The era of AI pilots ends for banks in 2027

Every bank has run an AI pilot. Very few have scaled one. That gap is where the Backbase 2027 Banking Predictions Report starts, the third edition of our annual outlook, built with banking leaders and experts from around the world. Its first prediction sets the tone for the other nine: in 2027, AI has to leave the lab. Below is the full chapter, with the data, the industry view, and what to watch.

Why the pilot era is ending

The era of AI experiments in banking is over, whether banks are ready or not.

Last year’s report predicted 2026 would be the year AI moved from pilots to productivity, and the industry got only partway there. Over the past year, a few banks have scaled AI with measured value while the majority remained stuck. Only 23% of financial institutions say they’re in production for agentic AI, 34% are piloting, and 26% are still planning, mostly for internal operations, employee knowledge search, and customer chatbots.¹ Meanwhile, only 30% of incumbents have reached advanced AI adoption, meaning the “scaling” or “transforming” stage, versus 47% of fintechs.² Pilots have multiplied, but mass rollout has yet to follow.

There are three key reasons behind this:

1. Every pilot starts from scratch

Many banks still treat AI as tactical point solutions, launching isolated use cases with dedicated teams, fragmented data pipelines, and custom scaffolding. Without a shared foundation, every initiative reinvents the wheel. Talent constraints deepen the problem: 39% of banks and insurers across EMEA cite internal skills as an adoption barrier, while 49% remain crippled by legacy data silos.³ Despite an annual global tech spend exceeding $600 billion, productivity remains flat because institutions add bleeding-edge AI models directly onto brittle core systems.⁴ Leading tier-one banks have broken this cycle through centralized agentic orchestration platforms. They decouple AI capabilities into reusable enterprise services that standardize security, context retrieval, and regulatory guardrails across every customer journey.⁵

2. Automating whole journeys at once

Many banks set out to automate entire journeys at once and stall. The faster route starts with the steps where the gains are clearest. At Bank of Singapore, an AI agent embedded in the know-your-customer (KYC) process now handles source-of-wealth checks, and tasks that once took up to 10 days can now be completed in about an hour.⁶ Every banking journey has steps like these, where automation pays off quickly, and others where judgment, empathy, or regulation calls for a human. The real skill is in calibrating how much of each process to automate, proving it step by step, and then redesigning the journey around what works.

3. Compliance comes last

Too often, compliance is checked at the end of an AI project, after it’s built, which slows every launch and turns every new rule into rework. Regulatory compliance is among the main barriers to further progress with generative AI, cited by 26% of U.S. banking leaders, and only 16% of generative AI use cases reach full deployment.⁷ The rules are still changing, too. Among 27 financial authorities surveyed in emerging and developing countries, only a third have issued formal regulations or guidance on AI,⁸ and even the EU has pushed back its high-risk AI obligations to the end of 2027.⁹ When controls are designed in from the start, a bank can show supervisors how its systems behave at any moment, and absorb new rules as they arrive without rebuilding what it’s already launched.

None of this buys financial institutions more time. AI is a horizontal technology that touches every part of the bank, and it’s fast becoming table stakes. Digital transformation may have taken the industry 15 years, but AI is moving much faster, and competitors aren’t waiting. Fintechs already lead incumbents on adoption, and the banks that scaled first are compounding their advantage with every use case they put into production. The question for 2027 is which banks can make AI work at scale, and how far behind the rest will fall.

“Organizations need to understand that the shift to agentic AI is happening much faster than they think, and they don’t have the luxury of waiting. I hear this in many discussions: ‘Okay, we’ll wait and see how model A or B develops.’ No. The 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. And that’s not something they can accomplish overnight.”
‍
– Panagiotis Kriaris, Senior Advisor - Fintech, Payments, Banking

The industry view

In 2027, the AI pilot era will finally end. Banks must build the production model that agentic AI needs: the skills, processes, and controls to run it across the bank, automating the parts of each journey where it pays off. In a production model, AI capabilities are built once and reused across the bank, so every new use case starts further ahead than the last. Automation is calibrated, with agents taking the steps where they pay off and people keeping the rest. And controls are designed in from the start, so AI becomes part of how the bank runs every day.

A handful of banks are already showing what production looks like in practice. DBS runs more than 2,000 models across over 430 use cases, worth about SGD 1 billion in economic value.¹⁰ In Australia, Commonwealth Bank measured about A$200 million in AI benefits in fiscal 2026, a number it expects to double in 2027.¹¹ And in Africa, Standard Bank Group says more than 39,000 employees, 72% of its workforce, are active users of generative AI, with early use cases showing productivity gains of about 20%.¹² What all of these institutions share is a model for running AI, beyond any single use case.

For most banks, getting to that level will mean three changes in 2027, one for each blocker. Those that don’t make them risk joining the more than 40% of agentic AI projects Gartner expects to be canceled by the end of 2027.¹³

1. Build once, reuse everywhere

The banks that run central AI platforms treat every project as a building block for the next. Shared data, models, guardrails, and agent capabilities, such as document reading, identity checks, or case summaries, can be used by any team, so what one team learns carries into the next project and each new use case gets cheaper and faster to launch. Agents can even speed up the build itself: Citi’s AI-powered tools have run 1.5 million automated code reviews, generating 100,000 hours of capacity weekly.¹⁴

2. Automate the right steps first

Start where the gains are clearest, prove them, and then redesign the journey around what works. In Mexico, Banorte cut document review time on credit and payroll applications by 86.6%, and automatic approvals rose from 70% in 2022 to 91% in 2024.¹⁵ With the appetite for open-ended experimentation shrinking, each step a bank automates needs a clear purpose before it starts.

3. Design for continuous compliance

Supervisors audit systems, so compliance has to live inside them. That means controls built in from day one, so every action an agent takes is checked, logged, explainable, and reversible, and a new rule becomes an update to existing controls. Partners can help: 77% of financial institutions already work with third-party AI platforms.¹ The question is whether those platforms come with access controls, audit trails, and data protections built into every agentic workflow, so banks don’t have to add them case by case.

The shift from pilots to production is ultimately about how a bank organizes itself around AI. With capabilities built once and reused, automation aimed at the steps that matter, and controls designed in from the start, the industry will finally turn a scattered set of pilots into a lasting capability. In 2027, the banks that treat AI as a way of operating will pull away, and the rest will spend the subsequent years catching up.

“2027 is the year bank boards stop accepting AI pilots as evidence of progress. Pick a journey your customers complain about most and redesign it with agents doing the work from the start. Challenge every step, handoff, and delay. Remove the work that no longer needs to happen, and let agents execute what remains, within rules, permissions and oversight set by the bank. Agents do the work. The bank stays in control. Then measure the result: faster resolution, lower cost, and a better customer experience. If you’re still running the same process with an AI assistant alongside it, you haven’t gone far enough.”
‍
– Jouk Pleiter, Founder & CEO, Backbase

Here’s what to watch in 2027

• Boards stop funding pilots that have no path to production
• Banks report how many AI use cases are in production, with value attached
• Central AI platforms become standard at the largest banks
• Agents help build and upgrade core banking software
• New AI guidance arrives from regulators outside Europe and the U.S.

“Each new AI use case should be cheaper and faster to launch than the one before. If yours aren’t, you’re running a collection of experiments. In 2027, build the foundations every use case needs once, then track the time and cost of each new launch. Those two numbers will tell you whether you’re building a capability or just collecting pilots.”
‍
– Mayur Vichare, VP Strategy, Value Consulting & GTM Lead, Backbase

Frequently asked questions

Why do so many banks get stuck in AI pilots?

The report names three blockers. Every pilot starts from scratch, with its own team, data pipeline, and scaffolding, so nothing carries over. Many banks try to automate entire journeys at once and stall. And compliance is checked at the end of a project, which slows every launch and turns each new rule into rework.

What does a production model for AI look like?

AI capabilities are built once and reused across the bank, so every new use case starts further ahead than the last. Automation is calibrated, with agents taking the steps where they pay off and people keeping the rest. And controls are designed in from the start, so every action an agent takes is checked, logged, explainable, and reversible.

Which banks have already moved AI into production?

DBS runs more than 2,000 models across over 430 use cases, worth about SGD 1 billion in economic value. Commonwealth Bank measured about A$200 million in AI benefits in fiscal 2026. Standard Bank Group says more than 39,000 employees, 72% of its workforce, actively use generative AI.

What happens to banks that stay in pilot mode?

They risk joining the more than 40% of agentic AI projects Gartner expects to be canceled by the end of 2027. Fintechs already lead incumbents on adoption, and the banks that scaled first compound their advantage with every use case they put into production.