If a bank has decades of customer data, a strong technology budget and a leadership team keen on AI, you'd expect productised AI to arrive quickly. New research from Oxford shows the opposite.
In 2025, the Future of Finance and Technology Initiative at Oxford's Saïd Business School published a report on how established financial institutions are moving to AI. Written by Wendy Kilminster and Pınar Özcan, the study is based on interviews with banks, fintech companies and industry advisers. A quote at the start of the report sums it up well: the finance director of a European bank says the technology and the data are great, everyone is very interested, but they don't know where to start.
In this post I put Oxford's findings side by side with the 2025 results of the Evident AI Index, which ranks banks by AI maturity.
Who leads? Evident AI Index 2025
Every year Evident scores 50 large banks from North America, Europe and Asia using public data. Almost half of the score comes from talent; in other words, a bank's AI maturity is measured mostly by the people working in this area.
Three things stand out in the table:
- The top two banks are in a league of their own. The gap between JPMorgan and Capital One is under 1 point; RBC in third place is 20 points behind. The leaders were also the fastest movers in 2025: one added 9 points, the other 9.6.
- Everyone is improving, and the leaders are improving faster. 47 of the 50 banks raised their scores, but the top 10 grew twice as fast as the rest. So the gap isn't closing; it's widening.
- Six of the top 10 are in the US. Europe's best are UBS (7th) and HSBC (8th).
This ranking answers the question "who is preparing more", not "who is gaining more". For the answer to the second, we need to look at Oxford's interviews.
Three preconditions for readiness
In Oxford's framework, everything starts with the characteristics an organisation has before it begins with AI. The report calls these preconditions, and the preconditions shape the organisation's AI foundations: its strategy and organisational infrastructure.
The picture painted by one of the interviewed institutions is striking: AI spending is a tiny part of an enormous technology budget. They could perhaps spend ten times as much; but there are other things they need to build first. Until cloud migration, data integration, data governance and APIs are done, AI can only exist as small islands of experimentation.
The real cost of a use case
One of the most instructive lines in the report is what an executive says about a use-case budget: 60–70% of the spending has nothing to do with AI; it goes to core middleware integration.
This is something developers know intuitively but that is often forgotten in investment decisions. The model is the most visible but smallest part of a system. The real work is getting data there reliably, connecting the result to existing processes, authorisation, logging and monitoring. That's why Oxford recommends that organisations build reusable building blocks instead of building separately for each scenario: a shared technology stack, shared security rules, and organisation-wide standards that decide which decisions can be automated and which stay with people.
There's another pattern the report keeps coming back to. First, business units experiment freely; then someone notices that company data has been uploaded to a public chat tool (ChatGPT, for example). What follows is always the same: the tool is blocked, an organisation-wide usage policy is published, and a central body is set up to approve use cases.
Why the efficiency maths doesn't add up
Most AI business cases at banks rest on efficiency: reducing the human time needed to do a job. An experience described in the report explains why this stays on paper. When 5% to 15% of every operations employee's work is automated, a spreadsheet at the end of the year shows savings of a million or two. But in real life those people can't be taken out of the system, so most business cases don't deliver what they promise.
5–15% of every employee's work is automated, and total savings reach millions by the end of the year.
The small slivers of time saved show up in neither role design nor headcount planning. An organisation that doesn't fill the freed-up time with valuable work never sees the savings.
According to the report, value can only be realised in one of two ways: a role redesign that replaces the work AI takes over with more valuable work, or a reduction in headcount. Where the second isn't discussed, employees stay reluctant to adopt the tools. That's why Oxford recommends that leaders openly discuss the impact on roles when they set efficiency targets.
The caution paradox
This is the report's most striking concept. Financial institutions need to be cautious: regulatory and reputational risks are real, and data security and resilience are a must. So the typical approach is to start with small steps: use cases that face the organisation rather than the customer, that cut costs rather than drive growth, and that improve a single step of an existing process.
The trouble is that this caution creates its own risk.
An executive's experience in the report makes this concrete: a project done wrong the first time makes every following initiative harder to adopt. Once people have been disappointed, they don't show interest a second time.
What works?
Oxford's answer isn't "drop the caution". The report recommends five practices so that what is learned from the first initiatives turns into self-reinforcing momentum, like a flywheel:
To me the fourth point is the most important. Instead of improving one step of a process with AI, redesigning a customer journey or a business task from start to finish produces more radical solutions and also makes it easier to overcome reluctance to adopt. The report also stresses that this doesn't require building the applications in-house; what matters is gaining enough depth to evaluate what technology vendors offer.
The fifth point is aimed at investment committees. Investments such as infrastructure and data quality, which don't produce profit directly but make everything else possible, always get cut in use-case-by-use-case return calculations. Evaluating the programme as a whole makes these investments visible and stops the same role from being counted as "savings" in more than one project.
My own note
The reason I find this report interesting is that the problems of large institutions aren't that different from the problems of small teams. Choosing a model is easy; the data flow, integration, authorisation and monitoring around the model are the real work. Then there's the question of measurement: saying "employees spend less time" doesn't show what that time turned into.
Financial institutions have three things their competitors can't easily copy: trust, data and customer relationships. That's the conclusion Oxford reaches too: when these three strengths are combined with properly built AI foundations, established institutions are still well placed in a changing competitive landscape. But first they need to get out of the pilots.
Sources
- W. Kilminster, P. Özcan, AI Preparedness in Established Financial Services Firms, Oxford Future of Finance and Technology Initiative, Saïd Business School, 2025.
- Evident, Evident AI Index, 2025.
- Teradata, Which banks are leading in AI? (summary of the Evident AI Index 2025).
- JPMorganChase, JPMorganChase tops the Evident AI Index for the fourth year, 2025.
- McKinsey & Company, Capturing the full value of generative AI in banking.