AI in Financial Services 2026: Adoption Is Near-Universal. Profitability Isn’t

For most of the last decade, writing about digital transformation in finance meant making a case for it. Mobile banking was still displacing branches, fintech was still the disruptor, and the argument was about whether incumbents would move fast enough.

That argument is settled. Adoption happened. What did not happen, for a substantial share of the industry, is the return.

This article looks at where digital transformation in financial services actually stands, using the most substantial evidence base currently available, and at what separates the institutions seeing results from the ones still waiting.

628 organisations, 151 jurisdictions, one clear pattern

The Cambridge Centre for Alternative Finance published its 2026 Global AI in Financial Services Report in April, produced with the Bank for International Settlements, the IMF, the World Economic Forum, the World Bank Group and others. It surveyed 628 organisations — 203 fintechs, 149 traditional financial institutions, 146 AI vendors and 130 central banks and regulators — across 151 jurisdictions.

Four in five financial services firms are deploying AI at some level. But the maturity distribution underneath that headline is where the useful information sits.

Forty-one percent are still at pilot stage. Fourteen percent describe their deployment as transformational. The report’s own framing is an industry “in mid-transition, marked by a significant execution gap between widespread early-stage adoption and true strategic transformation.”

Adoption figures are easy to celebrate. Outcome figures are harder, which is why they appear less often.

The report notes that profitability outcomes correlate with AI investment levels and workforce preparedness — not with which models or vendors were selected. That is a useful thing to know before signing the next platform contract.

The Cambridge report identifies persistent constraints in data quality and legacy systems as the structural divide separating institutions that transform from institutions that pilot indefinitely. Deloitte’s banking and capital markets research puts numbers on how bad that constraint is.

A model cannot outperform the data it is given. When nine out of ten data users inside a bank cannot reliably get at the information they need, the constraint on AI is not the AI.

This is the same pattern that shows up across every sector deploying AI at scale, and it is why data readiness deserves assessment before model selection rather than after. Financial services has a harder version of the problem than most, because the systems holding the data are frequently decades old and were never designed to share it.

The most quoted statistic in this category comes from the McKinsey Global Institute: generative AI could add between $200 billion and $340 billion in annual value across the global banking sector, equivalent to roughly 9–15% of operating profits, largely through productivity gains.

McKinsey’s narrower near-term estimate is more useful for planning: up to 20% in net cost reductions as adoption scales. Their banking review also finds that first movers stand to gain roughly a 4% return-on-tangible-equity advantage, while slower movers end up with an uncompetitive cost base. The distribution matters more than the total.

The use cases that survive contact with production

A consistent pattern runs through the research: the use cases that reach production are the ones where the value is measurable and the failure mode is contained.

The common thread is that each has a measurable baseline. Institutions that cannot say what a process costs today will struggle to demonstrate that AI improved it — which is the measurement discipline covered in our guide to measuring ROI from generative AI projects.

The Cambridge research finds fintechs ahead of traditional financial institutions on every maturity metric, and the agentic AI figures show it most clearly.

Two findings from the same report complicate the usual explanation that incumbents lack budget. First, 53% of respondents spend under $100,000 a year on AI and still report high maturity in generative and agentic AI. Second, firms in emerging markets and developing economies report higher deployment levels than those in advanced economies.

Neither pattern is consistent with spending being the constraint. Both are consistent with legacy systems and organisational readiness being the constraint — which is also why legacy modernization keeps appearing on the critical path of AI programmes that were not supposed to be modernization programmes.

One of the report’s more uncomfortable findings is a perception gap between the firms buying AI, the firms selling it, and the regulators supervising it.

On both measures, the vendors selling AI into financial services rate the risk materially lower than the institutions running it and the regulators supervising it. That is worth factoring into vendor conversations.

Bryan Zhang, who leads the Cambridge Centre for Alternative Finance, characterised the sector as one where adoption is remarkable but “accountability for AI failures is unresolved.” For regulated institutions, that is not an abstract concern — it is a question of who answers when an autonomous system gets a decision wrong.

Four things follow from the evidence above.

None of this is unique to finance. What is unique to finance is the regulatory exposure when it goes wrong, and the age of the systems holding the data — which is why the sector shows a wider spread between leaders and laggards than most.

Digital transformation in financial services is no longer a question of intent. Four in five institutions are deploying AI, and the technology has demonstrably worked in fraud, compliance and onboarding.

The open question is why that has translated into profitability for two in five and nothing measurable for a similar number. The evidence points consistently at the same places: data that cannot be retrieved, systems that will not integrate, and workforces that were not prepared for what arrived.

None of those are AI problems. They are the problems AI made visible.