What most banks get wrong about agentic AI: the four bets that actually pay off

Agentic AI is reshaping how banks operate, just not in the way most executives assume. A dominant narrative holds that agentic AI will replace people with automation, but that assumption misses where the real impact is landing. Research emerging through the middle of 2026 points to a clearer answer: not the banker, but the process, review and hand-off machinery sitting between the banker and the customer.

That distinction matters for anyone deciding where to spend the next quarter of AI delivery capacity. It shows up on the revenue side too. We have argued elsewhere that the biggest problem in FX sales isn't market volatility, it's decision latency: the highest-value use of agentic AI is rarely replacing the relationship. It is clearing the operational noise around it.

‍

Where banking leaders expect the impact

Surveys of banking executives point to compliance, risk and audit as the sharpest edge of near-term agentic AI deployment, not customer service. Over the next three years, a majority expect AI agents to be fully embedded in risk, compliance, audit, fraud detection and transaction monitoring, with similarly strong expectations for credit assessment, loan processing and know-your-customer functions. That is not a customer-experience story. It is an operations and control-function story.

What is driving that expectation is a capability shift, not just faster automation. Earlier generations of robotic process automation could execute a fixed sequence of steps. Agentic systems can decide which sequence applies, managing less structured, judgment-dependent tasks like exception handling and one-off analysis, the kind of work that used to require a human to read context and decide what to do next. Gartner projects that at least 15 percent of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from 0 percent in 2024, and that 33 percent of enterprise software applications will include agentic AI by 2028, up from less than 1 percent in 2024.

That is a different animal from the copilots most banks piloted first. We have written before about why the next phase of AI demands reinvention, not experimentation, and the same line separates rules-based automation from agentic decision-making in where each works, where each breaks, and why it matters.

‍

Where the evidence actually is: compliance, not customer service

The most concrete proof point so far is not the customer-facing chatbot. It is compliance review, historically one of the slowest, most document-heavy functions in any bank.

What "agent skills" actually are

A pattern showing up across recent industry analysis is the rise of what is often called agent skills: modular units of encoded expertise that sit between a model's general reasoning ability and a bank's actual compliance rules and workflows. Not the chat interface the user sees, but the regulatory logic and operational systems the agent has to work within. A methodology skill, a validation skill and a reporting skill can each govern a different dimension of an agent's behaviour, and, critically for a regulated industry, each is version-controlled, reviewed and auditable the same way code is.

The practical effect is speed of a specific kind. Because these skills are written and reviewed as natural-language documents rather than application code, updating a compliance workflow when a regulation changes can take hours instead of the weeks or months a code change would typically require. For a compliance function that lives or dies by how quickly it can respond to a shifting regulatory landscape, and 2026 has given every bank plenty of shifting regulatory landscape to respond to, that is a fundamentally different operating tempo, not just a productivity bump. It is the same architecture principle behind our Sanctions L1 Triage work: encode the expertise once, review it like code, and let the workflow keep pace with the regulation instead of trailing it by a quarter. We go deeper on why L1 triage needs a fundamentally different AI model than legacy screening tools, and on whether AI can be trusted in compliance decisions at all.

‍

The pattern separating pilots from production

The gap between AI activity and AI outcome is now well documented, and it is a large gap. Gartner projects that more than 40 percent of agentic AI projects will be abandoned before the end of 2027, due to escalating costs, unclear business value or inadequate risk controls, and largely because legacy systems were not built to support autonomous execution in the first place. It is the same gap we unpacked in why AI pilots in banking rarely scale, where early proof-of-concept success rarely survives contact with a live production environment.

MIT's Project NANDA put a sharper point on the same problem, reviewing more than 300 enterprise AI initiatives. Despite 30 to 40 billion dollars in enterprise generative AI spending, roughly 95 percent of organisations report no measurable P&L impact, while a small minority, about 5 percent, extract real, sustained value. The differentiator NANDA's researchers identified was not model quality, budget or even talent. It was whether the system could learn, retain context and adapt over time within a specific workflow, rather than resetting with every interaction. It also tracks with what we have seen when banks measure the wrong thing to begin with: most AI metrics in banking are misleading, which makes the 95 percent easy to misread as a technology failure when it is usually a measurement and design failure.

95% of enterprise gen AI initiatives show no measurable P&L impact. The 5% that do share one trait: a coordination layer, not a better model.

The institutions in that 5 percent share a common trait. They are not running standalone bots bolted onto an existing process. They are running agents coordinated through a shared operational layer: common customer context, consistent governed authority and a complete audit trail across every agent touching the same transaction. Without that coordination layer, adding more agents just scales the mistakes alongside the throughput, which is a fair description of what breaks first when banks deploy AI agents and what stalled pilots look like from the inside. It is the same maturity gap we mapped on the manufacturing floor in where does your plant sit on the AI maturity curve and the same shift, from static reporting to systems that can actually decide something, that we walked through in From Dashboards to Decisions. The pattern holds across industries: coordination is what turns a pilot into an operating model.

‍

Four bets that actually pay off

If you are deciding where to spend your next quarter of AI delivery capacity, the research points in a consistent direction.

1. Target operations first, compliance sharpest. It is where structured judgment, heavy documentation and repeatable-but-not-identical decisions intersect, exactly the profile agentic AI handles better than either rules-based automation or a general-purpose copilot.

2. Build the skill layer, not just the model connection. Encoding institutional expertise as reviewable, version-controlled assets is what makes an agent's output defensible to a regulator and updatable at the pace regulation actually moves.

3. Invest in the coordination layer before adding more agents. Shared context and consistent governance across agents is what separates scaled throughput from scaled mistakes, and it is a prerequisite, not a nice-to-have, once more than one agent touches the same customer or transaction, as we saw in what breaks first when banks deploy AI agents.

4. Redesign the workflow, do not retrofit the agent into it. The 95 percent no-impact outcome is not a model capability problem. It is an architecture and organisational-design problem, and it shows up consistently across the research base on enterprise AI adoption this year. See what production-ready AI actually means in financial services and how to run an AI proof of concept in a regulated bank for how to sequence the rebuild without stalling the business you already have.

‍

Where this leaves your next quarter

Most banks chasing agentic AI right now will get an efficiency-sized result, if any. The ones treating it as a redesign of how operations, compliance and service actually work are the ones showing up in the 5 percent.

That is the architecture principle behind Sanctions L1 Triage: the skill layer and the coordination layer are the design, not an add-on. If you are mapping where agentic AI earns its place in your operating model, here is what to look for in an AI agent implementation partner, and that is the conversation worth having next.