QWE AI Academy Here’s an unpopular opinion: most AI KYC tutorials are selling you the wrong problem. They pitch faster onboarding and cost savings – as if speed were the bottleneck. It isn’t. The bottleneck in 2026 is that the same generative AI powering your KYC vendor’s liveness check is also powering the fraud attempts hitting it. And the attackers are winning the pricing war.
If you’re evaluating AI KYC (Know Your Customer) tooling for a fintech, crypto exchange, or any regulated onboarding flow, you need to read the vendor pitches with different eyes. This guide walks through what actually matters, using numbers from the last twelve months of incident data – not the recycled benefits list every other article gives you.
The problem: document + selfie is quietly broken
For a decade, the standard KYC stack looked like this: user uploads an ID document, snaps a selfie, the system compares faces, checks a watchlist, done. AI made each step faster. It also made each step attackable.
The numbers are ugly. According to iProov’s data, injection attacks jumped 783% in 2024, and Jumio reported an 88% year-over-year rise in 2025. Sumsub’s telemetry tells the same story from a different angle: deepfakes reached 7% of all fraud attempts in 2024, climbed to 11% by early 2026, and “complex multi-step” attacks jumped from 10% of identity fraud in 2024 to 28% in 2025.
The pricing gap is what should scare you. Based on 2024-2025 pricing data, a synthetic identity sells for up to $15, and deepfake image services run $10 to $50. AI-generated faces good enough to pass KYC cost under $20 and about 30 minutes to create. On the defense side, a competent enterprise KYC stack easily runs into six figures annually. Attackers pay pennies to fire hundreds of attempts. You pay a lot to catch one.
Why the standard AI KYC pitch falls short
Read three vendor landing pages back-to-back and you’ll see the same four benefits: speed, accuracy, cost savings, global compliance coverage. It’s true. It’s also not what determines whether you get compromised.
The gap most tutorials skip: injection attacks. This isn’t a fancier deepfake shown to a webcam – it’s the attacker bypassing the camera entirely and feeding a synthetic video stream directly into the verification pipeline. Most 2020-era liveness detection can’t see it, because the fake never touched a lens. The new European technical specification CEN/TS 18099 was created specifically to define Injection-Attack Detection (IAD) requirements, complementing ISO 30107-3. If your vendor can’t tell you which of those two standards their liveness engine tests against, that’s your answer.
And the human fallback? Also broken. In a 2025 iProov study, only 0.1% of participants correctly identified every real and fake image or video shown. “We escalate suspicious cases to human review” used to be a comforting line. Now it’s a compliance risk.
A better way to evaluate AI KYC vendors
Stop asking “how fast is your onboarding.” Start asking three specific questions.
- What layers of liveness do you run? Passive alone is a red flag in 2026. You want passive + active + injection-attack detection tested against CEN/TS 18099 or ISO/IEC 30107-3.
- What’s your capture-path integrity? Can the vendor prove the video reaching their servers came from a real camera on a real device, not a virtual camera pipe? Ask for the technical mechanism, not marketing language.
- What’s your regulator-facing audit trail? If FATF or your national supervisor asks for evidence a specific onboarding wasn’t a deepfake, what do you hand them? An answer that starts with “we log the decision” is not enough.
That last one matters more than most buyers realize. FATF’s December 2025 Horizon Scan explicitly identifies deepfakes as capable of bypassing AML controls, CDD systems, and digital ID verification, and signals that supervisors will scrutinise deepfake controls as part of standard AML reviews. Under the EU AI Act (in force since August 2024), remote biometric identification systems are classified under Annex III as high-risk – meaning documentation obligations, not just detection ones.
Pro tip: Ask any vendor to demo a live face-swap attempt against their system while you watch. Not a slide about their detection rate – an actual attempted bypass. If they refuse or reschedule, that’s your evaluation.
A real example: the Arup case, read differently
The Hong Kong Arup incident (early 2024) gets cited in every fraud article, usually to make a point about deepfakes being scary. But the actionable lesson is buried. A finance worker moved US$25.6 million across 15 wire transfers off a single video call where multiple deepfaked “colleagues” pressured him into approving payments.
The KYC-relevant detail: the video call bypassed identity verification not by defeating it, but by making it feel unnecessary. The victim already “knew” the people on the call. This is the pattern to watch for in your own stack – the fraud that succeeds isn’t the one that beats your best control, it’s the one that routes around it. When you deploy AI KYC, map every decision point where a human trusts a face without invoking verification. Those are your real gaps.
Where AI KYC does earn its keep
None of this means the AI-heavy tools are a bad buy. Used honestly, they solve real problems. KYC automation combines AI, ML, OCR, and biometric verification for automated onboarding, real-time transaction monitoring, and Enhanced Due Diligence (EDD) on high-risk clients. The wins are concentrated in three places:
- Document parsing at volume. OCR on a passport is a solved problem. Let the machine do it.
- Perpetual KYC. Continuous risk re-scoring based on transaction patterns catches things a point-in-time onboarding check never will.
- Watchlist and PEP screening. Fuzzy matching across sanctions lists, in dozens of languages, is where ML outperforms rules engines by a margin no manual process can close.
Notice what’s not on that list: the selfie step. That’s the layer where you should be paranoid, layered, and willing to add friction.
What to do this week
Pull your current KYC provider’s most recent technical whitepaper. Search it for two terms: “injection attack” and “CEN/TS 18099” (or the older “ISO 30107-3”). If neither appears, email your account manager today and ask which standards their liveness engine has been tested against, when, and by whom. That single email will tell you more about your fraud exposure than any vendor comparison table you’ll read this quarter.
In Poland, by the way, this same process is often labeled PSK – Poznaj Swojego Klienta, per the karierawfinansach glossary – the regulations and the deepfake risk are identical regardless of what you call the process.
FAQ
Is AI KYC actually cheaper than manual review?
At volume, yes. For edge cases, no – and the math flips fast once you factor in deepfake defense costs, which often eat the savings from automating the easy 95%.
Do I need a dedicated deepfake detection vendor, or is my KYC provider enough?
Depends on your risk profile and transaction volume. A regulated crypto exchange onboarding thousands of users a month should assume its primary KYC vendor’s liveness alone won’t hold – the sector has been a documented top target for deepfake fraud, and the attack volume numbers bear that out (DeepStrike tracked a 704% rise in deepfake-capable attacks in 2023 alone). Turns out the gap between “our liveness passed certification last year” and “our liveness stops current injection attacks” is bigger than most buyers expect. A low-risk B2B SaaS doing occasional identity checks probably doesn’t need a second vendor yet. Re-evaluate at least twice a year – this changes fast.
What’s the single biggest mistake buyers make?
Treating the pilot as the evaluation. Vendors calibrate for pilots – their system will look its best under controlled conditions with known test inputs. What you actually want is a red-team exercise: give someone a $50 budget and a Telegram channel, tell them to try to break the onboarding flow, and see what they find. One afternoon of adversarial testing will surface things six months of normal usage won’t. The misconception here is that “passing the pilot” and “passing a real-world attack” are the same test. They’re not even close. If your vendor won’t allow red-team clauses in the contract, that refusal is itself data worth acting on.