BankingNewsAI Daily Brief ·
UK FCA mandates banks prove AI testing works with documented evidence.
Banking AI
Financial institutions & fintech technology
UK FCA shifts from AI principles to proof: banks must show testing evidence that actually works
The UK FCA is explicitly pressing banks to demonstrate — with evidence, not theory — that their AI testing and controls are effective. This signals a move from high-level governance talk to supervisory scrutiny of validation, monitoring, and outcomes for AI systems already in use. Expect more pointed requests for artifacts (test results, model changes, incident learnings) rather than policy documents.
Action
Inventory production AI models and produce an “FCA-ready” evidence pack per model (testing methodology, stress/edge-case results, drift monitoring, human override metrics, and audit trails). Mandate pre-deployment and post-change validation gates that generate repeatable, regulator-consumable outputs.
TD puts agentic AI into secured lending operations (end-to-end real estate lending workflow)
TD Bank Group announced launch of its first agentic AI model to transform real-estate secured lending end to end. Unlike copilots, this is positioned as an autonomous workflow driver across a lending process, implying deeper integration with core systems and operational decisioning. It’s a concrete example of a top-tier bank moving from pilots to agent-led execution in a regulated credit domain.
Action
Pick one lending workflow with measurable cycle-time and error-rate pain (e.g., document intake, conditions clearing, renewals) and build an agentic pilot with strict human-in-the-loop controls and full decision logging. Benchmark against TD’s move: target operational KPIs (turnaround time, rework, exception rates) and establish guardrails before scaling.
National Bank of Canada backs Sardine with $25M Series C extension for AI fraud/compliance/credit risk
National Bank of Canada is leading a $25M Series C extension in Sardine, an AI risk platform spanning fraud, compliance, and credit underwriting. A bank-led round here is a signal that AI risk vendors are consolidating around unified platforms rather than single-point fraud tools. The ‘one platform across fraud + AML/compliance + credit’ pitch is gaining institutional credibility.
Action
Reassess your fraud/AML/credit tooling map for duplication and gaps, and test whether a unified AI risk layer can reduce false positives and manual review across multiple risk functions. Use the financing signal to pressure incumbent vendors on measurable outcomes (alert reduction, approval lift, investigator productivity) and integration timelines.
General AI
Large language models & AI infrastructure
Coding agents are shifting from chat to always-on automation with observability, remote execution, and CI/CD-style feedback loops
LangChain is pushing LangSmith Engine + SmithDB as an agent CI/CD + observability stack that detects failures from production traces, clusters issues, and drafts fixes/evals. Cognition launched Devin Auto-Triage as an always-on incident/bug first responder that can generate PRs, while OpenAI and Microsoft expanded remote execution/control for Codex and GitHub Copilot tooling. Net: vendors are standardizing on persistent agents tied to traces, memory, and verification loops—not one-off chat sessions.
Action
Pressure your SDLC/platform teams and key vendors (core, fraud, servicing) to show an “agent ops” plan: trace retention, eval gates, rollback controls, and human approval for any code/config changes before you allow autonomous remediation in production.
AWS makes SageMaker endpoints OpenAI-API compatible, lowering switching costs and reducing vendor lock-in
Amazon SageMaker added OpenAI-compatible API support for real-time inference endpoints, meaning teams using the OpenAI SDK/LangChain-style clients can invoke models on SageMaker largely by changing an endpoint URL. This is a structural change in how easily enterprises can shift inference between providers and host models in their own environment. It directly reduces migration friction and strengthens a multi-model, multi-host strategy.
Action
Standardize internal app integration on an OpenAI-compatible abstraction layer so you can swap between hosted and self-hosted models with minimal refactoring. Use the new compatibility to negotiate better commercial terms and to move sensitive workloads toward controlled hosting without rewriting client code.
Cohere releases Command A+ (218B sparse MoE) designed to run on as few as two H100s, pushing ‘sovereign/regulated’ deployment economics
Cohere released Command A+, a 218B sparse Mixture-of-Experts model positioned for enterprise/sovereign use with a relatively low GPU footprint (as few as two H100s). The important change is practical deployment: high-capability models are being packaged for smaller on-prem or tightly controlled environments. That makes “keep the model inside the perimeter” more attainable for regulated data than it was with prior heavyweight architectures.
Action
Re-evaluate on-prem/private-cloud LLM options for restricted datasets (customer data, investigations, legal hold content) with updated cost/performance assumptions. Pilot a private inference stack for 1–2 sensitive use cases and compare total cost per 1,000 tasks vs. public API models including compliance overhead.