60% of AI projects will be abandoned. Here's what decides the rest

The pilot that never scales

The pattern is familiar. A team builds a promising proof of concept. Leadership gets excited. Then deployment begins, and the whole thing grinds to a halt.

The reasons are almost always the same. The data the model needs is scattered across core, CRM, and mobile app databases that don't sync. There's no real-time data flow, just batch jobs running nightly or weekly. The model's output sits on a dashboard instead of plugging into the workflow where decisions actually happen. And nobody can explain to a regulator how the model reached its conclusion because the data lineage is unclear.

In the same Gartner survey, 63% of organizations said they either don't have or aren't sure they have the right data management practices for AI. More than half admitted their data simply isn't ready.

McKinsey's 2025 State of AI survey reinforces the point: 88% of organizations now use AI in at least one business function, but only about a third have begun to scale it enterprise-wide. The majority remain stuck in experimentation or piloting. Usage is up. Value at scale remains hard to find.

The constraint is almost never the algorithm. It's everything around it.

Data Maturity sets the ceiling

AI is only as useful as the data it can reach. And for most financial institutions, that data is fragmented, inconsistent, and slow.

Think about what a credit decisioning model actually needs to function. A unified view of the customer across deposits, lending, CRM, and digital channels. Consistent definitions of basic terms (what counts as an "active account" differs across departments at most banks). Data that arrives in time to matter, not last night's batch extract.

When these conditions don't exist, you can have the best model in the world and still get poor results. Or biased results. Or results nobody trusts.

Here's what data maturity looks like in practice:

• A unified data layer exists. Analytics, reporting, and AI are drawing from a common platform, whether that's a lakehouse, a customer data platform, or a governed data fabric. Without this, every team works from its own extract, producing its own version of the truth.
• Data quality is actively managed. Inconsistent definitions are one of the most common and most underestimated problems in banking. Master data management, deduplication, and shared definitions aren't exciting work. But they're what separate models that perform from models that mislead.
• Data is accessible without weeks of bureaucracy. If data scientists have to file tickets and wait for dataset access, your AI development cycle stretches into quarters. Leading institutions are building governed self-service catalogs where AI teams can find and pull approved data without bottlenecks.
• Real-time pipelines are in place. Fraud detection, personalization, and credit monitoring all need streaming data. Event-driven architectures make that possible. Without them, you're limited to looking backward. That's useful, but it's not where the competitive advantage sits.

This tracks directly with the framework in Zennify's Change Appetite Matrix: data agility and AI agility rise together. You don't get to skip a step.

Integration Maturity is the multiplier

Data is one constraint. Integration is the other, and it's the one that tends to get less attention.

AI in banking doesn't run in a vacuum. A credit scoring model needs to talk to the loan origination system, the online application portal, and the core. A personalization engine needs to pull from CRM, transaction history, and channel interaction data at the same time. A fraud model needs to push alerts into case management in real time.

If each of those connections requires a custom project, deploying a single model across the enterprise becomes a multi-quarter, multi-million dollar effort. If you have a modern integration layer with standardized APIs and middleware, deployment gets much closer to plug-and-play.

The economics tell the story. McKinsey reports that banks spend roughly $650 billion a year on IT globally, more than any other industry as a percentage of revenue. And yet a ThoughtWorks analysis found that nearly two-thirds of retail banks' IT budget goes to maintaining existing systems. That leaves very little room for the modern integration work AI demands.

Institutions carrying heavy technical debt often have the best AI ideas and the worst ability to execute them.

What does integration maturity actually look like?

• Systems are accessible through standardized APIs, not file transfers or direct database queries. When most of your systems expose clean interfaces, connecting new AI services becomes straightforward.
• Architecture is modular. New components, including AI engines, can be introduced without rewriting the systems around them. If your core banking platform is a monolith, inserting an AI step into a process might require a vendor engagement just to get started.
• Workflows can incorporate AI as a step in the process, not just a standalone alert. When a model flags a risky transaction, can your systems automatically route it to case management or a human reviewer? Or does someone have to notice it on a dashboard?
• Infrastructure handles the volume. AI workloads that scan millions of transactions for anomalies need low-latency, high-throughput pipelines. If your current integrations choke under load, certain AI use cases are off the table.

What this means for your AI investment strategy

Aligning AI ambitions with data and integration reality isn't about slowing down. It's about spending money in the right order.

Assess before you build

Use a structured maturity framework to score your data, integration, and AI capabilities. If your data is still siloed and your integrations are brittle, a major AI rollout will stall regardless of the vendor you choose. Investing in the foundation isn't a delay. It's the fastest path to production-grade AI.

Treat data modernization as a strategic program

Appoint clear ownership. Fund it explicitly. Consolidate customer records, implement governance, move to cloud data platforms. These efforts compound. They improve reporting, regulatory compliance, and customer experience alongside AI readiness.

Build reusable integration patterns

Standard APIs for common banking functions (customer lookup, transaction posting, account status) mean every new AI module plugs into existing infrastructure instead of requiring custom work. This is where integration platform and API management investments pay off fastest.

Pilot end-to-end, not in isolation 

When testing an AI model, test the full chain. Does the data flow correctly in production? Does the model connect to the target workflow? Does the output trigger the right downstream action? Measuring model accuracy in a sandbox tells you almost nothing about whether it will work in the real world.

Bridge the gap between engineering and AI teams 

Cross-functional collaboration catches infrastructure problems before they become deployment blockers. One bank created a combined team spanning data engineering, integration architecture, compliance, and business units alongside its data scientists. The result was a smoother path from prototype to production for an AI-powered underwriting system, because the integration lead identified early that near-real-time data feeds required a change data capture pipeline.

The compound return

When the foundation is right, the payoff extends beyond any single AI use case. You build a pipeline for continuous AI deployment, not a collection of one-off experiments.

And here's what makes the investment case even stronger: the data and integration work that enables AI also improves everything else. Better reporting, faster product launches, smoother digital experiences, stronger regulatory posture. You're building infrastructure that supports every strategic priority on the roadmap.

McKinsey's 2025 Global Banking Review projects that AI adoption could drive up to 20% in net cost reductions across the industry. But that value only flows to institutions that can actually operationalize AI, which means institutions that have already done the data and integration work.

The institutions that scale AI successfully are the ones that invested, often unglamorously, in the capabilities underneath it.