GenAI investment decisions have moved firmly into the boardroom. Capital is committed, pilots are running, and boards are asking one question: where is the return?
A 2025 study of 809 financial institutions found that early GenAI adopters experienced an average 428-basis-point decline in return on equity as they absorbed integration costs. For smaller institutions, that figure reached 517 basis points. Researchers call this the Implementation Tax: the near-term drag from infrastructure, data engineering, change management, and governance build-out before productivity gains materialise. The pattern follows an Innovation J-Curve. Losses come first. Gains follow, but only for institutions that sequence the transition with discipline.
A Phased Blueprint to Reduce the Implementation Tax
Phase one is data and governance readiness. No model should enter a customer-facing or risk-sensitive workflow until data is clean, accessible, and well-governed. Skipping this step compounds costs downstream.
Phase two targets internal, lower-risk use cases: document summarisation, compliance review support, internal knowledge retrieval. These deliver measurable productivity gains without customer-facing model risk and build institutional confidence before higher-stakes deployment.
Phase three expands into higher-value journeys: customer onboarding, credit decisioning support, and relationship manager tools in corporate and SME banking. The implementation tax falls hardest on institutions that attempt everything simultaneously. A phased model distributes that cost across a managed timeline.
Institutions that engage specialist partners for deployment consistently outperform those that build entirely in-house. Banking-specific GenAI carries regulatory, architectural, and change management complexity that benefits from prior pattern recognition.
An AI Value Scorecard: CX, Risk, and Operations
Ninety-five percent of enterprises struggle to demonstrate measurable AI return. Gartner estimates that 40% of agentic AI projects will be cancelled by end of 2027 due to escalating costs and unclear business value. A structured scorecard, reviewed at board level from the outset, directly addresses both problems.
Track across three dimensions:
- Customer experience:digital adoption rates, contact centre escalation rates, time-to-resolution, and self-service completion across key journeys.
- Risk and compliance: fraud interception rates, reduction in manual review hours, and model accuracy in credit assessment. Well-implemented AI fraud systems intercepted over 90% of fraudulent transactions before approval as of late 2025.
- Operational efficiency: time saved per task category, processing throughput, and cost-to-serve. All metrics should run against a pre-deployment baseline and be reviewed quarterly.
Procurement: What to Demand Beyond Accuracy
Accuracy metrics are necessary but not sufficient. Four criteria deserve equal weight.
- Architecture compatibility: Assess whether the solution operates on top of existing infrastructure or demands structural change beneath it. A platform requiring deep core integration will erode projected returns before the first use case goes live.
- Governance by design: Require vendors to demonstrate that model explainability, bias monitoring, and audit trail integrity are dependencies of the core product, not optional configuration layers. Regulators across multiple markets are tightening scrutiny on this.
- Data residency and sovereignty: This is one of the most common gaps in vendor proposals and one of the most consequential, particularly for institutions operating across multiple jurisdictions.
- Build versus buy: Finally, the buy-versus-build decision deserves rigorous analysis rather than a default position. Building proprietary models requires sustained investment in talent, infrastructure, and iteration cycles. Buying accelerates time to deployment but creates dependency on vendor roadmaps and pricing. A hybrid approach, combining pre-built banking-domain capabilities with internal fine-tuning on proprietary data, has become the pattern most institutions with mature AI programmes are converging on.
The Bottom Line
The institutions that show genuine GenAI ROI over the next two to three years will be those that sequenced deployment with discipline, built measurement in from the outset, and held procurement to a standard beyond demo performance. The implementation tax is real. Treated as a planning variable rather than an unexpected cost, it is also manageable.
Frequently Asked Questions
- What is the GenAI implementation tax in banking?
The near-term financial and operational cost banks incur during GenAI adoption: infrastructure, data engineering, change management, and governance build-out. Research across 809 financial institutions found an average 428-basis-point ROE decline before productivity gains materialise. - How should banks measure GenAI ROI?
Across three dimensions: customer experience (adoption rates, resolution times, self-service completion), risk and compliance (fraud interception, manual review reduction, model accuracy), and operational efficiency (time saved per task, cost-to-serve). Track all metrics against a pre-deployment baseline, reviewed quarterly. - What should banks demand from GenAI vendors beyond accuracy?
Architecture compatibility with existing infrastructure, governance controls embedded in the core product, clear data residency and sovereignty commitments, and a deployment model that avoids core system disruption. - Should banks build or buy GenAI capabilities?
Most mature programmes use a hybrid: pre-built banking-domain capabilities with internal fine-tuning on proprietary data. The right balance depends on data maturity, regulatory context, and internal capacity.