Agents everywhere, agentic banks nowhere

More than 80 per cent of companies still see no measurable return on their AI investments. That is the central figure Alexis Krivkovich, Senior Partner at McKinsey, placed at the opening of her early-April analysis "AI is everywhere. The agentic organization isn't—yet". For banks, this is no footnote. It is the question of whether Agentic AI becomes an internal efficiency lever – or the point at which the classical competitive advantages of banking begin to slip.

The paradox the industry has been building towards for two years

The gap between investment volumes and realised business impact is by now well documented. McKinsey calls it the "great paradox": boards have released budgets in anticipation of a major transformation, the pilots are running, the models are improving – yet the profit and loss statement remains conspicuously quiet. Alexis Krivkovich, in her conversation with Lucia Rahilly, Global Editorial Director at McKinsey, identifies the real fault line: the challenge does not sit in the technology. It sits in the redesign of workflows, leadership roles and culture.

This matches the pattern emerging in German banks. The models work, the use cases are identified, the pilot units deliver impressive demos. Yet the moment a prototype is expected to move into regulated production, it gets stuck in process ownership disputes, governance committees and operating-model questions for which the traditional org charts hold no clear responsibility. The difference between pilots with impact and pilots in the museum, as McKinsey puts it, is not a question of the model – it is decided by the organisation around it.

Five pillars, read through the banking lens

McKinsey structures the "agentic organisation" along five pillars: business model, team structures, workflows, leadership, culture. For financial institutions, the implications can be mapped onto each of these in turn.

On the business model, Alexis Krivkovich uses the image of "near-zero marginal cost of delivery" – and offers an explicit banking example: a customer whose agent moves funds frictionlessly between institutions to capture the best available rate. What reads as a thought experiment in the McKinsey text is, at the European level, already regulatory programme: the Financial Data Access Regulation (FiDA) obliges financial institutions to release defined customer data via standardised interfaces to authorised third parties – the regulatory birth certificate of Open Finance. It is in precisely this intersection of open-finance rules and agentic customer tools that Alexis Krivkovich's scenario becomes operational. What used to stabilise the deposit base as structural inertia would then move in days rather than years.

At the team level, the implications are less dramatic but equally consequential. The "operating-model shift" McKinsey describes is about the rituals of the working day: how does a trading desk start the morning when an agent population has been aggregating data and preparing scenarios overnight? How does an underwriting team steer work when routine cases have already been decided by agents and only edge cases reach human attention? The answer is not "more AI" – it is a new cadence.

On workflows, McKinsey goes sharpest into the detail. The break with the past years of piloting lies in abandoning point solutions: instead of streamlining a single task, entire end-to-end processes are redesigned. For banks, that means not "agent for loan application review" but "Hire-to-Onboard", "Incident-to-Resolution", "Complaint-to-Redress" – the processes that today run across compliance, operations and front office, losing their time in handovers.

Leadership and culture, finally, form what McKinsey calls "the real bottleneck" – and here the analysis intersects with an uncomfortable truth for the German market: the era of episodic change programmes is over. Alexis Krivkovich speaks of a "perpetual state" of change. For banks that have only just finished their last reorganisation, that is not a reassuring message.

From Human in the Loop to Human above the Loop

The conceptual shift McKinsey introduces in this piece is arguably the most consequential: from Human in the Loop to Human above the Loop. "In the loop" means humans and agents take turns inside a process, with control distributed across them. "Above the loop" means a team of agents runs the entire core process, with the human deciding at the end.

Alexis Krivkovich illustrates the distinction with the American Arbitration Association. Agents there now sift through thousands of data points per arbitration case, construct timelines, examine both sides of the argument and propose a decision. Humans rule on the outcome. McKinsey reports that, in some cases, the agents perform better than the original process.

For banks, this strikes a structural nerve. Complaint management, suitability assessments under the Markets in Financial Instruments Directive II (MiFID II), case review in anti-money laundering (AML), sanctions screening, internal control reports under the Minimum Requirements for Risk Management (MaRisk), incident reporting under the Digital Operational Resilience Act (DORA) – each of these is a process built on data aggregation, pattern recognition and documented judgment. "Above the loop" is, for them, less a framing than an operating model. The regulatory question attached to it is no longer academic: how deeply are agents allowed to prepare decisions before human responsibility steps in? And how does an institution document that the "judgment layer" does not deteriorate into a rubber-stamping ritual?

Three in four roles are being redesigned

Seventy-five per cent of all roles, McKinsey writes, now require fundamental reshaping. That is a figure that produces reactions in board meetings – and one that lands in German banks on a very specific infrastructure of works councils, collective agreements and codetermination. The message is rarely delivered as bluntly as McKinsey puts it. But it is the same: almost every job description in the house will be a different one in two to three years.

Alexis Krivkovich names precisely which capabilities are rising in value: strategic thinking, systems orientation, leadership and judgment. And which – while still essential – are increasingly delivered with agentic support: research, quantitative analysis, data science. What matters is less the list than what it implies for the talent pipeline.

The genuinely uncomfortable question is the junior-to-senior gap. The tasks on which junior staff used to develop their judgment – slow case reviews, manual reconciliation, patient reading of files – are the first to move into agents' hands. Remove that layer without designing a new development path, and in ten years you have a house built only of expensive senior professionals, with no next generation behind them. Alexis Krivkovich calls this the "billion-dollar question". For institutions with their own talent programmes, it is the moment at which learning and development have to move out of the sidecar and into the operational core.

The real threat sits outside the balance sheet

The sentence Alexis Krivkovich delivers almost in passing is the most strategically important of the entire article: "That fundamentally changes the moat that has existed in financial services since the beginning of time." The moat – a term popularised by Warren Buffett as the image for durable, hard-to-imitate competitive advantage – describes, in banking, that mix of switching costs, information asymmetry and relational inertia which has so far protected the deposit base from daily optimisation. The point is the moment in which it is not the bank but the customer who deploys agents – agents that remove friction from a relationship which has, until now, lived precisely on that friction.

The regulatory lever that accelerates this movement is not a market phenomenon but European legislation. The Financial Data Access Regulation defines which categories of financial data institutions must release to authorised third parties – from payment accounts through savings products, pensions and insurance. Customer-side agents are thereby made not only technically possible but legally addressable. What the Revised Payment Services Directive (PSD2) opened up for payment accounts, FiDA extends to almost the entire financial portfolio of the customer. Opening an account, closing one, switching provider, comparing rates, rebalancing savings plans: the processes that keep the customer bound to their house bank today are not inert because they are objectively hard but because their friction costs rest on information asymmetry, form logic and switching effort. Remove that friction – because a customer agent absorbs it along standardised FiDA interfaces – and the inert deposit base turns into a daily optimisation problem.

That McKinsey makes this point in an HR and organisation piece is no accident. The response to competition for agent-equipped customers will not be given in product design. It will be decided by whether an institution moves fast enough internally to position its own interfaces inside the emerging customer-agent logic – and whether the organisation keeps pace.

Recommendations for financial institutions

For institutions seeking to turn McKinsey's findings into a programme, five steps are decisive over the next twelve months: