority explicitly: this person can require a callback on an independently sourced number, can halt a payment pending verification, and can pause a candidate in the pipeline. Name who they escalate to when treasury pushes back, and make that person senior. And commit to funding the boring half, which is procedure rewriting and staff training, because that is what actually reduces losses. If you also expect them to certify controls for regulators or auditors, say so up front, since that is closer to the work of an AI system auditor and it changes both the seniority and the paperwork.

What Does a Deepfake Fraud Defense Analyst Cost, and Should They Sit On Site?

No wage series covers this title, and no salary survey found for this piece prices it, so this paragraph stays qualitative on purpose. Any single dollar figure published for the role right now is a guess wearing a benchmark's clothes. Price it internally instead: start from your senior fraud investigator band, then adjust upward for the two things that actually differ, which are the authority to halt a payment and the technical depth to evaluate a detection vendor's claims honestly.

Two forces push the number in opposite directions, which is why quoted ranges vary so wildly. Candidates who can also build tooling and run detection infrastructure get priced against security engineering, and you will compete there. Candidates who are investigators first sit in the fraud band, which in most organizations is materially lower. Decide which one you are hiring before you write the offer, because hiring the investigator and then expecting the engineer is the most common way this role fails in year one.

One caution about titles. The market is new enough that scope is inconsistent between companies, and a candidate's current title tells you little. Ask what they were allowed to stop. That answer, and not the title, tells you which band applies.

On location, most of the work is remote-native. Case files, call recordings, procedure drafting and vendor evaluation all travel fine. Two parts resist it. Live incident response during an active attempt works better when the analyst can stand next to treasury, and the first ninety days benefit enormously from being physically near accounts payable, because the exceptions people actually grant are learned in hallway conversations rather than from a policy document. A sensible pattern is on site for onboarding and for a standing monthly day, remote otherwise.

On-premise constraints bite where the evidence is regulated: recorded calls, identity documents, biometric templates. Several jurisdictions impose specific duties on biometric and identity data, and those rules differ by jurisdiction and are still moving through 2026. Two practical consequences. Anything the analyst uploads to an external detection service is a data transfer, so decide the policy before the first incident, not during one. And if you plan to generate synthetic samples for training or testing, the handling rules are the same ones that govern a synthetic data quality specialist. Read the primary regulation for your jurisdiction and check with counsel rather than reasoning from a summary.

See a sample report >