The Evolution of AI Reasoning

The Evolution of AI Reasoning

How Deterministic Graph-Based Inference is Revolutionising Financial Services

The financial sector is increasingly recognising the need for AI systems that combine the power of Large Language Models with the precision and accountability of deterministic systems. As regulatory scrutiny intensifies, institutions are seeking solutions that offer both innovation and compliance certainty.

The AI Pendulum Swing: From Rules to Statistics and Back Again

The journey of artificial intelligence has been marked by significant shifts in approach. Early AI in the 1970s and 1980s relied heavily on expert systems with explicit rule-based logic to model human expertise. These systems evolved from simple if-then rules to more sophisticated knowledge representations including semantic networks and frames. Though limited by the technology of their era, these systems offered something invaluable: predictability and explainability.

By the 1990s, we saw the emergence of graph-based approaches as powerful tools for representing complex relationships between concepts. Knowledge graphs allowed machines to navigate interconnected facts and rules, drawing conclusions through logical pathways that could be tracked and verified-qualities that would prove increasingly valuable in regulated domains.

The 2000s and 2010s witnessed a dramatic pivot toward probabilistic models. Symbolic AI gave way to sub-symbolic methods as statistical machine learning and neural networks gained prominence. These approaches delivered remarkable performance but sacrificed explainability in the process, resulting in increasingly powerful but opaque systems. The financial industry, initially cautious, began exploring these tools for everything from fraud detection to trading algorithms, balancing the promise of performance against the growing "black box" problem.

Large Language Models (LLMs) represent the zenith of this probabilistic era. While they offer impressive language capabilities and seemingly intelligent responses, they also present significant limitations in consistency and precision-critical factors in regulated industries like financial services. These models, by design, generate outputs based on patterns and correlations rather than causal understanding, making them inherently probabilistic and sometimes unpredictably creative with facts.

Enter the Hybrid Era: Combining Strengths for Financial Compliance

We are now entering what industry experts are calling the "hybrid era" of AI-where organisations are combining the strengths of both paradigms. This new approach marries the flexibility and generative power of probabilistic models with the precision and explainability of deterministic systems.

For financial institutions, this hybrid approach offers a compelling solution to a pressing problem: how to leverage the language capabilities of LLMs while ensuring compliance, consistency, and transparency in a highly regulated environment. The UK's Financial Conduct Authority (FCA), among other regulators worldwide, is emphasising principles of fairness, accountability, and transparency in AI use, even as they map existing rules to new AI contexts.

The hybrid approach acknowledges that neither purely deterministic nor purely probabilistic systems can meet all the needs of modern financial services. Instead, carefully designed architectures that leverage both can create AI systems that are both powerful and governable.

The Challenge: LLMs in Financial Services

Banks and fintech firms are rapidly adopting LLMs for tasks ranging from customer service chatbots to assisting with complex regulatory analysis. According to recent industry surveys, over 70% of financial institutions are either implementing or planning to implement LLM technologies within the next 12-18 months. These institutions see the promise of AI in boosting efficiency, improving customer experience, and automating complex regulatory analysis.

However, deploying these models in a regulated environment introduces significant challenges that cannot be overlooked:

• Lack of determinism: LLMs can provide different answers to the same query on different occasions, making them unpredictable in situations that demand consistency. This variability complicates auditing and violates the determinism expected in finance IT systems.

• Hallucinations:They can confidently produce false or fabricated information. In a financial context, a hallucinated compliance rule or misinterpreted guideline could lead to wrong decisions and regulatory violations.

• Black box decision-making: It's difficult to explain exactly why a particular output was generated, creating issues for auditing and regulatory reporting. This opacity is problematic in finance, where firms must document decision-making processes for internal risk management and regulators.

• Bias & inconsistencies: LLMs learn from historical data, which typically embed societal or institutional biases. In finance, this can translate to discriminatory or inconsistent outcomes if not checked, potentially breaching laws.

• Prompt injection vulnerabilities: LLMs can be manipulated through specially crafted inputs that bypass built-in safety measures. For example, an attacker might structure a query that instructs the LLM to "forget previous rules" and then proceed to ask for guidance on circumventing AML regulations.

In financial services, where precision and compliance are paramount, even a single hallucinated detail (such as an incorrect interpretation of a regulation) can erode trust, cause legal trouble, and potentially result in substantial fines. The stakes are simply too high for probabilistic approaches alone.

Deterministic Graph-Based Inference: A Solution for Financial AI

Deterministic graph-based inference has emerged as a successful approach to guardrail LLM behavior in financial contexts. This methodology uses a model of the domain built from first principles, leveraging knowledge graphs with explicit rules and causal relationships to verify or replace LLM outputs.

At its core, deterministic graph-based inference refers to AI systems that derive conclusions via fixed logical rules encoded in a graph or network structure. Instead of relying on statistical prediction, these systems reason over a knowledge graph of facts and rules. Because the rules are explicitly programmed or learned from authoritative sources, the inference is deterministic: given the same inputs, the system will always produce the same output by following the logical connections, regardless of complexity.

This approach allows domain experts to encode decision logic (like regulatory criteria, compliance checklists, product eligibility rules, etc.) into a graph-based knowledge model. The resulting framework creates a structured representation of complex regulatory requirements and business rules that can be systematically applied.

The symbolic inference engine then uses this model to answer questions through logical deduction. Importantly, every step in the reasoning chain is recorded-providing complete traceability from question to answer. This is fundamentally different from how an LLM operates, which generates answers by predicting likely sequences of words based on patterns in training data.

Unlike traditional LLM approaches such as fine-tuning, retrieval-augmented generation (RAG), or prompt engineering-none of which are precisely deterministic-graph-based inference produces decisions that are:

• Traceable: Every decision can be mapped back to specific rules and data points

• Explainable: The system can provide a clear rationale for its conclusions

• Consistent: Identical inputs will always produce identical outputs

• Compliant by design: Regulatory frameworks are built directly into the logic

By encoding legal and regulatory expertise into deterministic systems, organisations can create a reliable framework where any advice or answer generated adheres to specific rules and regulations embedded in the knowledge graph. This approach transforms AI from a potential compliance risk into a literal compliance asset.

Implementation Approaches: How It Works with LLMs

Financial institutions can implement deterministic graph-based inference with LLMs through two primary architectural patterns:

Graph-First Reasoning

In this approach, the deterministic inference engine serves as the primary decision maker while the LLM acts as an interface layer, handling natural language understanding and communication. Critical decisions are made exclusively by the symbolic inference engine that traverses the knowledge graph.

The workflow typically follows these steps:

• The LLM processes the user query, extracts relevant information, and injects this data into the graph-based inference system

• The deterministic inference engine processes the data using the knowledge graph created from regulations

• When additional information is needed, a human can intervene to supply missing data or context

This architecture operates on a "deny all" principle, limiting answers to what can be generated safely from the graph. It guarantees precise outputs and eliminates hallucinations, though it is limited by the scope and complexity of the graph that can be produced.

Post-Generation Validation

In this alternative approach, the LLM generates complete responses that are subsequently verified and potentially corrected by the symbolic inference engine before being delivered to the user. The graph acts as a compliance checkpoint.

The workflow typically includes:

• The LLM drafts a response to the user's query

• Before reaching the end-user, the response is routed through the deterministic inference engine for verification-a sort of compliance officer within the AI

• If violations are detected, the system can block the answer, correct it, or escalate to human review

This approach marries the free-form generative ability of LLMs with strict rule adherence, creating more flexible and natural-sounding responses while maintaining compliance.

Benefits for Financial Institutions

Integrating deterministic inference guardrails with LLMs offers several compelling benefits:

• Transparency & Auditability: Every decision made by the graph-based system can be traced back to specific rules and data points, satisfying the growing demand for AI explainability in finance

• Elimination of Hallucinations: The knowledge graph serves as a grounding mechanism, preventing the LLM from introducing novel claims that aren't verified against established facts

• Regulatory Compliance Assurance: With regulatory knowledge encoded, the system actively enforces compliance, shifting AI governance from reactive to proactive

• Consistency and Reliability: Deterministic systems eliminate the randomness inherent in LLMs, ensuring customers and employees get consistent answers every time

These benefits address the core challenges that have made financial institutions cautious about widespread LLM adoption, potentially accelerating responsible AI implementation across the sector.

Learn More: Implementing Deterministic Guardrails

This approach to AI guardrailing is becoming increasingly crucial as financial institutions seek to balance innovation with regulatory compliance. For organisations looking to implement deterministic guardrails, a structured approach is essential-covering identification of high-risk use cases, knowledge capture methodologies, integration patterns, and testing strategies.

Rainbird's comprehensive whitepaper explores this topic in depth, detailing practical steps for operationalising deterministic graph-based inference in financial AI systems. The paper includes detailed implementation frameworks, architectural patterns, case studies, and even addresses how modern LLMs can assist in automating the knowledge engineering process itself.

Download the full whitepaper: "Deterministic Graph-Based Inference for Guardrailing Large Language Models: An Approach to Compliance and Control in Financial AI"

techUK - Unleashing UK Tech and Innovation

The UK is home to emerging technologies that have the power to revolutionise entire industries. From quantum to semiconductors; from gaming to the New Space Economy, they all have the unique opportunity to help prepare for what comes next.

techUK members lead the development of these technologies. Together we are working with Government and other stakeholders to address tech innovation priorities and build an innovation ecosystem that will benefit people, society, economy and the planet - and unleash the UK as a global leader in tech and innovation.

For more information, or to get in touch, please visit our Innovation Hub and click 'contact us'.

Upcoming events

Latest news and insights

Other forms of content

Receive our Tech and Innovation insights

Sign-up to get the latest updates and opportunities across Technology and Innovation & AI.