Why banks are stuck at AI pilots and how to break out

Most banks have run several AI pilots by now, across fraud detection, underwriting, customer service, and more. There has been a lot of talk at industry conferences and in boardrooms about how AI will drive efficiency in these areas, and the pressure to avoid falling behind peers makes standing still feel like a risk in itself.

But most of those pilots have followed the same pattern where the use case works in the test environment, is presented to leadership, and then stalls. That happens for several reasons, such as a governance review taking longer than expected, enterprise data being siloed, and a new priority emerging that demands resources. Six months later, the pilot is technically still active with a handful of users but frozen for wider rollout, like hundreds of other AI pilots across the bank.

Here, we will walk through why banks face this problem with their AI programmes and the steps they can take to move past AI pilots in ways that drive business growth.

The pilot paradox

The more AI pilots a bank runs, the harder it becomes to get any of them into production. Each pilot team independently builds out its own data infrastructure and compliance review process, and every new initiative adds complexity without solving the problems that stalled the previous ones.

The most common problem is data. Customer and transaction information is often scattered across 40 or more disconnected systems at a bank, so AI cannot access a single source of truth to generate outputs. For example, a fraud detection model that cannot see the full picture of a customer's activity will likely produce false positives because it does not fully understand the customer's typical financial interactions.

Compounding this, compliance and explainability are often only raised after the AI pilot is complete. Regulators require AI-driven decisions to be auditable and explainable, requirements that can take months of retrofitting to add once the AI system is established.

Beyond the structural problems, there is also the fear that poor AI outputs will negatively affect customers. An incorrect credit decision or a miscategorised transaction can undermine trust in ways that extend well beyond the AI programme itself. And that’s just the small-scale version. A rollout that poorly affects a large subset of customers can lead to reputational damage that takes years to rebuild.

Start at the foundation

The question most banks are asking is which AI use case to build first. But the first question they should be asking is whether the data layer underneath is built to support AI at all. Banks that skip to use cases without establishing a governed, unified data foundation are setting themselves up for failure down the road.

For most banks, building that foundation means rethinking how their underlying systems connect. Core banking modernisation sounds like a daunting undertaking, and for good reason. Banks that have tried to overhaul all their core processes at once have learned that the associated risk is high, which is exactly the approach banks should avoid. The most successful banks today are taking a composable approach to modernisation. In other words, they are picking one capability that needs an upgrade and modernising it in a way that connects cleanly to existing systems. If customers are asking for instant payments, that is where the work starts. Then, once the first capability is modernised, they begin thinking about the next one.

Using that approach, each upgrade strengthens the underlying data foundation, since every functionality at the bank shares the same central data storage. Data cannot be considered after the fact; it is a key element of any modernisation. Then the AI that follows draws from that foundation rather than having to reconstruct it from scratch. The AI model, working from clean, unified, governed data, makes better decisions than one reasoning from fragments. When a compliance team can see exactly how a model reached its conclusion from day one, the organisational confidence to expand what the AI is trusted to do follows naturally. And let’s not forget the customer, as that transparency extends to them too: when asked why a decision was made, the bank can explain it clearly rather than pointing to a computer decision.

Build the first use case

The institutions that succeed in selecting their first use case tend to start narrow, selecting a single application in which the AI output connects directly to a decision someone already has to make. For example, a bank might want to deploy AI to flag suspicious transactions in real time, routed directly to the fraud team with the context they need to act. The reason this is a good first use case is that the output connects to a decision that was already being made, just faster and with better information. In this case, the use case is determined by how the organisation actually operates rather than by what the model can do.

Once the first use case is right and working, confidence in AI outputs grows, making the next use case faster and less contested. Start with something too broad or too disconnected from how the bank actually operates and the pilot stalls, the budget gets questioned, and the foundation work gets blamed.

Rolling out that first use case well means starting with a subset of the organisation rather than a company-wide launch. A smaller group allows the team to validate that the model produces reliable outputs before expanding access, and it gives the people using it time to build familiarity with how to act on its outputs. Human review plays a practical role here, enabling the bank to build the evidence base it needs to expand its scope with confidence. Once the outputs are proven in that initial group and the surrounding workflows are working, widening the rollout is straightforward because the process already exists.

Banks that have built this way are making credit decisions in real time, a process that previously required a human review cycle measured in days. They are catching fraud patterns before transactions clear rather than after. Those are capabilities that simply did not exist before the foundation was in place.

What it takes to scale

Once the first use case is in production, the path to the next one is materially shorter. The data infrastructure already exists, and the compliance review process is already in place. A bank that starts with AI-assisted fraud flagging, for example, has already built the unified transaction data layer and the audit trail that a credit risk model needs next. The credit risk work then lays the groundwork for real-time portfolio monitoring. Each use case extends the same foundation rather than starting a new one, and that is where the economics of AI in banking change.

The banks that are scaling AI today stopped treating each use case as a standalone project and built something underneath that could support more than one. That shift, from isolated pilots to a shared foundation, is what separates the institutions catching fraud before transactions clear from the ones still debating which pilot to run next.

The starting point is smaller than most banks expect. One use case, chosen because it connects directly to a decision someone already has to make, built on data clean enough to audit from day one. Get that right, and the second use case is not another pilot. It is an extension of something that already works. 

About Andrew Steadman

 

Andrew Steadman is Chief Product Officer at SBS, the global financial technology company serving more than 1,500 financial institutions across 80 countries. 

 

 

About SBS

 

SBS is a global software company helping banks and the financial services industry reimagine how to operate in an AI-driven world. Trusted partner to more than 1,500 financial institutions and large-scale lenders across 80 countries, including Santander, Société Générale, BNP Paribas, Groupe BPCE, Crédit Agricole, La Banque Postale, HSBC, Attijariwafa Bank, Nationwide, NextGear Capital, Mercedes-Benz, and Toyota FS, SBS combines 50+ years of banking domain expertise with a suite of award-winning solutions built for the AI era. SBS's composable architecture extends across core banking, lending, payments, compliance, open banking, and asset finance.