McKinsey has published one of the most substantial analyses of the year on the banking sector, “The Paradigm Shift: How Agentic AI is Redefining Banking Operations”. The headline finding is not about technology. It is about why most banks cannot move past experiments.
What Makes Agentic AI Different from Generative AI
Let’s be clear. Generative AI, the kind behind tools like ChatGPT, answers questions. It produces text, summarizes, and explains. Agentic AI does something different. It takes action. It can plan a sequence of steps, carry out tasks across multiple systems, make decisions in the middle of a process, and adapt to the outcome, without continuous human oversight.
Picture an AI agent that does not simply show you a client’s data, but checks the documents, verifies compliance, assesses the risk, drafts the report, and notifies the responsible person, in a few minutes instead of days.
Why Banking Operations Are the First Target
In banks, 50 to 60% of the workforce is engaged in operations. That means processing requests, checking documents, managing transactions, providing customer service, and compliance. McKinsey estimates that Agentic AI can radically transform that work and cut operating costs by more than 20%. The first use cases already tested show a 30 to 50% reduction in manual workload for specific processes.
Why the Cost Advantage Will Not Last
This advantage will be competed away quickly, exactly as happened with internet banking, mobile banking, and first-generation chatbots. What sets the winners apart is not whether they adopt the technology. It is when, and how deeply. If the sector as a whole fails to transform, global banking profits risk falling by up to 10% over the next 10 years, as fintechs and neobanks that start out AI-native take market share.
Why Most Banks Are Stuck in Pilot Purgatory
So why have most banks not already moved? This is where the real problem McKinsey identifies sits, what it calls pilot purgatory. Banks test AI on isolated use cases, a chatbot here, an automation there. They see results, but they do not scale. They stay stuck in an endless stage of experimentation that never delivers meaningful value.
The question is not where to put AI. It is how to organize work and decisions in a company where software can act autonomously.
The Real Challenge Is Governance, Not Technology
Here is the harder version of that question. How do you scale Agentic AI in a sector where every autonomous decision has to be fully auditable (which agent acted, when, and on what data), explainable (why the decision was made), and compliant with DORA, MiFID II, GDPR, and Solvency II?
This is not a technology problem. It is a problem of governance architecture, the system that governs how AI operates inside a strictly regulated environment.
