Industry
Banking
Banking Services Transformation: Unlocking Exponential Value
Highlights
- The use of generative artificial intelligence (GenAI) in mortgage decisioning is growing. While this greatly improves efficiency and approval cycle times, it introduces ethical, regulatory, and operational risks.
- Unlike deterministic systems, GenAI output can deviate from prompt instructions and underwriting policies, potentially resulting in unfair lending decisions. Regulators are beginning to mandate that lenders must demonstrate the fairness of AI-driven decisions, which in turn demands transparent and auditable AI-backed mortgage systems.
- Mortgage lenders must adopt an ethical-by-design AI framework that incorporates guardrails spanning—behavioural transparency, instruction discipline, policy conformity, fairness surveillance, and governance and accountability—from the outset.
On this page
On this pageinpage
Overview
AI is increasingly being introduced into digital mortgage decisioning workflows to assist with document and guideline interpretation, risk summarisation, and underwriting support.
Large language models (LLMs) are particularly attractive because they can analyse unstructured borrower documents and generate contextual recommendations that traditionally required significant manual effort.
However, the use of GenAI comes with both operational and regulatory risks. Traditional solutions are deterministic and rule-driven, with logic tightly constrained to produce output within predefined criteria. However, GenAI models are trained using large volumes of historical data. Consequently, these models can sometimes deliver output that deviates from the explicit instructions given as prompts. In real-world situations, the model often leverages patterns learned during training, relying on data that could be inaccurate, incomplete, or outdated, resulting in output that may violate underwriting guidelines.
This behaviour raises serious ethical concerns around compliance with fair lending practices and can result in bias towards specific classes of borrowers. Mortgage lenders must ensure that AI-generated output does not result in unfair decisions for home loan applicants, especially given their obligation to comply with fair lending practices and the rising regulatory activity in this space. Though regulations on the use of AI in digital mortgage are evolving, some states such as California, Utah, and New York are emerging as frontrunners in introducing laws on AI use in mortgage origination. This shift places new demands on lenders to prove the transparency, explainability, and auditability of AI systems used in mortgage.
GenAI risks in mortgage decisioning
Traditional mortgage decisioning systems rely on deterministic logic where eligibility criteria are encoded into systems in a predictable manner, making them transparent and auditable
GenAI introduces a fundamentally different behaviour pattern. When an LLM receives a prompt, it does not retrieve predefined rules; it synthesises an answer based on a combination of prompt instructions and knowledge patterns embedded in its training data, occasionally deviating from explicit prompt instructions. For example, even when the prompt specifies certain evaluation criteria, the model may introduce additional reasoning derived from patterns learned during training, which may not align with underwriting policies. This is because LLM outputs are shaped by statistical relationships learned during training rather than strictly by the prompt.
This behaviour of an AI-powered digital mortgage solution can result in:
- Recommendations inconsistent with underwriting guidelines
- Lack of traceability in model reasoning
- Discriminatory practices and bias
With increasing regulatory oversight, explainable AI (XAI) or the ability to explain the reasoning behind AI-driven decisions is emerging as an urgent imperative. This requires mortgage lenders to design AI-driven underwriting workflows that are auditable–this will help demonstrate compliance with fair lending obligations and adherence to policies. AI output must be traceable, explainable, and capable of being independently reviewed—incorporating fairness safeguards into the AI solution architecture is therefore becoming mandatory.
How bias enters
Understanding how exactly bias enters a system is key to designing appropriate controls.
Bias can enter AI-assisted mortgage decision systems through various paths (see Figure 1).
Figure 1: Current AI-assisted mortgage decisioning flow and bias exposure
Training data bias: Historical enterprise data and information from public websites are used to train LLMs. If the training data is outdated, inaccurate, or contains incorrect interpretations, the model may adopt such patterns, which will corrupt its reasoning.
Prompt misinterpretation: LLMs use probabilistic techniques to interpret prompts. In the absence of clear guardrails, the model may augment prompt instructions with the assumptions or contextual knowledge learned during training. In mortgage decisioning, this may result in recommendations that appear reasonable but do not adhere to policy guidelines.
Feature proxy effects: Model inputs typically do not include protected attributes such as race, gender, religion, or ethnicity. However, this information can be inferred from attributes present in other documents, which serve as indirect proxies. Examples of such attributes are geographic indicators, employment patterns, or demographic information present in borrowers’ documents. If the model relies on such proxy attributes, its reasoning will be corrupted, leading to biased outcomes.
Output amplification: The output of GenAI models includes detailed narratives that can amplify reasoning. Even though the output appears persuasive and solid, the data underlying the reasoning may be inconsistent with fair lending guidelines.
Regulatory agencies require lenders to actively monitor lending bias in real time, not as a post-facto compliance check, which means that fairness must be embedded within the architecture and governance processes of the AI model. When bias is identified, mortgage lenders are expected to quickly intervene and mitigate it. Subsequently, lenders must implement controls to prevent repetition of such events.
Ethical-by-design architecture
The way forward for mortgage lenders is to treat bias mitigation as a core design principle.
This means incorporating guardrails across multiple layers of the AI-assisted mortgage decisioning workflow. Mortgage lenders will need to take targeted action to make this a reality.
Controlled instructions: To ensure AI systems operate within set boundaries, prompt engineering becomes highly critical. Prompt instructions should constrain the LLM model to evaluate borrowers’ loan applications strictly in accordance with underwriting policies. Using approved prompt templates reduces the likelihood of the model introducing reasoning inconsistent with established policies.
Policy alignment: Verifying AI recommendations against policy guidelines is key. If anomalies are detected, the system should flag the output for human review rather than implementing the recommendation.
Fairness monitoring: Lenders must put in place monitoring processes to evaluate the model output across multiple borrower scenarios to identify patterns that may indicate bias. The objective is to identify behavioural patterns in the model’s reasoning that may result in unequal treatment instead of relying on final outcomes alone.
Human oversight: Intuitive human reasoning is an imperative for sound mortgage decisions. AI-based decisioning systems should be used to assist and augment human underwriters by processing large volumes of borrower data, summarising risk factors, and offering decision recommendations. The final decision on rejecting or approving a loan must rest with trained and certified professionals who understand policy and regulatory expectations.
Mortgage lenders must incorporate auditability into prompt engineering to ensure traceability of instructions, model output, compliance with policy, and overrides by human reviewers. This will equip mortgage lenders to demonstrate that the AI-assisted loan decisioning engine operates within controlled boundaries without independently influencing lending decisions.
Actioning ethical AI
Creating a structured AI-powered mortgage framework involves ensuring that model behaviour aligns with underwriting guidelines, fair lending practices, and emerging regulatory requirements.
Mortgage lenders committed to transitioning to an ethical-by-design AI framework must traverse five interconnected phases (see Figure 2).
Figure 2: Ethical-by-design AI framework for mortgage
Phase 1 – Behavioural transparency: Monitor model behaviour to understand how it interprets borrower information to generate output. Identify output that violates explicit prompt instructions, and assess if reasoning is being influenced by inaccurate training data rather than established policy.
Phase 2 – Instruction discipline: Use pre-approved prompt templates underpinned by an underwriting policy to prevent the AI system from using generic knowledge learned during training.
Phase 3 – Policy conformity: Build policy verification mechanisms to ensure that the model’s reasoning is consistent with established lending practices. This will help flag policy violations.
Phase 4 – Fairness surveillance: Establish surveillance mechanisms to continuously monitor model behaviour across diverse borrower scenarios gleaned from a large number of loan transactions. This will help identify patterns of uneven treatment arising from flaws in the model’s reasoning and prevent discriminatory lending decisions.
Phase 5 – Governance and accountability: Maintain records of prompts, outputs, validation checks, and human actions to ensure traceability and transparency. This will prove invaluable in demonstrating ethical mortgage lending to regulators.
Future forward
The role of AI will evolve into an integrated decision support system, enabling human underwriters to quickly analyse complex loan scenarios and retrieve relevant policies.
As adoption increases, the focus will shift from AI as a tool for operational efficiency optimisation to broader aspects of trust, transparency, and accountability. Mortgage lenders will need to demonstrate that AI-backed processes are consistent, align with underwriting guidelines, and can withstand scrutiny around the explainability of AI recommendations.
The next major milestone will be the use of GenAI to analyse huge volumes of borrower data available with lenders. We foresee the evolution of an intelligent analytical assistant that can interpret borrower information within the limits imposed by policies and governance standards. In addition, the assistant will analyse borrower information to help lenders predict borrower behaviour, identify signs of delinquency and take proactive action to prevent it, retain existing borrowers, and enable perpetual customer engagement.
Industry watchdogs’ expectations around transparency and oversight will continue to evolve. In mortgage lending, the ability to combine advanced analytics with responsible oversight will become imperative to unlock AI value over the long term. Adopting an ethical-by-design AI system will equip mortgage lenders with the ability to continually adapt to emerging governance expectations. Bringing this vision to life, however, will require significant effort—mortgage lenders must consider partnering with a service provider with the requisite domain expertise and implementation experience.