The AI-Agent Era in Banking Has Barely Begun

Most banks are still piloting agentic AI — not deploying it. That is the central finding of a new analysis from the American Bankers Association, which warns that despite headline-grabbing announcements from Goldman Sachs, Lloyds, and JPMorgan, the financial industry has only started to enter the agentic AI era — and governance frameworks are lagging even further behind than deployments.

The stakes are getting harder to ignore. Lloyds Banking Group alone projects that AI agents will generate £100 million in value in 2026 by autonomously handling fraud investigations and resolving complex customer complaints. That figure is a preview of what becomes available across the industry when more institutions reach full deployment — and a measure of how far most banks have yet to travel.

What Agentic AI Actually Means in Banking

There is a meaningful distinction between AI tools and AI agents that most coverage blurs. An AI tool helps a human analyst: it summarizes documents, flags anomalies, or drafts responses. An AI agent acts: it receives a fraud alert, pulls transaction history, cross-references behavioral patterns, contacts the customer, escalates if needed, and logs the resolution — without a human touching any step.

That autonomy is the value proposition. It is also what makes governance so difficult.

Goldman Sachs has deployed AI tools broadly — its internal LLM Suite serves tens of thousands of employees across trading, legal, and research functions. JPMorgan's comparable system serves more than 60,000 employees. BBVA, Santander, and several large regional institutions have automated customer service and document workflows at scale. These programs are real and growing.

But the ABA's analysis is clear: these are mostly AI tools, not AI agents. Institutions running genuinely agentic systems — where AI takes consequential multi-step actions with minimal human involvement in the loop — remain a small minority. Most are in structured pilots on narrow, low-stakes workflows.

The Governance Frameworks Are Not Ready

The deeper problem, according to the ABA, is not the pace of deployment. It is that most banks' governance infrastructure was designed for deterministic, rule-based software — not for systems that reason and adapt.

A rule-based fraud detection system is straightforward to audit: did it follow the rules? An agentic system does not follow rules. It weighs evidence, makes probabilistic judgments, and takes actions that vary by context. Auditing that process requires capturing the agent's reasoning at each decision point, building override mechanisms so humans can intervene when the agent misjudges, and ensuring staff understand the system's behavior well enough to catch errors.

Most banks do not have those capabilities in place. Audit frameworks built for deterministic software are being stretched to cover probabilistic reasoning systems — a mismatch that creates real compliance exposure as the OCC, CFPB, and Federal Reserve move toward formal AI model governance guidance.

Who Is Ahead and What They Are Doing

Lloyds is arguably the clearest case study in what agentic deployment looks like at meaningful scale. The bank is not automating simple FAQ responses — it is putting agents on fraud investigations and complex complaint resolution, high-stakes multi-step workflows that require context, judgment, and regulatory defensibility. The £100 million value projection for 2026 reflects that these are not narrow toy use cases.

Goldman and JPMorgan have broad AI infrastructure — data pipelines, LLM access, and technical staff — giving them a shorter path to agent-scale deployment than most. What they are building now is the governance layer: the oversight mechanisms, audit trails, and human-override protocols that would let agents operate at full autonomy in regulated workflows.

What the Rest of the Industry Risks

Falling behind on agentic AI in banking is not a neutral position. Institutions that deploy agents for fraud monitoring, regulatory reporting, loan processing, and customer operations will build a structurally lower cost per transaction. The advantage compounds: faster fraud resolution reduces chargebacks, lower analyst overhead creates capacity for complex work, and better customer service outcomes improve retention.

Banks waiting for governance frameworks to mature before deploying risk waiting for a problem that gets solved in practice, not in advance. The institutions writing the governance playbook are the ones currently running agents in production and learning what supervision actually requires.

The competitive gap that results from a two-year deployment lag will not be easily closed once it opens.

What to Watch

Formal AI model governance guidance from federal regulators — OCC, CFPB, and the Federal Reserve — is the variable that will determine how fast the rest of the industry moves. Specific requirements around audit trails, explainability, and human override protocols will give risk-averse compliance teams the framework they need to approve production deployments. Until that guidance lands, watch which banks are building their own governance standards and whether those standards are holding up under examination.

By Hector Herrera