KYC / AML Automation: What Production Deployments Show

Yes — at volumes and pass rates that survive regulators. North America's largest digital identity provider processes over 10 million ID cards a year through automated extraction at under 15 seconds per document, replacing a ten-contractor manual review operation. Banorte cut document validation from 20 minutes to eight and — the number that matters most — raised its automatic approval rate from 70% to 91% over two years, returning 2,000 hours a month to the business while adapting to market changes in under 24 hours. FinTrU runs a 99% first-time pass rate in compliance review on 96% document-classification accuracy.

The build side documents the same pattern with its economics exposed: Sun Finance, processing 80,000 monthly microloan applications with roughly one in ten fraudulent, lifted extraction accuracy from a 79.73% baseline to 90.8% while cutting per-document cost 91% — projecting manual review down from 60% of documents to 30%. And Grab's custom vision model closed the gap general models couldn't: Thai document accuracy up 70 percentage points from baseline, Vietnamese up 40.

Every deployment keeps the same shape: the machine reads, validates and drafts; the analyst decides, on an audit trail. In this category that's not caution — it's the product's compliance posture.

The frontier, in three distinct ways this record documents better than any vendor deck. First, the language gap: general and open-source vision models were trained on Latin scripts, and it shows — Grab's fine-tuning attempts still struggled with Thai and Vietnamese documents and dense unstructured layouts until they trained a custom vision model with language-specific stages. If your customers' IDs aren't in the training data's languages, benchmark on your regions before believing any accuracy claim.

Second, the safety-alignment collision — this record's most unusual failure. Sun Finance's replacement attempt using a frontier model alone reached only 61.8% accuracy, with ID-number extraction collapsing to 43% — because the model's built-in privacy safeguards refused to extract PII from identity documents. The task was legitimate; the model declined it anyway. The production system that works routes around this with purpose-built pipelines — a lesson in why "just call the API" fails in exactly the domains where the API is most cautious.

Third, the analyst-replacement mirage: Adyen's DABStep benchmark — over 450 real-world multi-step analysis tasks — has state-of-the-art reasoning agents scoring around 16%. The drafted-narrative-plus-human-decision pattern isn't regulatory theatre; on current evidence, it's where the capability actually ends.

Buy the regulated core — screening, IDV and compliance workflow platforms carry the sanctions-list plumbing, the audit posture and the calibration that regulators have already seen elsewhere, and the record's bank-grade outcomes (Banorte's 91% auto-approval, FinTrU's 99% pass rate) ride on that class of product. The evaluation criterion this record adds to the standard list: demand accuracy evidence on your document geography specifically, because the language gap above is where bought and built systems alike quietly underperform.

The build lane is real but narrow, and it's defined by exactly two conditions the record's builders share: document volume in languages the vendors serve poorly (Grab's Southeast Asian custom vision model, Sun Finance's multilingual microloan pipeline), and the engineering bench to own a regulated ML system's lifecycle. Both builds published their economics — 91% per-document cost reduction at Sun Finance is the kind of number that justifies the lane — and both kept the analyst checkpoint intact.

What nobody in this record does is hand the decision to an agent, and the 16%-on-real-analyst-work benchmark is the reason stated with unusual honesty by a team that would profit from the opposite finding. Automate the reading, the screening and the narrative; keep the decision human, evidenced and auditable — that's not the compromise position here, it's the entire documented state of the art.