Data readiness: the number one AI obstacle for 48% of banking executives
The AI disruption dividend, part two of McKinsey's Global Banking Annual Review published in November 2026, contains one figure we find more telling than any productivity promise.
48% of the banking executives surveyed rank data readiness as their top obstacle to AI. Technology comes last. We set out below what this means for a bank or a large group in Switzerland and Europe.
What McKinsey's November 2026 report says
McKinsey surveyed 85 banking executives in its Banking Executive Survey on AI 2026. The obstacles they cite rank in this order:
data readiness, for 48% of respondents;
talent, for 41%;
change management, for 37%;
operating model and strategy, for 33 to 35%;
technology, for only 27%.
The report sums up this hierarchy in one sentence: technology is the smallest problem. Every bank can buy the same models. No bank can buy another one's data, rules and past decisions.
The same report puts a number on what is at stake. Across 52 agentic AI use cases analysed, the average return reaches about 2.9 times the investment over three years, with a payback period of roughly ten months.
It adds a warning we take seriously: a late start may not offer a chance to catch up, depending on how early gains are reinvested. The window, McKinsey writes, will start closing.
Why data is the blocker, not models
In a bank, client data is not a reservoir, it is an archipelago. The core banking system holds positions, the CRM holds interactions, the compliance tool holds KYC documents, and the relationship manager keeps notes in personal files.
Each system has its own version of the client, its own identifiers, its own quality rules. An AI agent asked to prepare a portfolio review therefore queries four different truths, and none of them is authoritative.
This is what we see at our clients: AI projects do not stall on a model problem. They stall the day the team discovers it takes six weeks to assemble a clean, authorised dataset.
The McKinsey report confirms this in its own way. Across seven agentic systems in production, agents cut task time by roughly 60%. Yet the bank captures only part of that, because the gains scatter across dozens of processes and silos.
The bottleneck is not the algorithm. It is reliable, permissioned and traceable access to data.
The three layers of institutional intelligence
McKinsey proposes a term we find useful: institutional intelligence. The idea is to connect the data, policies, past decisions and expert judgment currently scattered across the bank, so that every agent and every employee works from what the whole bank knows.
The report describes three layers:
data: reliable data products such as customer 360 or credit 360, governed as products rather than one-off extracts;
context: a map of how a client's accounts, obligations and regulatory status connect, with permissions enforced before anything reaches a model;
encoded judgment: the decision rules of experts, captured in a format an agent can reuse.
The second layer is the decisive one in our view. McKinsey stresses that the identity and access frameworks built for employees do not yet cover software acting on their behalf. An agent can read a client's accounts, but needs sign-off before it moves money.
For a bank supervised by FINMA, this context layer also answers the expectations of Guidance 08/2024 on artificial intelligence: knowing which data fed which decision, under which access right, and being able to demonstrate it.
The example of a Geneva private bank
Consider a situation we meet regularly, anonymised. A Geneva private bank with 400 employees and 30 billion francs under management wants to deploy an assistant that prepares client meetings for its relationship managers.
The model is chosen in three weeks. The project then stalls for five months. The reason fits in one sentence from the head of data: for a given client, we cannot say which information a relationship manager is allowed to see.
Client data lives in the core banking system, the CRM, the document repository and a data warehouse inherited from a merger. Identifiers match for only 70% of relationships.
The bank decides to reverse the order of work. Before any agent, it builds a customer 360 data product covering the 1,200 most important relationships first, with a single identity reference and a rights matrix by role.
The result after four months: preparing a meeting drops from 90 to 20 minutes, and every summary produced by the assistant cites the sources consulted, along with the access right that authorised them. Compliance signs off on that basis.
This pattern is not isolated. According to its CEO on the third-quarter 2026 earnings call, Northern Trust used an onboarding agent to cut data preparation time from two and a half weeks to three days, 83% less.
What we take away
For a Swiss or European bank or group, the McKinsey report confirms a conviction we have held for a long time: an AI strategy starts with a data strategy, and that strategy is measured in access rights as much as in volume.
Three decisions strike us as priorities before the end of the year:
appoint a business owner for each critical data product, starting with the client view;
enforce permissions upstream of the model, not in the agent's interface;
measure the time it takes to assemble a clean dataset, and make it a steering indicator on a par with token costs.
A late start costs more than it did two years ago. Nothing is lost, however, for a bank that puts its data in order now, one domain at a time.
If you want to position your organisation on these three axes, our AI Barometer gives you a first reading in a few minutes. To discuss it with us, the contact page is open.





Comments