BankingNewsAI Daily Brief ·
Bank of Ireland appoints Citigroup executive Prag Sharma as its first chief AI officer.
Banking AI
Financial institutions & fintech technology
Bank of Ireland appoints Citigroup executive Prag Sharma as its first chief AI officer
Bank of Ireland hired Prag Sharma from Citigroup as its first chief AI officer. Creating a dedicated executive role signals that AI ownership is shifting from technology experimentation toward an enterprise operating agenda spanning business deployment, risk and governance.
Action
Clarify whether current AI accountability sits at sufficient executive level; assign one leader authority over portfolio value, model risk, data controls and workforce execution rather than dispersing responsibility across functions.
General AI
Large language models & AI infrastructure
Moonshot’s Kimi K3 brings frontier-class open weights, 1M-token context and lower-cost agent performance into enterprise model selection.
Moonshot AI launched Kimi K3, a 2.8T-parameter multimodal model with a 1 million-token context window, API availability, and open weights promised by July 27. Independent Artificial Analysis scored it comparably to Claude Opus 4.8 and GPT-5.5 at an estimated $0.94 per Intelligence Index task, while Arena ranked it #1 in Frontend Code Arena. Its reported 51% hallucination rate on AA-Omniscience is a material constraint despite improved accuracy.
Action
Add Kimi K3 to your next controlled vendor-model evaluation for document-heavy, coding and workflow-automation use cases; measure accuracy, hallucinations, latency and total task cost against your incumbent models.
OpenAI is documenting new failure modes from long-running AI agents
OpenAI published deployment lessons from long-horizon models that can work over extended periods rather than answer a single prompt. It identifies distinct safety risks and observed failures arising when models plan, use tools and pursue multi-step objectives, alongside safeguards refined through iterative deployment.
Action
Red-team agentic workflows against long-horizon failures—goal drift, unsafe tool use and compounding errors—not just prompt-level hallucinations, and require intervention checkpoints for consequential tasks.
Anthropic’s $1.5 billion copyright settlement has received final approval
A court approved Anthropic’s $1.5 billion copyright settlement, closing one major case over the use of copyrighted works in AI training. It does not settle the broader legal question, but it establishes a material financial benchmark for training-data disputes facing frontier-model developers.
Action
Reassess model-provider concentration and contract protections for IP claims; require clear indemnity, data-provenance disclosures and exit options in strategic AI procurements.