Certain technologies are so compelling that they soon develop a life on their own. Generative AI (gen AI) leaped from being a laboratory to mainstream use in the latter half of 2022 and was able to make the leap after OpenAI launched a public version of its ChatGPT service. In just two short months, it was home to over 100 million customers and was the fastest-growing service in the history of mankind.
At the beginning of 2023’s first quarter, major tech companies were integrating Gen AI features into their products and providing programmatic use of generative models to business customers. In the year that has passed and Gen AI is making waves across a variety of sectors, including those that typically have taken a cautious approach to the implementation of new technologies. Credit risk, for instance.
McKinsey recently surveyed senior executives in credit risk at 24 institutions comprising nine out of the best ten US banks. We surveyed these executives on their organizations’ implementation of gen AI, as well as the current uses for it and their plans for the future of it, and the issues they anticipated.
20 percent of respondents have implemented at least one generation AI application in their companies, as well as a further 60% anticipate doing it within the next calendar year (Exhibit 1). However, even the cautious executives are convinced that the gen AI technology will become part of their businesses’ credit risk processes in two years.
Table of Contents
Use instances in credit risk
As banks gear up to implement Gen AI, they are evaluating possible applications throughout the entire credit lifecycle. These applications generally make use of the large-language models (LLMs) to mix to summarize, and analyze unstructured data as well as natural language. They can also generate complicated forms of natural language (such as emails, reports, and summaries of documents) and produce structured data or instructions for various software tools. Our research has revealed many possible uses for gen AI in the field of credit risk.
Most Popular Insights
- The agentic organisation is a new model for the AI time
- Global Banking Annual Review 2025: Why accuracy is not the same as heft. What the banking future will be
- The potential economic value of Generative AI: The next frontier of productivity
- Agentic AI deployment with security and safety Playbook for technology-focused executives
In terms of customer engagement, gen AI systems could give customers hyper-personalized products according to their profile and their activity history. Gen AI systems could assist relationship managers by writing individual outreach messages, analyzing meetings, and recommending future steps. Gen AI-powered virtual specialists could assist customers in identifying and selecting the most appropriate products.
During credit decision or underwriting processes, Gen AI tools can look over documents and alert authorities to policies that are not in compliance or have missing data. They can also draft outreach emails asking for clarifications or information that is missing for customers. They could also assist in compiling details about customers and conduct credit analysis ,and write different parts of credit memos prior to having credit officers read the memos. Agent-based Gen AI systems are able to independently follow task sequences to gather data from various sources, calculate pertinent ratios, evaluate results against standard thresholds, and then summarize the results on credit notes. These capabilities can be created in natural language with basic English without the requirement for programming or advanced modeling abilities.
After credit approval after approval, Gen AI can speed up and streamline the process of contracting. Gen AI systems are able to draft legal contracts, for instance or develop outreach messages to inform customers of the credit decision and any next steps to be taken.
For monitoring portfolios, Gen AI tools are able to assist portfolio managers in a variety of ways, including automating the preparation of routine risk and performance reports or writing summaries (based on the analysis of portfolio managers) of options for optimizing portfolios. Gen AI systems can even create specific strategies for optimizing subsegments in accordance with the organisation’s risk-aversion, and also optimize the current EWS. (EWS) by taking in real-time, unstructured data (such as market or news reports) to detect borrowers with high risk or segments of borrowers that require extra attention.
Additionally, Gen AI tools are able to assist customers with assistance procedures, such as crafting customized outreach communications to customers in case of problems. Gen AI systems are also able to determine the best restructuring options and help customers navigate the process of restructuring. Furthermore, certain institutions use Gen AI to guide the agents’ interactions with customers in real-time, as well as with post-call analysis.
Participants in our study report that they are looking into new AI applications in all of these areas. Portfolio monitoring is currently the top field of interest for respondents, with nearly 60 percent of them currently pursuing these uses. Credit applications are the second-largest category of activity reported, as are controls and reporting. About 40% of our respondents have plans or ongoing projects in both of these areas (Exhibit 2). In all business sectors, the respondents are able to see slightly more potential for Gen AI in wholesale credit than in retail credit.
The current state of Gen AI for credit risk
Gen AI has landed in the world of credit risk, but has yet to change it. Executives who were surveyed spoke candidly about the situation of their Gen AI-related applications. They are typically specific, non-customer-facing solutions that address specific operational issues.
For instance has created a proof-of-concept Gen AI tool that is able to prepare climate risk questionnaires for commercial customers. Relationship managers at the bank have to complete the questionnaires in order to conduct their monitoring of climate risk. Gen AI system, which is based on an LLM, is able to extract relevant data from annual reports as well as other disclosures. The documents are processed to determine relevant sections, which are then presented to the model, along with carefully-crafted prompts requiring the model to locate and summarize the most important details. The model gives a synthesized response, with pertinent citations to the source information. In the end, human experts on the subject evaluate and confirm the model’s results.
Another possible use-case, which many banks have investigated them is the use of Gen AI to draft credit memos. For commercial banks, the bank’s first line of business is often required to spend significant amounts of time acquiring data, conducting analyses, and writing memos to be used for credit decisions as well as underwriting requirements. Gen AI tools are able to perform tasks like gathering, extracting, and sourcing data; analyzing financial data; visualizing information; and writing sections of memos based on preset guidelines. Portfolio managers are then able to review the memo they have written, along with an estimate of the confidence level that is provided by the gen AI tool before approving the memo. Apart from freeing the space to perform other tasks the tool can also increase the accuracy and consistency of memos created and could accelerate the credit decision-making process.
The programmers of these systems is executed in natural language using agents-based systems, with no requirement for advanced programming capabilities. Meta-agents are able to coordinate precisely designed agents that specialize in certain jobs, to get results through multi-step tasks. For credit memos, they could include the extraction of information, calculation of ratios, as well as summarizing the details. A second layer of agents can enhance risk management mechanisms and assist in reducing common AI-related pitfalls, like hallucinations.
Since the introduction and use of next new AI systems, banks that have adopted them have reduced the time needed to respond to questions on climate risk by about 90 percent, ranging from nearly two hours down to just 15 minutes. The system’s answers are accurate 90 percent of the time.
Challenges
Executives recognize that scaling up the use of Gen AI for credit risk is going to be a challenge. The main obstacles that are cited by a majority of our respondents, are related to the governance and risk. The most significant risk categories that are that are associated with the use of generation AI include:
- Inadequate fairness of algorithms that mislead or confuse users
- IP infringements can be a result of copyright infringements or plagiarism
- privacy breaches resulting in the utilization of sensitive or personal data to model models
- the creation of malicious content
- security threads and security holes and
- Issues with performance and explainability
- the potential risk of the use of proprietary data from third party
- environmental social, as well as governance (ESG) consequences for example, an increase in carbon emissions, or worker disruptions
The main risks can result in legal, regulatory reputational, or business consequences if not managed properly.
Seventy-seven percent of participants identified the potential lack of new AI capabilities within the company. Other issues, as outlined by more than 50 % of participants are difficulties in defining the use instances and the stakes.
We wanted participants to discuss the necessity of frameworks or guardrails to control risks in Gen AI applications. The most important concern, as stated by the majority of respondents was data quality, which was followed by the risk of model (mentioned by 58 % of respondents) which include transparency and fairness, audibility and the ability to explain.
Lack of institutional support that is formal and coordinated for Gen AI within credit risk firms can make a difference in the way these issues are addressed. A mere third of institutions who were asked to provide an expertise center (CoE) for managing the gen AI usage instances. Only 10% of respondents say that their institutions have defined the gen AI usage cases centrally. They are usually developed in a decentralized way therefore, the common practices and lessons learned aren’t leveraged.
Building a gen AI ecosystem
To fully realize the potential of Gen AI in the area of credit risk, banks must transcend the current ad-hoc method and create a standard set of guidelines to prioritize, design and deploy, manage, and reuse AI applications. Eight of these practices are crucial:
- A AI roadmap: It should be in line with the business’s overall strategy, describe the needed capabilities and solutions, and offer an estimated timeframe to develop, launch and deployment on a large scale.
- Aligned processes for building gen AI tools: They should allow for rapid, but secure, end-to-end testing, extensive validation and the implementation of the solutions.
- Secure: AI-ready technology stack which supports hybrid cloud environments, ensuring that companies have the computing power needed to build models and use them in a massive manner. This stack must be able manage the unstructured and in-between data sources, develop and run models, and also pre and post-process data.
- Integration with foundation models for enterprise and tools: These are huge deep-learning neural network, like large language models that are the basis of advanced generation AI systems, and software toolkits that facilitate the customization and deployment of these models. Gen AI applications employ these models directly or draw on their models to create proprietary solutions.
- Automated and robust tools for supporting: They include machine-learning operations (MLOps systems that manage training, as well as the creation of models) as well as the right pipelines for processing and data infrastructure to aid in the development releases, maintenance, and development of applications.
- A talent and governance: Model that allows cross-functional skills to aid in gen AI development. The people providing this expertise might include software developers, natural-language-processing (NLP) specialists, teams to run reinforcement learning based on human feedback (RLHF), cloud computing specialists, AI product leaders, and legal and regulatory experts.
- A modular solution architecture: This allows parallel development as well as flexible connections across various layers, including the UX layer as well as that of the business logic layer.
- The result of these methods is a library that contains high-quality, reusable AI solutions and services. The items within the library are able to be connected to many scenarios for business and can be used across the value chain of credit.
Implementing the eight practices is a time-consuming task for many institutions and effort, however, deploying just a few practices can increase effectiveness and efficiency significantly. For instance, organizations that have embraced two practices say that successful gen AI installations have speeded up by 30 to 50 . The first is that these institutions use an architecture of modular solutions which comprises three layers including a user-experience layer and an enterprise logic layer as well as one layer of infrastructure.
They are facilitated by the operating model of the company. In addition, institutions can reuse components that are already in use and implement open-source libraries. for instance, developers are able to select and pick from a myriad of pre-made components (such such as prompt libraries, data-retrieval pipelines and guardrails) using open-source software to create an end-to-end AI solution in a short time, usually within 2 or 3 weeks.
