Banks Are Moving AI from Advisor to Decision-Maker — and Regulators Haven't Caught Up

Banking is entering what American Banker's exclusive research calls an "agentic era" — a shift from AI systems that recommend actions to AI systems that execute transactions autonomously. Lloyds Banking Group is planning enterprise-wide agentic deployment. JPMorgan Chase is already using agentic AI for investment banking workflows. McKinsey estimates the technology could cut bank operating costs by up to 20%. And regulators have not yet defined what guardrails autonomous financial agents must operate within.

What "Agentic" Means for Banking

In its previous form, AI in banking was advisory: a model analyzes a credit file and surfaces a recommendation; a compliance tool flags a suspicious transaction for a human to review; a chatbot answers a customer question and escalates when it can't. A human remains in the loop for every consequential decision.

Agentic AI removes that loop. An agentic financial system can:

• Execute fund transfers based on pre-authorized criteria without per-transaction approval
• Route and structure commercial loans using AI-generated analysis without a junior banker reviewing the file
• Rebalance portfolios in real time based on market conditions within a defined risk envelope
• Generate, revise, and internally approve pitch-deck materials for investment banking without analyst intervention

The American Banker research frames this as a shift from AI as an assistant to AI as an agent with transactional authority — the ability to commit the bank to actions, not just recommend them.

What the Major Banks Are Building

Lloyds Banking Group is planning enterprise-wide agentic AI deployment, according to the American Banker findings. Lloyds is one of the UK's largest retail banks, with roughly 26 million customers. An enterprise-wide deployment means agentic systems operating across customer-facing products, back-office operations, and risk management simultaneously — not a single pilot project.

JPMorgan Chase is already using agentic AI inside its investment banking division for pitch-deck generation and internal workflow automation. The specific application — having AI draft, revise, and prepare presentation materials that previously required associate-level work — is a narrower deployment than Lloyds' enterprise scope, but it represents JP Morgan's first step from AI assistance to AI execution in a revenue-generating business unit.

The McKinsey estimate of up to 20% net cost reduction is the figure banks are using internally to justify the investments. For a bank like JPMorgan, with total non-interest expenses exceeding $80 billion annually, even a 10% realization would represent billions in savings. That economics case is overriding whatever caution might otherwise slow a bank's risk appetite toward autonomous systems.

The Regulatory Gap

This is where the story gets complicated.

Banking is one of the most heavily regulated industries in the world. Every system that processes customer transactions, evaluates credit, or executes trades operates within a framework of rules — capital requirements, fair lending laws, fiduciary duties, KYC (know your customer) and AML (anti-money laundering) obligations. Those rules were written for systems where a human being remains accountable for each consequential decision.

Agentic AI does not fit that framework. When an AI agent executes a transaction autonomously, the question of accountability — who approved this, who verified it was compliant, who bears liability if it causes harm — does not have a clean answer under existing regulations.

Neither the Federal Reserve, the OCC (Office of the Comptroller of the Currency), the FCA in the UK, nor the European Banking Authority has published guidance specifically covering AI agents with transactional authority. The American Banker research notes that regulatory guardrails for autonomous financial agents remain "undefined" — a significant gap as deployment scales.

Banks are managing this by building their own internal guardrails: hard limits on transaction size, mandatory human review thresholds, explainability requirements for AI-generated credit decisions. But these are voluntary internal controls, not regulatory requirements. Different banks are setting different limits, creating an uneven playing field.

What Could Go Wrong

The risk scenarios that regulators are most focused on, based on public statements from banking supervisors in 2026:

Fair lending violations at scale. If an agentic credit system encodes bias — even unintentionally — it can issue biased decisions across millions of applications before anyone notices. A human loan officer making biased decisions is a compliance problem. An AI agent doing so systematically is a systemic risk.

Liquidity cascades. Multiple agentic systems from different banks responding to the same market signal simultaneously — all programmed to reduce exposure to the same asset class — could accelerate a market move rather than simply react to it. This is the AI equivalent of the portfolio insurance strategies that contributed to the 1987 stock market crash.

Accountability gaps in disputes. When a customer disputes an AI-executed transaction, who is responsible? The model vendor? The bank's deployment team? The product manager who set the authorization envelope? That question does not have a settled legal answer.

What to Watch

The US Consumer Financial Protection Bureau is expected to issue preliminary guidance on AI decision-making in consumer finance before the end of 2026, and the OCC has signaled it will address agentic systems in its 2027 supervisory priorities. In the UK, the FCA's AI Lab has been running sandboxed agentic deployments with select banks since mid-2026. When regulatory frameworks do arrive, banks that have been operating ahead of them will need to retrofit compliance — a cost that could partially offset the savings they are currently banking on.

By Hector Herrera