McKinsey: Agentic AI Could Compress Banking Profit Pools by 9–10%
McKinsey's 2026 banking research frames agentic AI as a margin event, not an efficiency story — and warns slower adopters about 'pilot purgatory'.
What Happened
McKinsey's 2026 banking research — the Global Banking Annual Review alongside its 'Move first or fall behind' analysis — estimates that agentic AI could compress industry profit pools by 9–10% as automation shifts value between banks and toward technology-enabled challengers. The firm notes that 50–60% of a typical bank's full-time equivalents are tied to operations, and that early agentic deployments have cut manual workloads by 30–50% in zero-touch processes. Its warning for slow movers has a name: 'pilot purgatory' — an endless cycle of narrow experiments that never scale.
Why It Matters
This is one of the first mainstream estimates that frames agentic AI as a repricing of banking economics rather than a productivity side-story. A 9–10% profit-pool compression is not evenly distributed: institutions that industrialize first capture the cost advantage and the market share; the rest effectively donate margin. It also reframes the AI budget conversation — when the downside of inaction is measured in profit pool, AI stops competing with other innovation projects and starts competing with strategic survival.
Banking & Fintech Implications
Move from use-case counting to operating-model redesign: pick two or three end-to-end journeys — onboarding, disputes, credit operations — and rebuild them agent-first with explicit human oversight gates. Put agent governance (mandates, audit trails, model-risk treatment) in place while scaling, not after. And measure success in unit cost and cycle time, not in the number of pilots — pilot count is precisely the metric that purgatory optimizes.
My Take
Pilot purgatory is a governance failure, not a technology one. In every organization I have watched escape it, the difference was ownership: the executive committee owned the scaling decision and the operating-model consequences, instead of leaving both to an innovation lab. The 9–10% number will be debated — the direction will not.