AI in lending: From process efficiency to strategic reinvention

A loan application still requires processing numerous documents submitted in a variety of formats (PDFs, scans, emails, pictures, portal uploads, multiple languages, etc). Their manual extraction and validation are time-consuming for relationship managers, causing delays and increasing error risks. Intelligent document processing (IDP), powered by machine learning, natural language processing (NLP) and large language models (LLMs), automate document and information processing. First, document identification automatically classifies the document types (e.g. salary slip, tax return, bank statement, annual report) regardless of the format and quality. Second, document data reading uses NLP/LLM capabilities to locate and extract relevant data. For example, for mortgages it reads tax returns, salary slips, pension statements and fund documentation; for SME loans it extracts balance sheet tables, P&L statements and cash flow projections. Third, data extraction populates the loan origination system fields automatically. The system also generates modular document checklists based on credit type, displayed in the digital banking journeys to show customers which documents are required, were received, or are still needed. These first three steps can be fully automated. The fourth, data validation and interpretation remain semi-automated. While inconsistencies and complex cases can be flagged, or covenants or footnote details can be included, human review and judgement will be included. (e.g. discrepancy between salary slip and tax statement). Importantly, IDP capabilities are journey-agnostic and can be jointly developed with e.g. KYC/ onboarding automation efforts, increasing ROI across multiple customer journeys. The system learns from historical patterns as well as manual adjustments, improving accuracy with each document processed.

The credit analysis team spends approximately 40% of its time manually analysing data and assessing client eligibility against organisational standards and credit risk appetite. This creates bottlenecks, inconsistent decision quality and missed opportunities to guide the client towards approval – where a competitive Time-2-Yes is decisive. AI-powered credit assessment solutions transform this process by ingesting all organisational policies, rules, and case details (including client documentations) to support transparent and fast credit decisions. When a client application arrives, the AI enhanced system analyses comprehensively the financial profile against policy requirements, identifying both strengths and gaps. Rather than simply approving or rejecting, the system generates a complete assessment that highlights critical decision and internal control points and provides actionable guidance, especially for edge cases. For example, if a client's debt-to-income ratio is an Exception-to-Policy, it is not just flagged as a rejection trigger; but it proactively identifies concrete options: "Debt-to-income ratio is 40% (policy limit: 35%). Client could improve by: (1) increasing monthly income by CHF 500, (2) reducing existing expenses by CHF€500, (3) requesting a lower loan amount, (4) increasing down-payment or collateral by CHF 50’000." These insights feed directly into a comprehensive decision proposal for the relationship manager. The entire process is automatically documented, creating a complete audit trail for transparency and traceability. As a result, clients receive faster decisions and - when additional information is required - clear guidance on how to improve their applications. The system automatically updates when organisational policies or internal controls change, ensuring consistent decisions across all applications (and employees).

60-70% of operational costs in lending are attributed in portfolio maintenance. Monitoring individual loans still often relies on monthly payment signals or quarterly reports, which creates significant delays in risk identification and response times. By the time risk signals emerge, it’s often too late to mitigate. The process remains highly manual and time consuming with limited, fragmented data, preventing financial institutions from seeing the complete view of the portfolio health. AI-powered monitoring transforms this by analysing portfolio data in real time, leveraging a centralised data lake encompassing traditional loan performance metrics, transaction and payment data from client accounts, financial data (annual reports, balance sheets), as well as public data (press releases, regulatory filings). This integrated data foundation enables the AI system to monitor continuously for early warning signals, based on internal data (e.g. concentration risks, payment pattern changes, income volatility), external/ public record data (e.g. change in job, address relocation), as well as market trends (e.g. sector-wide stress affecting the client’s industry), before a problem materialises. Some FinTechs and banks even experiment with including exotic data source such as social media (e.g. user density in shopping areas to derive walk-in revenue trends, or user comment sentiments). Mitigating actions, impairment processes and Margin calls can be triggered automatically based on real-time risk signals rather than waiting for quarterly ex-post assessments. Additionally, any trigger is directly considered in the context of the wider portfolio to allow for a new level of risk assessment and management. This ultimately enables banks to better steer and lower their loan loss provision and RWA costs. The monitoring data is consolidated through interactive risk report dashboards that provide both Relationship Managers as well as Credit Risk Officers with immediate visibility into risk status, enabling rapid decision-making and intervention. This data enrichment can also be leveraged for client onboarding, enabling a more comprehensive risk assessment. Beyond early warning systems, the AI conducts continuous stress testing to model how the portfolio would perform under various economic scenarios (interest rate changes, market volatility, employment disruption). This enables institutions to identify vulnerabilities proactively and adjust risk management strategies.

Clients or bank employees will no longer need to navigate fragmented workflows for a credit request. Instead, their initial, and in the future likely sole, interaction point with(in) the financial institution will be a single agent, e.g. in a chat or dialogue style interface. This agent will orchestrate the end-to-end process seamlessly in the background, automatically triggering direct support from other specialised agents and the right sequence of next action steps (e.g. document collection, eligibility check, credit analysis and decision making). The user experience is transformed into a coherent journey and is no longer a sequential series of disconnected interactions.

Currently, decisions are made by humans based on a point-in-time snapshot of the client’s data, sometimes several months old. Agentic AI continuously ingests real-time data and dynamically enhancing recommendations, enabling decision-making based on up-to-date and enriched information. For example for many loans, especially Lombard and SME, the theoretical credit limit can be reassessed daily in the background, enabling instant approval for standard low-risk cases. Today this is not possible due to lack of computing power and capability to process data sets in real time. Additionally, the decision is made by leveraging not only the client information but also financial metrics, behavioural signals, the institutions risk appetite and current utilisation (i.e. for real-time risk-based pricing), market context and predictive indicators, increasing the robustness and accuracy of the decision. For straightforward cases, agents will make decisions autonomously in the future. For complex or Exception-to-Policy situations, they provide human experts with a complete analysis and recommendations. Overall this capability, dramatically accelerates decision-making and by design increases the transparency, traceability and governance of any decisions and escalations.

Through instant task execution by agents, the client experience can be transformed. Rather than requiring the client to provide complete information upfront, agents can generate preliminary offers based on the available information (e.g. initial onboarding data, public records, transaction history). As the client provides additional information or documentation (e.g. collateral information), the agent continuously reassesses eligibility and refines the offer. The agents enable a real-time reassessment, rather than a lengthy reapplication process. Therefore, based on their needs, clients can choose between a faster, standardised offer or supplying extra information to receive a more personalised offer with potentially better terms; agents support both routes.

Financial institutions can shift from reactive support to proactive partnership. Agents can continuously monitor the client’s situation and proactively engage, rather than asking questions to escalate a problem or merely fulfil regulatory review requirements. For example, they can identify when a client might benefit from refinancing, predict payment difficulties before they occur on existing products, as well as identifying potential cross-selling opportunities on solutions before the client realises that they need it. Additionally, moments-of-truth or important stages in life (e.g. move out of clients, approaching retirement) will be better leveraged.

Agentic AI systems don't just assess risk at a point in time; they continuously learn and adapt. As market conditions change, customer situations evolve, and new risks emerge, agents update their risk models and decision frameworks in real time. They identify emerging patterns in the portfolio that might signal new risks. They recommend preventive actions before problems materialise. They optimise the balance between growth and risk continuously, not just through periodic portfolio reviews. Through collateralization of loans or B2B lending platforms which enable refinancing of loans, institutions can flexibly adjust their RWA allocation in real-time in the background. Risk and capital management becomes dynamic and predictive, not static and reactive.

Comprehensive, real-time data across all stages of the lending journey, leveraging data lakes, and in addition encompassing public and third-party resources. Legacy data siloes must be broken down and data quality (if needed with prior data cleansing) becomes non-negotiable.

This is not only a technology shift: Job profiles, skills, interaction channels, hand over points, internal KPI and standards, etc. and ultimately the financial institution’s employees must evolve together. The overarching structure of a bank’s operating model and organisation may need to be redesigned to ensure value long-term as potentially not all value chain steps are anymore performed by humans or even in-house. Success requires a clear mindset towards embracing and adopting the change.

Where autonomous agents make more decisions, robust governance and risk controls are essential to provide consistency, traceability and model-risk management. All agent decisions must be auditable with clear documentation of data inputs, assumptions and logics applied (including escalation rules). Validations must be in place to manage AI-specific risks: data hallucinations (generating plausible but false information), model drift (performance degradation over time), and unintended biases. Robust controls are needed to detect and flag these issues before they affect the credit decisions.

The credit officer role is essential but will shift fundamentally: rather than investing time in administrative data entry or making routine credit decisions, credit officers become decision validators and risk managers. Their focus shifts to reviewing agentic recommendations for complex cases, validating logic and data quality, monitoring agent performance for emerging risks or biases, and continuously refining the models and rules underlying agent decisions. They act as the human safeguard ensuring agentic AI systems operate within risk tolerance and regulatory requirements. Not to be underestimated is also their responsibility to train the future cohort of Credit Officers so that they too will be able to challenge the system outputs with the equally quality and experience, instead of blindly relying on them.

Agentic AI requires robust orchestration platforms, secure APIs, real‑time data pipelines, explainability and monitoring capabilities, and strong security and privacy controls. Delivery demands both strategic vision and engineering rigour.