Adoption Was the Easy Half

Financial services has stopped asking whether to use AI. The harder question — how to industrialise it — has a binding constraint, and it isn't the model. It's the operating model.

NVIDIA's sixth annual State of AI in Financial Services survey, its largest yet at 839 respondents, reads on the surface as a victory lap for AI adoption. Active use has climbed to 65% from 45% a year earlier, while the share of firms merely "assessing" pilots has halved to 24%. Nearly every respondent plans to hold or grow AI spending in 2026, with 44% raising budgets by more than 10%. If arrival were the question, the survey would be an answer.

But arrival is settled. The more useful way to read this report is as an early map of who will convert AI into durable advantage — and on that score, the most revealing number is where the money is now going. The top spending priority has flipped to "optimising AI workflows and production cycles" (41%, up from 26% the year before). That is the vocabulary of an industry moving out of the lab and onto the factory floor. Three shifts follow from it.

The contest is industrialisation, not adoption

Adoption is becoming table stakes; the real competition is converting pilots into production at scale, repeatably. The evidence sits in the friction points. The leading obstacles to agentic AI are overwhelmingly operational — performance and reliability (34%), a shortage of internal skills to manage and monitor agents (33%), data issues (30%), and integration with existing workflows (28%) — not gaps in the underlying technology. And of the 42% of firms "using or assessing" AI agents, only about half have actually deployed them (21% of the total sample).

The pattern is a sector rich in experiments and poor in production lines. The firms that win will be the ones that build a pipeline to move use cases from pilot to live, governed operation. The laggards will keep curating a portfolio of demos that never compound into anything.

The constraints are organisational, not computational

The two largest barriers to scaling AI are data-related issues (40%) and the talent shortage (35%). The most instructive trend underneath those: having enough data to train models has nearly vanished as a concern, collapsing from 49% in 2023 to 16% this year. The problem was never volume. It is now governance, privacy, data sovereignty, and fragmentation across disparate systems — the question of whether an institution's data is coherent enough to be trusted in production.

This is the through-line from our earlier analyses. AI amplifies whatever operating logic an institution already runs on. Firms whose data and organisation are already coherent will compound their lead; disorganised ones will simply automate their disorder faster. The bottleneck to AI value is not compute. It is organisational order.

The architecture choices reveal a control agenda

Two trends look technical but are strategic. Hybrid infrastructure has jumped to 47% from 26%, with cloud-only deployment falling to 42% from 57%. And open-source models are now rated important by 84% of respondents, with 48% of those in management calling them very important.

Neither is really about plumbing. Both are about ownership — controlling cost-per-token as usage scales, keeping proprietary data in-house, and not exporting enterprise value to a third party for the use cases that matter most. Institutions are deliberately repatriating control over the core of their AI stack while renting commodity workloads from managed providers. That build-versus-buy boundary is hardening into a genuine strategic fault line, and it deserves to be drawn on purpose rather than by default.

Where the value is actually landing — and a caveat

The return on investment clusters in unglamorous places: document processing (named by 32% as a top-ROI use case), customer experience and engagement (30%), and document management (23%). The flashier applications — algorithmic trading, portfolio optimisation — rank lower and are confined to specific segments. The lesson is to chase the boring, compounding back-office wins first, because that is where evidenced return actually lives.

A caveat a careful reader should price in: this is a vendor-sponsored survey, drawn largely from NVIDIA's own distribution lists, so the sample skews AI-committed. Treat the most triumphant figures with that in mind. The headline that 89% of respondents saw AI both lift revenue and cut costs sits uneasily next to independent work from the same period — the Cambridge Centre for Alternative Finance found only around 40% reporting any profit boost and 43% reporting no change, with most large institutions unable to measure the value at all. The truth lives in the gap between vendor optimism and measurement difficulty.

What BridgeUp is doing about it

The work follows directly from the data, in order of leverage. Industrialise two or three high-ROI back-office use cases into a real production pipeline instead of adding more pilots. Fix data governance before scaling further, because it is the genuine bottleneck. Make a deliberate stack-ownership decision — own the core, rent the commodity — rather than drifting into one. And stand up the measurement discipline that lets an institution prove what its AI is actually worth, the capability most of the market still conspicuously lacks.

Adoption was the easy half. The advantage is in the industrialisation, and it is won on the operating model — which is exactly the terrain BridgeUp works.

Source: NVIDIA, "State of AI in Financial Services: 2026 Trends" (839 respondents, fielded August–September 2025). Comparative figures from the Cambridge Centre for Alternative Finance, "2026 Global AI in Financial Services Report."