Only 13% of Financial Firms Are On Track With AI Initiatives as Scaling Stalls

Three-quarters of financial firms report positive return on investment from their AI pilots — but only 13% are on track to scale those initiatives into production, according to a BearingPoint study published October 1. The gap between successful proof-of-concept and operational deployment has become the defining AI challenge in financial services, and the numbers suggest it is widening rather than closing.

The 13% Problem

The BearingPoint study surveyed financial services firms across banking, insurance, and asset management on their AI program status. The finding that captures the sector's dysfunction is not the 13% on-track figure in isolation — it is the combination: near-universal reports of positive pilot ROI alongside near-universal failure to convert those pilots into scaled systems.

Something is systematically blocking the path from "this works in our sandbox" to "this runs in our operations." According to Global Banking & Finance Review's coverage of the study, the two dominant blockers are regulatory hurdles and legacy IT integration barriers.

What "Pilot Purgatory" Looks Like

Analysts have started using the phrase pilot purgatory to describe the state where AI projects are too successful to kill and too complicated to ship. In financial services, the pattern is familiar:

• A business unit builds an AI tool that demonstrably reduces manual processing time or improves risk flagging accuracy
• The pilot clears internal evaluation
• The project then stalls at integration: connecting the AI tool to core banking systems (CBS), loan origination platforms, or claims management software built in the 1980s and 1990s
• Compliance review adds months as legal and risk teams evaluate regulatory obligations under frameworks not designed with AI in mind
• The pilot is quietly extended, re-scoped, or deprioritized while business-as-usual takes precedence

The result is a portfolio of AI initiatives that look healthy on internal dashboards — ROI proven, team enthusiasm maintained — but generate no operational impact.

The Regulatory Layer

Financial services is among the most heavily regulated sectors trying to deploy AI. Banks operating across multiple jurisdictions face overlapping AI governance requirements: the EU AI Act's financial services provisions, US banking regulators' guidance on model risk management (SR 11-7, the Federal Reserve's model risk management guidance), and emerging state-level rules. Each adds documentation, validation, and audit trail requirements that can double the time-to-production for any AI system touching credit decisions, fraud detection, or customer advice.

The EU AI Act classifies several financial AI applications — credit scoring, insurance pricing, fraud detection — as high-risk systems subject to mandatory conformity assessments before deployment. For firms with EU operations, this is not optional paperwork; it is a legal prerequisite.

The mismatch between regulatory frameworks designed for static, explainable models and the dynamic, probabilistic nature of modern AI is the technical root of the compliance friction. Regulators built oversight frameworks around models that could be fully documented and tested against defined input-output relationships. Large language models and ensemble systems do not fit that documentation paradigm cleanly.

The IT Integration Problem

Legacy IT is the second blocker, and in many institutions it is the harder one. The core systems running most banks — payments processing, loan books, general ledgers — were built on COBOL, mainframe architecture, or early client-server stacks. They are stable, audited, and deeply integrated into daily operations. They are also architecturally incompatible with the API-first, cloud-native designs that modern AI tools assume.

Connecting an AI model to a core banking system typically requires middleware layers, data extraction pipelines, and careful management of the operational/technology (OT/IT) divide — a gap between systems designed for real-time transaction processing and systems designed for analytical computation. Building and validating that bridge is expensive, slow, and requires specialized talent that most firms do not have in-house.

The Market Pressure

The study arrives as bank stocks face AI-related selling pressure. Some investors have been pricing in disruption ahead of the incumbents' ability to adapt — a pattern that emerged in the fintech era and is repeating with AI. The irony is that the firms most exposed to AI disruption are also the ones most structurally blocked from deploying AI quickly.

Pure-play fintech competitors and neobanks, operating on cloud-native stacks without legacy IT debt, face none of the integration barriers that stall incumbents. The 13% scaling rate among traditional firms may reflect a structural disadvantage, not just an execution problem.

What to Watch

The firms that break out of pilot purgatory will likely do so through one of two paths: isolated greenfield deployments on new product lines (no legacy integration required) or aggressive legacy modernization programs that treat AI-readiness as the forcing function for core IT upgrades. Watch for announcements in Q1 2027 from major banks on either track — those will indicate whether this year's pilot portfolio translates into next year's operational AI.

The regulatory environment is also shifting. US banking regulators are updating model risk management guidance to address AI specifically; the shape of that update will either accelerate or further complicate the path from pilot to production for the 87% of firms currently not on track.