Economic Essay Contest AI in financial services: Efficiency is not enough
The financial industry has embraced artificial intelligence (AI) with remarkable speed. Algorithms approve loans in seconds, chatbots resolve complaints around the clock and fraud detection systems process billions of transactions daily. By almost every operational metric, AI is delivering. Yet there is a growing risk that the industry is measuring the wrong things. The dominant narrative — centered on cost reduction and automation — is dangerously incomplete. AI in financial services must be reframed, from an efficiency tool to a trust infrastructure. That reframing has implications across customer service, product development, and security.
Customer service: The efficiency trap
The most common measure of AI success in customer service is cost. McKinsey estimates that AI-driven automation could reduce customer service costs in banking by up to 30 percent. This is a legitimate gain — but institutions that optimize purely for cost reduction risk hollowing out the relationship that banking fundamentally depends on.
The contrast between two approaches is instructive. When Wells Fargo deployed its AI assistant Fargo, early customer feedback centered not on the technology's limitations, but on its inaccessibility to human escalation — a deliberate cost-containment design choice. Bank of America's Erica, by contrast, logged more than 2 billion client interactions by 2024 — not by replacing human advisors, but by triaging complexity: handling routine inquiries while routing sensitive conversations to human agents. The difference is philosophical before it is technical. Erica was designed around the question, what does the customer need? Cost efficiency followed as a consequence, not a goal.
The implication is clear: Institutions must stop measuring AI success by deflection rates — how many calls avoided a human — and instead measure trust indicators: resolution satisfaction and long-term retention.
Product development: Inclusion as strategic obligation
AI's most transformative potential lies in credit underwriting, where traditional models have systematically excluded entire populations. The Federal Reserve estimates that approximately 45 million Americans are "credit invisible," while the World Bank places unbanked adults globally at 1.4 billion.
Upstart has demonstrated that this exclusion is not an inevitable feature of responsible lending — it is a limitation of outdated models. By incorporating more than 1,600 variables including employment patterns and cash flow volatility, Upstart produced default rates approximately 75 percent lower than comparable traditional benchmarks while approving significantly more borrowers from underserved demographics. AI-driven inclusion and sound risk management are not in tension. Institutions that treat them as such are rationalizing inertia.
However, inclusion without explainability is not trust — it is opacity with better outcomes. The EU AI Act and Korea's emerging AI governance framework both mandate that automated credit decisions be explainable in plain, auditable language. Institutions must therefore invest in "explainable AI" frameworks as a structural commitment, not a compliance checkbox. When a customer is denied a loan by an algorithm, they deserve to understand why. An institution that cannot explain its own decisions has not earned the authority to make them.
Security: The asymmetry problem
AI has altered the security landscape, and not symmetrically in the industry's favor. Cybercrime targeting financial institutions cost an estimated $6 trillion globally in 2023, and was projected to reach $10.5 trillion by 2025. Adversarial AI — deepfakes, AI-generated phishing, model poisoning — is already outpacing legacy defense frameworks.
JPMorgan Chase has responded with investment that signals the existential dimension of this risk: cumulative cybersecurity expenditure exceeding $15 billion over five years, with AI processing more than 5 billion data events daily. Mastercard's Decision Intelligence Pro, launched in 2024, applies generative AI to over 1 trillion data points annually, reducing false transaction declines by up to 85 percent while improving fraud detection rates — proving that security and customer experience need not be traded against each other.
The deeper argument is this: security is not a risk management function — it is the foundation of financial trust itself. Institutions that underinvest in AI-native security are not simply accepting operational risk. They are gambling with their institutional identity.
Customer service, product development, and security are often treated as separate silos, but they should not be. Each is a different expression of the same challenge: How does a financial institution earn and sustain trust in an era of radical technological change? AI that obscures human access erodes trust. AI that makes opaque decisions erodes trust. AI security that fails to evolve erodes trust. Conversely, AI designed around explainability, inclusion and genuine customer outcomes builds the institutional credibility no marketing campaign can manufacture. The institutions that will lead the next decade are those that understand the principle their industry has always known. In finance, everything — ultimately — is built on trust.