# Mapping AI across the banking value chain
A customer deposits $3,000 into her checking account through a mobile app at 11:47pm on a Sunday. By the time that deposit clears, is reconciled against the bank's general ledger, and shows up in a regulatory report on Monday, it has passed through at least six systems where AI is quietly making decisions, and three more where a vendor's slide deck promised AI but a human, or a simple rule, is actually doing the work. This lesson traces that journey.
Follow the $3,000 from her phone to the bank's books. At each stop, ask: is AI genuinely operating here, or is this a marketing claim?
Before she can deposit anything, she opened the account. AI-powered document verification tools (from vendors like Onfido or Jumio) use computer vision to check whether her driver's license is real, matches her selfie, and hasn't been Photoshopped. This is mature, genuinely deployed technology across most digital-first banks and neobanks in the US and Europe.
Where it's real: liveness detection, document forgery checks, matching a face to an ID photo.
The moment she hits "submit," a fraud model scores the transaction in milliseconds. It weighs device fingerprint, geolocation, deposit velocity (has she deposited unusually often this week?), and behavioral biometrics (typing rhythm, swipe pressure). This is one of banking's oldest and most successful AI use cases, machine learning fraud models have been in production since the 2000s, long before "AI" became a boardroom buzzword.
Where it's real:
If she deposited a check via mobile capture, optical character recognition (OCR) and computer vision extract the amount, date, and signature, then flag mismatches between the handwritten and printed amount fields. This is a narrow, well-solved AI problem, mature since the mid-2010s.
Where it's real: narrow computer vision tasks with a single clear input and output.
She has a question about when the deposit clears and messages the bank's chatbot. Here the picture splits. Simple intent-routing chatbots (answering "what's my balance," "when does this clear") genuinely work and reduce call center volume. But vendors routinely oversell generative AI chatbots as capable of complex dispute resolution or nuanced advice. Independent testing, including examples from the Consumer Financial Protection Bureau, has flagged chatbots giving customers inaccurate or unhelpful answers on account-specific questions.
Where it's oversold: "conversational AI that resolves complex disputes end to end." Most still escalate to humans for anything non-routine.
Overnight, her deposit needs to be matched against the ledger entry, the payment network record (ACH, the US Automated Clearing House, or SEPA in Europe, the Single Euro Payments Area), and internal accounts. Machine learning models increasingly handle "fuzzy matching," reconciling records that don't line up perfectly due to timing differences or formatting mismatches. This is a genuine, growing application, though most large banks still run substantial rule-based logic alongside it.
Where it's real: pattern matching across large, messy datasets with tolerances for near-duplicates.
Her deposit is swept into an AML monitoring system required under the US Bank Secrecy Act and, in Europe, the EU's Anti-Money Laundering Directives. AI models flag unusual patterns for review. This is real AI, but it's also the site of one of banking's most persistent AI disappointments: false positive rates.
Estimate (widely cited across industry sources, exact figures vary by institution): traditional rule-based AML systems generate false positive rates as high as 90-95%, meaning only 5-10% of flagged transactions are genuinely suspicious. Newer ML-based systems claim to cut this substantially, but banks report mixed results in production, and regulators still require explainability that many "black box" models struggle to provide.
Where it's oversold: "AI that eliminates false positives." No vendor has credibly demonstrated this at scale; incremental reduction, not elimination, is the honest claim.
If she has a loan relationship, AI models score her ongoing credit risk and may trigger a marketing offer ("you may qualify for a higher limit"). Credit scoring augmented by ML is real and regulated (in the US, subject to the Equal Credit Opportunity Act and Regulation B's requirement for adverse action explanations). But "hyper-personalized AI banking" that anticipates every customer need is still mostly aspirational; most deployed systems are simpler propensity models, not the sophisticated personalization engines marketed in vendor demos.
Where it's oversold: fully autonomous, real-time personalization across all channels. Most banks run batch-scored models updated daily or weekly, not live adaptive AI.
Come month-end, her deposit is one of millions rolled into reports for regulators, in the US the Federal Reserve and the Office of the Comptroller of the Currency (OCC), in Europe the European Central Bank (ECB) and national regulators. Generative AI tools that draft regulatory narratives are being piloted, but production use remains limited. Vendors pitch "AI that writes your regulatory filings," but compliance teams still require full human review, and no major regulator has signaled comfort with unreviewed AI-generated filings.
Where it's oversold: end-to-end automated regulatory reporting without human sign-off.
When evaluating any AI claim in the value chain, ask three questions:
1. Is the task narrow and well-defined? (Document verification, fraud scoring, OCR: yes. Open-ended advice or judgment: usually no.)
2. Is there a large, labeled historical dataset to train on? (Fraud has decades of labeled examples. Novel product recommendations often don't.)
3. Does a wrong answer carry regulatory or reputational risk that requires human sign-off anyway? (If yes, "AI-powered" often really means "AI-assisted, human-approved.")
Simple due-diligence checklist for any vendor AI claim:
1. What % of decisions are fully automated vs. flagged for human review?
2. What's the false positive / false negative rate, and versus what baseline?
3. Can the model's output be explained to a regulator or an auditor?
4. What happens when the model is wrong, who is accountable?Knowledge check
1. The lesson frames the deposit's journey as a way to teach a specific analytical skill. What is that skill?
2. Why does the lesson highlight that fraud scoring models have been in production since the 2000s, well before 'AI' became a boardroom buzzword?
3. A vendor claims its onboarding tool uses 'AI-powered identity verification.' Based on the lesson's framework, what would best confirm this is genuine AI rather than a marketing label?
4. Select ALL correct answers describing characteristics of genuine, mature AI fraud-scoring in banking as described in the lesson.
Select all the correct answers.
5. Select ALL correct answers about the purpose of tracing a single transaction through multiple systems in this lesson.
Select all the correct answers.
When a vendor or internal team pitches "AI across the customer journeycustomer journeyThe full sequence of touchpoints a customer has with your brand before, during and after purchase, spanning awareness, consideration, decision, retention and advocacy.View full definition →," the deposit trace above gives you a mental mapmapUsing software to automate repetitive marketing tasks and campaigns, enabling personalisation at scale across channels like email, web, and social.View full definition → to press on specifics. Ask which of the eight stops they mean. Fraud scoring and document verification have mature ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition → cases (fraud losses reduced, onboarding time cut from days to minutes). Generative AI in regulatory reporting and fully autonomous personalization do not yet have mature, broadly-verified ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition → cases as of 2026, they're pilots, not production at scale.
For a grounded look at where banks are actually spending AI budgets versus where the hype concentrates, the McKinsey Global Banking Annual Review is a useful, free, regularly updated source.
🎬 [VIDEO: "How AI is Used in Banking (Real Examples)" - youtube.com - search for recent explainer content from banking technology channels covering fraud detection, credit scoring, and chatbot deployment with concrete institutional examples]