Compliance AI projects stall as data gaps slow bank rollouts

Only 34% of financial institutions have taken an artificial intelligence project past the pilot stage, while nearly 80% plan to use AI for financial crime compliance by 2026, according to research cited by SymphonyAI Financial Services. The gap points to a practical constraint for banks and other regulated firms: AI ambitions in anti-money laundering and compliance are running ahead of the data, workflow and governance systems needed for production use.

Jason Shane, head of strategy and innovation at SymphonyAI Financial Services, said the issue is often not model performance but the architecture around it. In one example he described, a compliance team had tested a transaction-monitoring AI system for eight months, but deployment was held back because three upstream data systems used different customer identifiers, schema mapping remained unresolved and each additional data source required weeks of manual reconciliation.

Shane identified three recurring causes of stalled compliance AI programmes: fragile integrations between systems, uncontrolled growth in AI agents across teams, and internal builds that underestimate the infrastructure required before deployment.

Data fragmentation remains the first bottleneck

Many compliance environments depend on custom links between individual systems, according to Shane. These point-to-point pipelines may function in testing, but can break or require substantial maintenance when a schema changes, volumes rise or a new data source is connected.

The FinCrime Frontier Report cited by SymphonyAI found that 11% of organisations are very confident in their data quality. It also found that 66% of compliance professionals see data quality and integration as the largest drag on anti-money laundering efficiency, ahead of model capability and regulatory complexity combined.

For compliance teams, fragmented data can undermine the core purpose of AI-based monitoring. If a transaction-monitoring model and a know-your-customer system draw on different records for the same customer, the model is not operating from a single customer view. Shane argued that a shared, entity-centred data foundation can give AI components a common understanding of customers, accounts and counterparties.

Agent growth raises governance questions

Shane also warned that separate AI agents can emerge across transaction monitoring, suspicious activity report support and sanctions screening without a common view of the same customer or counterparty. In that situation, two agents can assess the same party differently, while compliance teams may lack a clear audit trail showing which agent made a decision and on what basis.

He said firms need an orchestration layer that routes agent actions through shared domain context. In practice, that means workflow rules connect detection, investigation and reporting so that decisions can be traced back to consistent underlying data.

Governance is often added too late

Internal AI builds can also expand beyond their original budgets and timelines, Shane said, as teams end up constructing data pipelines, entity-resolution tools and model-governance controls that were not fully costed at the start.

The report cited by SymphonyAI found that 17% of institutions have fully operational AI governance frameworks, while 46% name cost and data challenges as their biggest obstacles to scaling AI. Shane said governance needs to be built into compliance AI from the start, including lineage, auditability and explainability, rather than added after deployment.

For regulated financial institutions, the operational risk is that an AI decision cannot be reconstructed during an examination. Shane said firms that move from pilot to production tend to be those that first establish common data context, coordinated workflows and governance controls before AI agents go live.

This story draws on original reporting from Finextra Research.