Anonymous Big Four UK Bank

As GenAI adoption accelerated across retail banking operations, a Big Four UK bank encountered compounding cost and compliance pressures. Unchecked token consumption from multiple competing LLMs created budget overruns and opaque cost attribution, undermining the business case for continued AI investment. Simultaneously, the bank's publicly stated ESG commitments demanded measurable carbon tracking across AI workloads — an area where standard cloud billing offered little visibility. Regulators and internal risk committees increasingly required evidence-based assurance that GenAI systems met emerging governance standards, not internal self-assessments alone. Without structured third-party verification, each new AI deployment carried elevated audit risk and reputational exposure in a heavily scrutinised sector.

The bank deployed a Travel Insurance Knowledge Assistant on Amazon AWS Bedrock, evaluating six model variants — GPT-4o, GPT-4o-mini, Llama 3.2 (1B and 3B), Mistral-7B, and Phi-3.5 — with Trustwise Optimize:ai serving as an intelligent optimization layer to manage token efficiency across workloads. A team of 75+ AI scientists and architects underpinned the implementation. The bank then engaged the RAI Institute's RAISE Pathways™ platform to complete a structured self-assessment across 125 controls spanning three verification domains: the Generative AI Foundations Badge (75 controls mapped to NIST AI RMF, ISO/IEC 42001, and OWASP); the Sustainable AI Verification Badge (25 controls per ISO 21031/SCI); and the AI Cost Efficiency Verification Badge (25 controls aligned with FinOps Foundation and TBM Council practices). Implementation rates across the three categories ranged from 72% to 87.3%.

The verification programme delivered measurable gains across cost, sustainability, and governance. Total Cost of Ownership fell 38% with maintained model performance and clear resource planning visibility presented to senior stakeholders. Carbon intensity of AI workloads dropped 66% through adoption of carbon-aware computing schedules, directly meeting internal ESG targets. The Generative AI Foundations badge reached a peak implementation rate of 87.3% across 75 controls.

• 38% TCO reduction while preserving model quality and output accuracy
• 66% carbon intensity reduction via workload scheduling aligned to grid carbon signals
• 87.3% peak controls implementation rate under the Generative AI Foundations badge

The bank also secured first-mover status as one of the earliest financial institutions to hold verified responsible AI credentials, with all three badges publicly recorded in the Responsible AI Registry.

• Third-party AI verification frameworks can simultaneously address cost, sustainability, and compliance — these need not be separate workstreams with separate budgets.
• Carbon-aware scheduling is a deployable, measurable tactic for reducing AI's environmental footprint; it does not require custom infrastructure beyond workload timing controls.
• Structured self-assessments mapped to established standards (NIST AI RMF, ISO/IEC 42001, FinOps) accelerate regulatory audit preparation and reduce last-minute scramble.
• Testing multiple LLM variants under a unified optimization layer surfaces real cost-performance tradeoffs that single-model deployments routinely miss.
• Publicly recorded verification badges create external stakeholder trust and competitive differentiation in regulated retail banking markets.