What is the speaker's vision for AI in finance?

The vision is to build a transparent, auditable, and sovereign ecosystem where AI can 'prove' rather than just 'predict', prioritizing governance and accountability alongside innovation.

Why is traditional AI like linear regression still relevant in finance?

Despite advancements, 80-90% of financial decisions still rely on models like linear regression due to their inherent auditability, control, and stable workflows, which are crucial for a regulated industry.

What are the four main pillars of AI in finance according to the speaker?

The four pillars are knowledge search/unstructured data, knowledge discovery/structured data, agent ecosystem development, and governance/interpretability.

How does the proposed platform handle legacy systems in finance?

The platform is designed to onboard a company's internal legacy knowledge, structured, and unstructured data, enabling AI to work with existing systems without requiring a complete overhaul.

What is the role of knowledge graphs in the proposed AI solution?

Knowledge graphs are crucial for integrating data, governing AI behavior, extracting features for financial analytics, and providing interpretable, temporally consistent insights.

How can organizations ensure trust in AI-generated outputs in finance?

By implementing systems that score every step of the AI process, from data retrieval to reasoning and API calls, providing users with transparency and control over the output's reliability.

What is 'domain guard' and its significance?

Domain guard is a small, dedicated model trained for governance tasks, aiming to perform model surveillance and ensure AI behavior aligns with regulatory and trust requirements.

AI Dev 25 x NYC | Stefano Pasquali: Building Trustworthy AI for Finance

DeepLearning.AI

Education3 min read25 min video

Dec 5, 2025|803 views|5|1

Save to Pod

Key Moments

TL;DR

Building trustworthy AI in finance requires a sovereign architecture combining knowledge graphs, LLMs, and agentic reasoning for transparency and accountability.

Key Insights

The majority of AI pilot projects in finance fail due to a lack of trust in AI outcomes for mission-critical applications.

A shift from AI hype to maturity necessitates treating data and governance as first-class citizens alongside LLMs.

Sovereign AI architecture, integrating knowledge graphs, LLMs, and agentic reasoning, is crucial for financial AI transparency and auditability.

Financial AI must prioritize regulation and risk management, demanding explainable and controllable AI models.

LLMs are a tool, not the sole solution; traditional ML models and knowledge graphs are essential components for robust financial AI.

Internalized, integrated AI platforms are vital for finance, ensuring security, compliance, and control over mission-critical use cases.

THE CHALLENGE OF AI ADOPTION IN FINANCE

The financial industry faces significant hurdles in adopting AI, with less than 10% of pilot projects reaching full-scale production. This low adoption rate stems primarily from a lack of trust in AI-generated outcomes for mission-critical applications. Past approaches have been hampered by vendor fatigue, the cost of large-scale models, and a tendency towards 'shadow AI' without proper governance. This highlights a critical need to move beyond the hype and embrace maturity, treating data and governance with the same, if not greater, importance as LLMs themselves.

THE VISION FOR SOVEREIGN AI ARCHITECTURE

Drawing from a childhood fascination with physics and graph theory, the speaker proposes a 'Sovereign AI' architecture. This unified framework combines knowledge graphs, Large Language Models (LLMs), and agentic reasoning to create a system that is transparent, auditable, and trustworthy. The goal is to build an ecosystem where innovation meets accountability, moving beyond mere prediction to the ability to prove and explain AI-driven decisions, which is paramount in a highly regulated industry like finance.

INTEGRATING KEY COMPONENTS FOR TRUST

The proposed architecture emphasizes a multi-pillar approach. Knowledge graphs are central for unifying structured and unstructured data, while agentic reasoning enables complex workflows and decision-making. Traditional machine learning models and existing APIs remain crucial, with LLMs acting as a powerful application layer rather than a complete replacement. This composite approach acknowledges that financial tasks like relative value calculations cannot solely rely on LLMs, necessitating integration with established analytical tools and data sources.

GOVERNANCE AS A FIRST-CLASS CITIZEN

A core tenet of this vision is a robust governance layer, akin to an 'MRI machine' for AI processes. This involves meticulously scanning every step, from data retrieval and reasoning to potential hallucinations and prompt injection. By scoring each process element and providing transparency, users gain a quantifiable understanding of their trust in the AI's output. This allows for policy implementation, such as automating execution for highly trusted recommendations.

THE IMPORTANCE OF INTERNALIZED PLATFORMS

The speaker argues against relying solely on centralized, external models for mission-critical financial use cases. Instead, a fully internalized, integrated AI environment is advocated. This approach ensures security, control, and compliance, allowing operations to continue even without external connectivity. While acknowledging the value of external tools for specific tasks, the focus for high-stakes financial decisions remains on platforms that leverage internal know-how and adhere to regulatory obligations.

ADVANCEMENTS IN KNOWLEDGE GRAPHS AND GOVERNANCE

Significant progress is being made in leveraging knowledge graphs, which are becoming more accessible with LLMs, though quality control remains key. New methodologies are being developed to clean and boost graph quality, enabling them to extract features for advanced analytics and govern LLM behavior. Furthermore, specific 'domain guard' models are being trained for governance tasks, such as model surveillance, aiming to create a powerful ecosystem for ensuringAI reliability and safety in financial applications.

FUTURE DIRECTIONS: CAUSAL REASONING AND AGENTIC PRODUCTIVITY

Future development is focused on next-generation analytics that move beyond correlation to causal reasoning. By merging knowledge graphs with LLM-based reasoning, the aim is to understand and trace the 'why' behind AI-derived insights. The ultimate goal is to enhance trader and portfolio manager productivity through multi-agent systems that provide a trusted, interpretable, and traceable environment for data retrieval, API interaction, and trade memo generation.

Mentioned in This Episode

●Companies

●Organizations

●Concepts

Common Questions

The primary challenge is the lack of trust in AI outcomes for mission-critical applications, leading to low production scalability and significant 'PC fatigue' from unfulfilled pilots.