23% of financial institutions already run agentic AI, Backbase report finds

Twenty-three percent of financial institutions already run agentic AI in production, according to Backbase's Banking Predictions 2027 report, which argues that banks must now rebuild their operating models around governed AI agents rather than continue running isolated pilots.

Backbase, an Amsterdam-based banking technology company, published the report on 13 November 2026. It is the firm's third annual predictions report and outlines 10 trends for the year ahead. Contributors come from commercial, retail and capital markets banking, as well as venture capital and fintech.

The report's central claim is stark: a widening divide is opening between institutions that embed AI in day-to-day operations and those that remain stuck in limited pilot projects. The stakes are practical, not theoretical. Lenders face margin pressure and tighter regulatory oversight at the same time, and in Backbase's view AI deployment has moved beyond experimentation to questions of governance, accountability and organisational design.

Why does the report say the pilot phase is ending?

Backbase frames the end of the pilot era as the first of its main industry predictions. With 23% of financial institutions already running agentic AI in production, the company expects leading banks to adopt centralised orchestration platforms. These platforms, in Backbase's view, should reuse tools across the business while embedding compliance from the outset, rather than bolting controls on after deployment.

The speed of the transition caught even advisors by surprise.

"The shift to agentic AI is happening much faster than banks think.... Models will continue to improve, but the work organizations need to do to integrate them into their operations and prepare for new distribution channels cannot wait," said Panagiotis Kriaris, Senior Advisor covering fintech, payments and banking.

The ellipsis in the original quote marks a pause, not a gap in meaning: Kriaris's point is that model improvement is the part banks can wait for, and operational integration is not.

Where will AI budgets go, and who must justify them?

A second theme is spending discipline, and here the report's numbers point in opposite directions. On one side, 36% of the world's major banks expect to more than double their AI budgets. On the other, only 30% of the use cases announced by the world's 50 largest banks included impact metrics.

That gap between rising budgets and thin measurement will put pressure on technology teams and suppliers, the report argues, forcing them to show a direct return on investment for each use case. Banks will need to link AI spending to baseline costs, running costs and measurable outcomes if they want to defend their budgets.

Backbase Chief Technology Officer for Digital Banking Greg Fahy put the test in investment-committee terms:

"In 2027, every AI line item in banking will have to justify itself like any other investment, with a baseline, a running cost, and a return. Leaders that can't show all three will see their budgets move to teams that can," said Fahy.

The implication for the supply side is direct. Vendors and internal platform teams that cannot produce impact metrics will compete for a shrinking share of a growing pool of money.

What does governance of AI agents require in practice?

The report identifies governance as a major operational issue, and it treats the shift from single assistants to multiple agents across workflows as the trigger. Written policies alone will not satisfy regulators or internal risk functions once agents act across many processes, Backbase argues.

Instead, institutions will need systems that do three things:

• Assign verified identities to individual agents
• Log agent actions in real time
• Allow the bank to disable an agent immediately if needed

The report situates this requirement in a broader shift in banking regulation towards traceability and auditability in automated decision-making. A bank that cannot reconstruct what a given agent did, when it did it, and under whose authority will struggle to pass that test.

Fatbardha Rino, Founder and Chief Executive Officer of Jet Bank, framed responsibility as remaining with the bank regardless of scale:

"2027 will be a big change when it comes to AI: how responsibly we use it and to what extent.... As we scale, accountability must remain clear: we are responsible for every decision and every customer outcome. Trust, security and regulatory discipline have to be built in from the start and by the bank," said Rino.

Can banks rent models but still own outcomes?

Backbase Founder and Chief Executive Officer Jouk Pleiter argued that the arrival of AI agents changes core management assumptions rather than simply adding another layer of software. Banks may be able to buy or rent access to models, he said, but they still retain responsibility for outcomes.

"Banks can rent AI models, but they have to own the accountability for what those agents do. That changes the operating model itself, as agents become colleagues and people stay accountable for the decisions that matter," said Pleiter.

The phrase "agents become colleagues" is the report's most compact statement of its operating-model thesis: the technology is treated as a participant in workflows, not a tool purchased and shelved.

How must the operating model itself change?

Another of the report's five main industry predictions is that banks will need to redesign roles, structures and processes as AI agents take on more preparatory work. The report expects the changes to begin in operational areas where processes are fragmented and manual intervention remains common.

Backbase describes a model in which agents handle tasks in the back-office "messy middle", while people continue to make final decisions. In that setup, productivity gains come less from automating individual tasks and more from changing how the organisation is structured.

That distinction matters for anyone measuring the return on agentic AI. A bank that automates tasks within an unchanged structure will capture a fraction of the gain available to a bank that restructures around the division of labour between agents and people.

Rosanna Fornarino, Chief Operating Officer of Judobank, located the difficulty precisely in that organisational work:

"Launching pilot initiatives and testing AI use cases is relatively straightforward. The greater challenge lies in transforming the organization itself by building the right culture, fostering the right behaviours, and creating the conditions that enable AI to scale sustainably and deliver meaningful business value," said Fornarino.

Are banks about to serve AI agents as customers?

The final major theme is the shift from app-based banking to conversational interfaces, and the report anchors it in a consumer number: 14% of consumers globally already allow AI to choose financial providers on their behalf.

That figure suggests banks may increasingly have to serve software agents as well as human customers. The consequence could alter how banks attract and retain customer relationships, especially if third-party assistants begin to sit between consumers and financial institutions. A bank that is invisible to a consumer's chosen assistant is, in effect, invisible to the consumer.

The report argues that banks able to respond, advise and act through consent-based conversational systems may be better placed to retain those relationships. Consent-based is the operative term: the conversational channel is framed as operating under explicit customer permission, not as open access to accounts.

Eugene Deeny, Lead Industry Advisor for Banking and Capital Markets at Endava, said the customer at the door may no longer be a person:

"There's already a high degree of comfort across generations with interacting through conversational AI. The bank has to be ready for the person at its front door to be an agent," said Deeny.

What should banks take from the 10 trends?

Taken together, the report's five main predictions and 10 trends reduce to one operating instruction: treat agentic AI as an organisational redesign programme with governance and measurement built in, not as a software procurement. The dividing lines it draws — production versus pilots, measured versus unmeasured spending, governed versus ungoverned agents, conversational versus app-based distribution — all reward institutions that moved early on structure.

With 23% of institutions already in production and 36% of major banks preparing to more than double AI budgets, the report's closing logic is that the cost of waiting is no longer technological but competitive: the operating-model work, in Kriaris's words, "cannot wait".