A Step-by-Step Playbook for Investigating Suspected AI-Generated Fraud in Onboarding

Generative AI didn’t invent identity fraud. It industrialized it. A face-swapped selfie that once required a compositing artist and a few hours in After Effects can now be produced by an app in under a minute, for less than the cost of a coffee. A forged utility bill that used to have telltale font substitutions can now be generated with correct kerning, correct paper texture, and a plausible address on the first try. The economics changed, and the volume followed.

For a KYC analyst, this doesn’t mean throwing out the fundamentals — it means applying them faster and with a different set of red flags in mind. This playbook is the method I use to work a suspected AI-fraud case from first signal to filed decision.

The three vectors worth knowing cold

Before the steps, it’s worth being precise about what “AI fraud” actually covers in an onboarding context, because the investigative approach differs for each:

• Synthetic identity documents. IDs, utility bills, and bank statements generated or edited with GAN- or diffusion-based tools, sometimes stitched onto a real template obtained from a data breach.
• Biometric spoofing. Deepfake video or photo injected into a liveness check — either a pre-recorded synthetic face presented to the camera, or a real-time face swap run through a virtual camera driver.
• Synthetic identity fraud proper. A blended identity — a real Social Security or national ID number paired with a fabricated name, date of birth, and AI-generated supporting documents — built to season an account over months before it’s used for a bust-out.

Each of these leaves a different evidence trail, which is why step one is never “run everything through the same check.”

The playbook

1. Recognize the signal. Most AI-fraud cases don’t announce themselves — they show up as a cluster of small anomalies that individually wouldn’t trigger review. Watch for: a liveness confidence score that passed but sits near the threshold; a document OCR match that’s perfect too perfect, with zero of the natural extraction noise you’d expect from a photographed physical document; an applicant who completes a multi-minute form in under twenty seconds; or a selfie whose lighting direction doesn’t match the ambient light implied by the ID photo taken seconds earlier.

2. Freeze and preserve before you triage. The moment a case is flagged, stop any automated pipeline step that would re-encode, re-compress, or overwrite the original files. Recompression destroys the compression-artifact and metadata evidence you need in step three. Hash the original files (SHA-256 is sufficient) and log the hash before anyone — human or model — touches them again. This is the single most-skipped step, and it’s the one that determines whether your eventual SAR narrative is defensible or just an opinion.

3. Run document forensics triage. With the originals preserved, look for the specific fingerprints generative tools tend to leave behind: inconsistent font kerning or letter spacing under magnification, especially around variable fields like names and dates; shadow and lighting mismatches between a document photo and its supposed physical environment; compression “ghosting” around edited regions when the file is examined at high zoom; and metadata that’s either suspiciously empty (stripped) or internally inconsistent (an EXIF timestamp that predates the account’s creation, or a device model that couldn’t have produced that image resolution). Reverse image search the document background and template — synthetic documents are frequently built from a small number of reused base templates, so the same watermark or table layout showing up across “different” applicants is a strong signal.

4. Cross-check the biometric layer independently of the document layer. Don’t let a passed liveness check close the loop on its own. Look for the absence of natural micro-motion — no visible pulse in the skin, no blink irregularity, unnaturally stable framing for a handheld phone capture. Passive liveness models score texture and depth cues; a well-executed face swap can beat one signal while failing another, so if your stack only checks one, treat a borderline pass as inconclusive rather than clean. Run the selfie through reverse image and face-similarity search against your own prior applicant pool — face reuse across “different” identities is one of the highest-precision indicators available to an analyst, because legitimate applicants essentially never share a face.

5. Correlate the network and velocity signals. Individual identity fraud usually leaves a lonely trail; synthetic-identity rings and AI-fraud farms leave a crowded one. Pull device fingerprint, IP/ASN, and referral-path data across the flagged case and its near neighbors in time. Look for shared devices or emulators across “unrelated” applicants, IP clustering on residential proxies or VPN exit nodes inconsistent with the claimed address, and submission cadences too regular to be human — form fills spaced at near-identical intervals are a signature of scripted or semi-automated attack tooling.

6. Make the escalation decision against a written threshold, not a feeling. Have a pre-agreed decision tree so the call doesn’t depend on who’s on shift: outright reject only when document forensics and biometric signals both fail independently; route to Enhanced Due Diligence when one channel is ambiguous and the other is clean; file a SAR with the NCA under POCA 2002 when the fact pattern meets the suspicious-activity threshold, regardless of whether the account was ultimately opened. The US equivalent regulator, FinCEN, issued a November 2024 alert on this exact typology (FIN-2024-Alert004) that’s worth reading regardless of which jurisdiction you sit in — it cites a case earlier that year where fraudsters used deepfake video to impersonate a multinational company’s CFO on a video call convincingly enough to get a finance employee to wire $25 million. No document was forged and no liveness check was fooled in the traditional sense; the entire attack happened after onboarding, through a real-time video call. The lesson most teams miss: your deepfake exposure doesn’t end at KYC. If your fraud program only tests generative-AI defenses at the onboarding gate, you’ve secured the door and left the window open.

7. Write the narrative like it will be read by someone with no context. A defensible SAR narrative states, in plain chronological order: what was observed, what tool or check produced the observation, and why that observation met the suspicious-activity standard — not your suspicion alone. “Applicant’s selfie video showed no observable blink over 8 seconds of passive liveness capture; reverse image search matched facial embedding to two prior applications under different names submitted within 48 hours” is defensible. “Seemed like a deepfake” is not.

8. Close the loop on the typology, not just the case. Every confirmed AI-fraud case is training data for your program, whether or not you touch a model directly. Log the specific indicator combination that caught it, update your red-flag registry, and flag the pattern to whoever owns rule tuning or model retraining. AI-generated fraud evolves fast precisely because the tooling is shared across attackers — a typology that worked against you this month will circulate, and the next hundred attempts will be slightly better at evading whatever caught the last one.

Why the fundamentals still do the heavy lifting

None of this replaces CDD, EDD, or a properly risk-rated customer profile — it sits on top of them. The core discipline of KYC has always been “does the evidence in front of me actually corroborate the identity being claimed, and can I explain why in writing.” Generative AI just raised the production quality of the evidence an attacker can fabricate. The analyst’s job hasn’t changed; the burden of proof for “corroborate” has just gotten heavier, and it’s why a preserved, hashed, well-documented case file matters more now than it did five years ago — not less.