The early adopters have published their numbers. JPMorgan Chase generated $1.5 billion in annual value from AI in 2025. Goldman Sachs put 12,000 developers on GitHub Copilot. Citizens Bank reported 20 percent productivity gains across AI-enabled workflows. The question is no longer whether AI delivers — it is why only 23 percent of institutions have made it to production.
KPMG's 2026 technology survey found that AI adoption among financial services firms doubled from 30 to 75 percent between 2024 and 2026, yet the majority of institutions remain stuck in pilot. The barrier is not technology maturity. It is governance infrastructure — and that gap is closeable.
Where the ROI Is Concentrated
The evidence from early adopters clusters in five deployment categories, each with documented and replicable returns.
Developer productivity is the clearest ROI category in the industry. Goldman Sachs's GitHub Copilot rollout produced measurable gains in code review, test generation, and documentation. Naranja X reported developers saving two to three hours per day on routine coding tasks. Galicia Bank cut test case development time by 60 percent. The mechanism is consistent: AI reduces the cognitive overhead of boilerplate and debugging, freeing developer hours for higher-value work. Institutions that have not deployed AI coding assistance are leaving documented productivity on the table.
Fraud detection and AML produce the largest absolute dollar savings. The sector saves an estimated $1.5 billion annually through AI-enhanced fraud detection, and 48 percent of institutions using AI for AML report saving $1 million or more per year. Two factors drive this: reduced false positive rates that lower manual review costs, and faster detection of novel fraud patterns that rules-based systems miss entirely. For institutions still running rules-only fraud detection, the comparison to AI-augmented peers is increasingly difficult to defend in examination contexts.
Customer service and contact center operations are generating 20–40 percent efficiency gains at institutions that have deployed AI at scale. The spread in outcomes is driven by integration quality — tools tightly connected to CRM and account management systems consistently outperform generic deployments.
Back-office automation across document processing, loan origination, compliance reporting, and reconciliation workflows produces consistent 20–30 percent efficiency gains at institutions that have moved to production. The key differentiator is ownership: high performers designate a business owner accountable for AI-assisted workflow outcomes, not just a technology team that shipped the tool.
Risk and compliance analytics — regulatory reporting, stress testing data preparation, model monitoring — reduce manual labor on compliance teams while improving output consistency. This is also where the governance investment required to deploy AI responsibly, the AI inventories, model documentation, and three-line accountability structures we described in our FS AI RMF coverage, pays operational dividends beyond regulatory compliance alone.
Why Most Institutions Are Stuck
Accenture's Q1 2026 survey found that 91 percent of financial institutions call AI strategically important, but only 23 percent have successfully moved to production at scale. The failure modes are consistent across the research.
Governance infrastructure is absent before deployment. Institutions that launch AI tools without documented ownership, monitoring, and oversight protocols face remediation under examiner pressure — a process that costs more than building governance correctly the first time. The agentic AI control frameworks we outlined in our piece on why SR 11-7 isn't enough address this directly.
Integration is underestimated. JPMorgan's and Goldman Sachs's productivity gains were not produced by deploying off-the-shelf AI tools. They were produced by AI systems tightly integrated with proprietary data, internal workflows, and institutional knowledge bases. Generic deployments consistently underperform.
Change management is treated as optional. The two to three hours per day Naranja X developers save with AI coding assistance required training, adoption support, and deliberate culture change. Without structured adoption programs, institutions see 20–30 percent utilization rates — a fraction of the available ROI.
Pilots are not designed to scale. A controlled pilot with 50 users often fails when expanded to 5,000 users across multiple business lines without the supporting infrastructure — identity controls, audit logging, monitoring, and exception handling — that production deployment requires.
What High Performers Do Differently
The institutions exceeding AI ROI expectations in 2026 share three consistent characteristics.
They start with high-volume, high-frequency workflows — developer productivity, fraud detection, customer service — where AI assistance produces compounding returns, rather than low-frequency, high-judgment tasks where the technology is less mature. They build governance infrastructure before deployment, not after, because that investment removes the bottlenecks that slow AI at scale. And they treat AI as infrastructure rather than a project: JPMorgan's $1.5 billion in AI value is the accumulated output of AI embedded across hiring, training, tooling, and governance — not a single initiative with a completion date. Institutions that replicate that infrastructure posture are the ones that replicate the returns.
Key Takeaways
- JPMorgan ($1.5B annual value), Goldman Sachs (12,000 developers on GitHub Copilot), and Citizens Bank (20% productivity gains) have published results that are documented and replicable.
- The highest-return deployment categories are developer productivity, fraud and AML detection ($1.5B in sector-wide annual savings), and contact center operations.
- Only 23% of financial institutions have moved AI from pilot to production at scale — the barrier is governance infrastructure, not technology maturity.
- High-performing institutions start with high-volume workflows, build governance before deployment, and treat AI as infrastructure rather than a project.
- The governance investment required by the FS AI RMF and examiner expectations is not a cost that competes with AI ROI — it is the enabling condition for achieving it at scale.
The Risk Dispatch covers the regulatory and technology developments that matter most to financial services technology leaders. For the governance framework that enables AI at scale, see our coverage of the FS AI RMF 90-day action plan and agentic AI governance beyond SR 11-7.