Agentic AI applied to risk management

What it is about

Diego Noguero, Director of Pre-Sales for Latin America at FICO, presented at the Fourth Annual Risk Conference (September 17, 2026) what agentic artificial intelligence is and how it changes risk management.

The 20 slides explain the evolution of AI in waves (rules, machine learning, deep learning and generative AI foundation models) and distinguish generative AI, which creates content and stops, from agentic AI, which pursues an objective, plans and executes actions within defined limits. They review the anatomy of an agent (reasoning model, memory, tools, orchestration and policies) and the challenges of AI: hallucinations, lack of transparency, data bias and cost.

They then place risk management in four steps (identify, measure, decide and monitor) and by use case, from credit risk and fraud to third-party risk, and set out what changes with agentic AI: speed, consistency, scale and traceability. The last part presents the FICO initiatives, with a platform of decision agents with governance, three purpose-built AI models and a demonstration of rule authoring with agents.

Objectives

• Distinguish generative AI from agentic AI and explain the anatomy of an agent.
• Relate agentic AI to the risk management cycle and its use cases.
• Present the FICO initiatives in agentic AI with governance and traceability.

Key takeaways

1. Agentic AI does not wait for an instruction at every step: it decides its path within defined limits.
2. An agent is made up of a reasoning model, memory and context, tools and actions, orchestration and policies with guardrails.
3. With agentic AI the decision goes from weeks or days to seconds, with the same policy applied the same way every time and every action documented and auditable.
4. According to the presentation, the FICO foundation model would be up to 1,000 times more efficient than a general-purpose language model.
5. The presentation cites up to 38% improvement in regulatory compliance cases with a focused model and more than 35% improvement in fraud detection with a sequence model.