Financial services is the sector that spent the most on AI between 2023 and 2025, announced the most ambitious AI programs, and deployed the most AI — but the deployment picture looks almost nothing like the announcement picture. The systems running in production at banks, insurers, and asset managers are largely not the systems that got press releases. The systems that got press releases are mostly still in extended pilots.
Understanding the gap requires distinguishing between two fundamentally different AI generations operating simultaneously inside financial services institutions. One is the new generation — the LLM-based tools for customer service, wealth management, compliance, and internal knowledge management that dominated announcements after 2022. The other is the old generation — machine learning and statistical AI systems for fraud detection, credit scoring, trading, and risk modeling that have been in production for fifteen to twenty years and predate the current wave entirely.
When banks report that they “have AI in production,” they are nearly always talking about the old generation, with incremental improvements. The new generation is, with real exceptions, still crossing the pilot-to-production gap described elsewhere.
The Old Generation That Actually Works
JPMorgan Chase’s AI program most cited by financial services journalists involves a system called IndexGPT, an LLM-based investment advisory tool. What is less reported is that JPMorgan’s actual production AI capability is much older: fraud detection models processing real-time transactions, credit risk models running across tens of millions of accounts, and trading algorithms operating in the markets every day. These systems have been running, iterating, and improving for over a decade. They are boring relative to an LLM. They are also consequential at a scale that makes the LLM pilots look like prototypes.
The same is true across the sector. Visa’s real-time fraud detection AI has been operating in production since at least 2016 and currently processes transactions at a rate that makes it one of the highest-volume AI systems in the world by any reasonable measure. American Express’s credit risk models incorporate hundreds of variables and have been updated continuously through multiple technology generations. These are not new-era AI — they are ML engineering achievements that the current LLM excitement has temporarily overshadowed.
The reason these systems work at production scale while the LLM-based ones struggle is instructive: they operate in well-defined decision spaces with labeled training data, clear success metrics, and feedback loops that allow continuous improvement. A fraud detection model can be evaluated against ground truth (was the transaction actually fraudulent?) with high precision and at high volume. The evaluation loop is tight and automatic. The model knows when it is wrong.
LLM-based financial services applications — AI advisors, document review systems, compliance chatbots — operate in messier spaces where ground truth is harder to establish, evaluation is more qualitative, and errors are harder to detect automatically. This is not a technical limitation of LLMs specifically. It is a property of the problem domain.
Where the New Generation Has Actually Deployed
The LLM-era AI applications that have genuinely made it to production in financial services cluster in specific places that share a common characteristic: they assist human judgment rather than replacing it, and their output is reviewed before it creates any downstream commitment.
Internal knowledge management tools — systems that help employees search across internal documents, policies, and training materials — have deployed relatively widely because the failure mode is low-stakes. If the system returns an incomplete or slightly wrong answer, the employee typically catches it and either uses other sources or asks a colleague. The tolerance for imperfection is high because the human is still in the loop.
Regulatory document analysis tools used by compliance teams have deployed at a meaningful number of institutions, for a similar reason. The compliance analyst uses the AI to do initial categorization and summarization of regulatory filings, then reviews the AI’s output before it informs any decision. This is the “AI as a really fast first reader” model, and it works because the value is efficiency rather than judgment replacement.
Earnings call and financial report summarization tools have been adopted fairly widely among research and investment teams, again because the output is reviewed by a human before it drives an investment decision. The risk tolerance is defined by the human review step.
Where the new generation has notably failed to deploy at scale is in the places where the announcement pitch was most exciting: customer-facing wealth management advice, automated compliance sign-off, real-time document review for M&A and lending, and anything involving regulatory liability.
The JPMorgan Case: What IndexGPT Actually Is
JPMorgan’s IndexGPT — announced in May 2023 with significant press coverage — was described as an AI thematic investment advisor that would “select securities that fit a given theme.” By late 2024, it had been trademarked, piloted, and partially deployed to wealthy clients in a very specific way: as a research tool that suggested thematic portfolios for human advisors to review and potentially recommend, not as a direct-to-customer AI advisor.
This is not a failure. It is a sensible deployment of an AI capability in a context where the regulatory environment makes fully automated investment advice a minefield (SEC Rule 15c3-5, Regulation Best Interest, and several state-level regulations create liability exposure for automated advice that does not meet specific standards). The human advisor in the loop is not just a product choice — it is a regulatory necessity.
The same dynamic plays out across wealth management at Goldman Sachs (whose “GS AI” internal toolkit is more widely used by analysts than by clients), Morgan Stanley (whose AI assistant, built on OpenAI technology, is actively used by financial advisors in a human-in-the-loop configuration), and smaller wealth managers who are waiting to see how the regulatory environment evolves before committing to customer-facing deployment.
The Compliance AI Time Bomb
The most significant unrealized AI potential in financial services is probably in compliance — anti-money laundering monitoring, suspicious activity reporting, know-your-customer documentation review — and the reason it has not deployed at scale is a combination of regulatory risk and vendor immaturity.
Large financial institutions generate compliance alerts at volumes that far exceed human review capacity. The vast majority of AML alerts are false positives. The human review process is expensive, slow, and subject to analyst fatigue and inconsistency. An AI system that could reliably reduce the false positive rate by 30-40% while maintaining detection quality would represent a significant operational improvement.
The technical capability exists to build such a system. Several vendors have demonstrated it in controlled environments. The reason it has not deployed broadly is that regulators have been cautious about approving AI for core compliance functions — for good reasons, given the liability and enforcement implications of compliance failures. The OCC, FinCEN, and Fed have issued guidance that acknowledges AI’s potential in BSA/AML compliance while requiring extensive model validation, explainability documentation, and change management procedures that most institutions have not yet completed.
This is the typical story of regulated-industry AI: the technology is ready before the governance framework is. The governance framework is not unreasonable — it is actually appropriate given the stakes. But it means that the sector where AI could have the most compliance impact is also the sector where deployment will take the longest.
The banks that have done this work — that have invested in model validation infrastructure, built explainability tooling into their AI systems, and worked proactively with regulators to develop acceptable deployment frameworks — will be significantly better positioned when regulatory clarity arrives. The ones that are waiting for clarity before starting the infrastructure work will be eighteen months behind.
One email a week with practical Mac workflows, useful comparisons and honest tech recommendations.