The 18 Percent Problem in Banking AI

On 24 June 2026, Personetics published a survey of 902 banking professionals drawn from more than 30 countries. Close to 80 percent of them described fully operationalised generative AI as significant or transformational for their institution. Only 18 percent said the technology was actually running inside day-to-day operations.

That distance between belief and practice is the real state of AI in banking today. Almost everyone is convinced. Very few have shipped.

So what is holding the other 82 percent back? The same survey points at something far less interesting than model architecture. 56 percent named data silos between business lines as a major barrier, and 31 percent said their hardest challenge was keeping AI outputs accurate and compliant. Neither of those gets solved by a better model.

The Bottleneck Sits Below the Model

Consider what a single personalised offer has to touch inside a retail bank. Core banking, the card processor, the mortgage book, the CRM, the consent register and the marketing suppression list. Each of those systems was bought in a different decade by a different team working from a different definition of "customer", and none of them were designed on the assumption that a language model would one day need to read all six at once and then explain its reasoning to a regulator.

Personetics found that the average institution takes 12 weeks to move a personalised offer from concept to launch. Twelve weeks. That number has very little to do with AI capability and almost everything to do with how long it takes six systems to agree on who the customer is.

This is why the institutions doing well here look unglamorous from the outside. Teams that build Enterprise AI solutions for financial services that survive an audit tend to spend their first months on entitlements, lineage, retention rules and a single customer definition, long before anyone argues about which model to use.

The Regulatory Clock Moved, but Not by Much

Europe handed out a small reprieve this year. On 6 May 2026 the EU institutions reached a provisional political agreement on the Digital Omnibus package, confirmed by member states on 13 May, which pushes compliance for stand-alone high-risk systems under Annex III of the AI Act from 2 August 2026 out to 2 December 2027. AI embedded in regulated products moves to 2 August 2028.

Does that buy banks breathing room? A little. But those changes only take legal effect once the Omnibus is formally adopted and published in the Official Journal, and 2 August 2026 remains live for the Article 50 transparency obligations, which is where most customer-facing banking use cases sit anyway.

Truth be told, sixteen extra months is not much when the obstacle is a data estate assembled over three decades.

What the Leaders Actually Did

So what separates the institutions in production from the ones still demoing? Three habits keep showing up.

They picked one workflow and finished it. A narrow, well-instrumented use case in collections or complaint handling beats nine simultaneous pilots that all stall at legal review.

They wrote down what "wrong" looks like before launch. Accuracy is not a feeling. If you cannot describe the failure you are watching for, your monitoring will not catch it.

And they gave the work an owner with a budget. Innovation teams produce demos. Line-of-business owners produce systems that people actually use on a Monday morning.

Yes, none of this is new advice. It is roughly the same advice that applied to data warehousing in 2010 and to mobile banking in 2015, which is probably why it keeps getting skipped.

The 18 percent are not smarter than everyone else. They just started at the boring end.

About Author

Shefali Vasave

Shefali manages content marketing at Opus Technologies, a domain-native engineering partner for banks, payment providers, and fintechs, and writes on the various aspects of financial institutions navigating change in a real-time, digital-first world.