Torras.ai

In my recent masterclass with ADEN Business School, I explored how artificial intelligence is reshaping the financial sector. From predictive analytics to generative AI and now agent-based systems, banks are navigating a complex landscape filled with opportunities and risks.

Watch the full session (in Spanish):

The bad: Banking Chatbots Snapshot — August 2025

The table below summarizes a cross-section of banking assistants around the world. It is not exhaustive and may contain some inaccuracies, but it illustrates the landscape: transactional capability is common; generative capability is emerging; the combination of both remains absent.

Many banks are piloting generative AI. As of now, 0% combine generative AI with live transactional capabilities in a single consumer-facing chatbot.

Why the “0%” Gap Persists

• Explainability: Banks must justify decisions and responses; LLMs remain probabilistic and opaque.
• Reliability & hallucinations: Even rare mistakes are unacceptable in financial interfaces.
• Transparency & governance: Documentation, traceability, and data-lineage requirements are strict.
• Third-party & ICT risk: Many GPAI models are delivered by external providers; integration raises resilience and privacy concerns.
• Model & data controls: Training data quality, retention, and IP/privacy constraints complicate deployment.

The good: Why General Purpose AI (GPAI) Matters More Than Ever

GPAI refers to AI systems and foundation models that can serve many purposes across a bank, from customer support to coding support. Supervisors now focus heavily on GPAI:

• European Banking Authority (EBA): monitoring GPAI tests and early adoption across EU banks, especially in customer support and internal process optimization; urging caution around consumer protection, explainability, and reliability.
• EU Code of Practice for GPAI: emphasizes transparency, documentation of capabilities and limitations, data-handling, and quality controls over time.
• U.S. OMB AI guidance (M-24-10): sets responsible-AI expectations for the public sector and vendors, reinforcing principles of transparency, risk management, and evaluation that influence financial services best practices.

In practice, banks are experimenting with GPAI for:

• Customer service: drafting replies, improving resolution, powering employee Q&A on policies and procedures.
• Call centers: reliable transcription, summarization, and quality assessment.
• Engineering: code generation, error detection, legacy migration (e.g., COBOL to modern stacks).
• Legal & compliance: monitoring changes, summarizing rulings, and analyzing contractual impacts.

Agentic AI: From Static Bots to Doers

Agent-based architectures pair LLM reasoning with tools and permissions. A banking agent can follow a goal, call internal APIs, fetch documents, fill forms, and hand off to specialized sub-agents. This is the likely path to trusted “genAI + transactions”—but only with strong controls, auditability, and risk boundaries.

The ugly: Predictive AI Still Runs the Core

Old-school predictive models remain essential. Tools like gradient boosting and tree-based ensembles are still the workhorses for churn prediction, risk modeling, and marketing uplift. They’re explainable, auditable, and robust—provided they’re retrained as the market shifts.

What Banks Should Do Next

• Separate concerns: keep generative advice and transactional execution distinct until guardrails are proven; use explicit human confirmation for sensitive steps.
• Invest in governance: model cards, red-team testing, evals, incident playbooks, and dataset retention policies (multi-year).
• Prefer retrieval-augmented generation (RAG): ground answers on approved internal content; log citations for audit.
• Go “on-prem” or VPC where needed: for high-risk workloads, keep models and data in bank-controlled environments.
• Start with staff agents: employee-facing agentic use cases (ops, compliance, IT support) are lower-risk and high-ROI.

Bottom line: No bank has yet shipped a chatbot that fuses generative reasoning with live transactions for consumers. But GPAI experimentation is accelerating, regulators are clarifying expectations, and agentic architectures are maturing. The first movers will set the pattern—much as online banking did two decades ago.