[Note: over the past decade I have taught a variety of courses primarily with a focus on fintech.  Recently I was given the opportunity to merge two areas of interest: teaching about AI-related topics and the financial services industry. In addition to slides, below is a long form set of examples and exercises of the blended topic assisted via agentic deep research (guess which parts!).  This is intended for a classroom setting, feel free to re-use.]

Section 1: Introduction – Key Concepts and Definitions

Artificial Intelligence is a broad field of computer science focused on creating systems that perform tasks normally requiring human intelligence. Think of AI as an umbrella term for technologies enabling machines to learn, reason, and make decisions. For example, an AI chess program can evaluate millions of moves to play at a grandmaster level. Machine Learning (ML) is a subset of AI involving algorithms that improve through experience – in other words, ML systems learn from data to get better at a task without being explicitly programmed for every scenario (AI vs. ML vs. Generative AI: Basics). If AI is the overarching concept of smart machines, ML is one of the key techniques allowing those machines to learn. Generative AI is another subset of AI that focuses on creating new content. Rather than just analyzing existing data, generative models can produce text, images, or other media that mimic human-created content. In plain terms, Generative AI (GenAI) models like GPT-4 or DALL-E don’t just answer questions – they generate brand-new essays, code, visuals, and more, based on patterns learned from vast training data.

Why do these concepts matter to financial services? Financial firms run on data and complex decision-making. AI and ML are already helping banks and insurers analyze mountains of information faster and more accurately than humans. For instance, AI in finance is used to automate tasks and improve decision-making – from scanning loan applications to flagging fraudulent transactions. A recent industry survey found that by 2024, 58% of finance functions were using AI in some form. The relevance is clear: AI can crunch numbers, spot trends, and execute routine processes in seconds, which frees up people for higher-level work. Machine learning models drive credit scoring, trading algorithms, and risk management systems, learning from historical patterns to predict outcomes. Generative AI is newer on the scene, but it’s gaining traction for tasks like drafting research summaries or answering customer queries in a human-like way. In short, AI isn’t sci-fi for Wall Street – it’s becoming a day-to-day tool.

Financial services executives are paying attention because the industry’s core functions align well with AI’s strengths. Finance is data-rich and relies heavily on information processing, from reading legal documents to monitoring transactions. The World Economic Forum notes that financial services, with their data-heavy, language-heavy operations, are uniquely positioned to capitalize on AI advances. And they are doing so – in 2023 alone, financial services firms spent about $35 billion on AI, with investments projected to nearly triple to $97 billion by 2027. The motivations range from boosting efficiency to improving customer experiences. A simple way to think of it is: AI and ML can be your new analysts and assistants, working 24/7 without fatigue, handling tedious tasks, and surfacing insights from big data. For example, an AI system might automatically review thousands of loan documents overnight to identify those that need human follow-up – something that would take a human team days. As we’ll explore, this translates into faster decisions, lower costs, and often better outcomes. Executives don’t need to code algorithms themselves, but understanding these fundamentals is key to leading in the age of AI.

(Key definitions: AI – machines performing intelligent tasks; ML – systems learning from data; Generative AI – models creating new content. In finance, these technologies are driving automation and insights across operations, risk, and customer service.)

Section 2: The Evolution from Blockchain to Generative AI in Fintech

Every few years, a new technology trend captures fintech’s imagination. Not long ago, blockchain and cryptocurrencies were all the rage. From roughly 2016–2020, one could hardly attend a finance conference without hearing how distributed ledgers would revolutionize everything from payments to contracts. We saw a frenzy of crypto startups, ICOs, and even established banks experimenting with blockchain. This followed a common “hype cycle” pattern: initial enthusiasm, skyrocketing expectations, and then a sobering phase when challenges emerge. In fact, by 2024, industry observers noted that the blockchain hype had nearly died, replaced by intense interest in AI. Gartner’s 2024 Hype Cycle for emerging tech showed most blockchain innovations sliding into the “trough of disillusionment,” meaning they hadn’t lived up to the early hype. Technologies like NFTs, Web3, and decentralized exchanges – once media darlings – hit roadblocks in scalability, regulation, or clear business value. As one analyst quipped, “Blockchain tech just really hasn’t hit the heights that were promised… this is not an overnight sensation”.

So why has the spotlight shifted to Generative AI now? A perfect storm of factors made GenAI the new fintech darling around 2023. First, we witnessed breakthrough AI models (like OpenAI’s GPT-3 and GPT-4) achieving human-level fluency in language tasks. The launch of ChatGPT in late 2022 was a watershed moment – within two months it reached 100 million users, the fastest adoption of any consumer app in history. That mainstream success demonstrated the practical potential of AI to millions, including banking customers and executives. Suddenly, “AI” wasn’t just an academic concept; it was something you could talk to for advice or have draft an email. Interest in generative AI soared past crypto – Google searches for “AI” overtook “Bitcoin” as ChatGPT fever took hold. In one telling example of market sentiment, Microsoft’s stock jumped 6% (adding $120 billion in value) after it announced GPT-4 integrations. Some investors clearly believe AI is the next big disruptor.

From a hype cycle perspective, generative AI is approaching the “peak of inflated expectations.” Massive venture capital is flowing into AI startups. In 2024, an astonishing 42% of all U.S. VC funding went to AI companies (up from 22% in 2022). Fintech funding overall hit an 8-year low as many investors redirected their bets toward AI innovations. One VC observed that capital is flooding to where the next breakthrough is expected, and right now that’s AI. Why the optimism? Generative AI promises tangible near-term gains. Unlike blockchain, which often required complex industry consortia or new infrastructure, GenAI can be piloted within existing processes today. Banks can deploy AI assistants to help call center reps, or insurers can use GenAI to generate personalized policy documents. These applications don’t upend the whole business model – they enhance it. In other words, AI feels more immediately useful to solve day-to-day problems, whereas blockchain sometimes felt like a solution looking for a problem.

It’s worth noting that hype doesn’t guarantee smooth sailing. We should recall the lessons of the blockchain craze – after excitement comes the grind of integration, regulation, and proving ROI. Some fintech experts caution that generative AI in finance, while promising, faces practical hurdles like data privacy and model accuracy in a highly regulated environment. For example, a fintech newsletter humorously asked, “WTF is generative AI in fintech?” – concluding that while better chatbots and automated wealth management are likely, truly game-changing uses will require overcoming data access challenges and ensuring personalization. In other words, the industry needs to avoid AI snake oil and focus on realistic wins.

So why is GenAI gaining traction now? In summary:

• Maturity of AI technology – Years of ML research yielded models (transformers, GPT, etc.) capable of human-like outputs, ready for commercial use.
• Mainstream demonstration – ChatGPT’s popularity showed even non-tech users what AI can do, creating pull demand in enterprises.
• Immediate use cases – Financial institutions can apply GenAI to existing data and workflows (customer service, report generation, coding help) without a complete tech overhaul.
• Competitive pressure – No bank wants to be left behind. Just as firms felt urgency during the mobile banking wave, now AI is seen as a competitive differentiator. Executives hear how rivals are piloting AI for, say, faster loan processing, and they don’t want to miss out.

Generative AI’s momentum in fintech reflects a broader cycle: technology hype shifting from one “next big thing” to another. We rode the blockchain rollercoaster up and (for some) back down; now we’re on the GenAI climb. The excitement is justified – AI can genuinely transform financial services – but managing expectations and focusing on value will be crucial to avoid a hype hangover.

Section 3: Key AI Concepts Relevant to Financial Services

Let’s dive into some core AI applications in finance. In this section, we’ll break down a few buzzwords – NLP, predictive analytics, fraud detection, risk scoring, and chatbots – and give real examples of how they’re used in banking and fintech today.

Natural Language Processing (NLP)

NLP refers to the ability of computers to understand and generate human language. Financial services deal with mountains of text: think of earnings call transcripts, legal contracts, customer emails, compliance documents, news reports, and more. NLP is the AI technique that makes sense of all this unstructured text data. For example, banks use NLP to analyze financial news and social media sentiment – an AI system might scan news feeds and tweets to gauge market sentiment about a particular stock or economic event. This helps traders or risk managers react faster. Another common application is in compliance: NLP can sift through communications to detect potential regulatory violations or flag suspicious behavior (like an employee discussing confidential info in an email). Crucially, NLP underpins chatbots and virtual assistants – the AI needs to understand a customer’s question and fetch the right answer. Thanks to NLP, a customer can type or ask, “What’s my account balance?” or “I lost my credit card, what do I do?” and the system can interpret the intent and respond appropriately. In short, NLP is how we get from raw text or speech to actionable data in finance. It turns documents into data and queries into results.

Real-world example: JPMorgan built a tool called COiN (Contract Intelligence) that uses NLP to parse thousands of commercial loan agreements in seconds, extracting key terms (like loan amount, covenants, etc.) that would take legal officers hundreds of hours to identify manually. That’s NLP saving time and reducing errors in a labor-intensive review process. Another example – sentiment analysis – uses NLP to determine if text is positive or negative. Hedge funds have employed sentiment analysis on news or social media to inform trading strategies (e.g., detecting a sudden negative sentiment around a company could signal a stock drop). All these NLP use cases leverage the fact that a huge portion of valuable information in finance is locked in language – and AI can unlock it.

Predictive Analytics

Predictive analytics in finance refers to using historical data and statistical algorithms (often powered by ML) to predict future outcomes. It’s basically using data to answer: “What’s likely to happen next?” Financial institutions have been doing this for ages with models and forecasts, but ML turbocharges the process with more complex pattern recognition. In practice, predictive analytics shows up in credit scoring and risk modeling. For instance, an ML model might analyze thousands of borrower attributes and past loan performances to predict the likelihood of default for a new loan applicant. Upstart, a fintech lender we’ll discuss later, famously uses AI models that consider non-traditional variables (like a borrower’s education or employment history) to predict credit risk more accurately than a traditional FICO-based model. The result? Lenders can approve more people with confidence. In fact, one study found an AI credit model could double the approval rate while keeping defaults lower than traditional methods – a huge win for expanding access to credit responsibly.

Predictive analytics also drives investment decisions. Many hedge funds use ML to predict stock movements or option prices by finding hidden patterns in market data. AI can ingest decades of price data, technical indicators, macroeconomic signals, etc., and forecast short-term price trends (this falls under the umbrella of algorithmic trading). Risk scoring is another vital use: banks predict things like the probability a transaction is fraudulent, or the expected loss on a loan portfolio under various economic scenarios (stress testing). These predictions help allocate capital and set aside reserves. Insurers predict claim probabilities to price policies – increasingly with ML models that might take into account satellite imagery (for property risk) or even IoT sensor data from cars (for auto insurance risk). In each case, the AI model is trained on historical examples (e.g., past customers who defaulted or not) and learns the combinations of factors that predict the outcome. The payoff is often better accuracy than simple rules. As an example, a traditional credit score might misjudge some applicants, whereas an AI model might find that an applicant with a slightly lower credit score but a stable job and college degree is actually a good credit risk – approving them when a legacy model would not. According to industry reports, AI-powered credit scoring can indeed evaluate creditworthiness with greater accuracy and speed than older methods.

In summary, predictive analytics in finance is about anticipating – who will repay a loan, where markets are heading, which transactions look suspect, which customers might churn – and doing so more accurately by letting ML find patterns we might miss. It’s essentially data-driven crystal ball gazing, and it is core to many AI applications in the sector.

Fraud Detection

Fraud and financial crime are perennial concerns for banks, credit card networks, and insurance companies. AI has become a crucial weapon in the fight against fraud. Traditional fraud detection relied on hard-coded rules (e.g., flag a transaction if it’s over $X and overseas). That catches some fraud but also throws false alarms and can be evaded by clever criminals. Machine learning-based fraud detection instead learns the patterns of legitimate versus fraudulent behavior from data. For example, an ML model can analyze millions of credit card transactions and identify subtle anomalies: maybe a certain sequence of purchases, or a timing pattern, that often precedes a fraud case. These models can then flag anomalies in real-time. The advantage is they can adapt as fraud tactics evolve, whereas static rules can become outdated.

In practice, when you get a text about a suspicious charge on your card seconds after it happens, it’s likely an AI model that triggered it. Mastercard reported that AI has increased its fraud detection capability significantly – by some measures, up to a 20% improvement in catching fraud compared to prior techniques. In fact, businesses that integrate robust ML tools have seen up to a 40% improvement in fraud detection accuracy. That’s huge when you consider global card fraud losses run in the tens of billions of dollars. AI also helps reduce false positives (legitimate transactions incorrectly flagged), improving customer experience.

Modern fraud systems often combine multiple AI techniques: supervised ML to spot known fraud patterns and unsupervised ML to detect new outliers, as well as graph analytics to uncover networks of fraudulent behavior. A fascinating example: Mastercard’s AI team recently used generative AI plus graph technology to identify stolen card numbers being sold online – their algorithm could predict full card details from partial info and double the detection rate of compromised cards. By doing so, banks can preemptively block cards before fraud occurs. Similarly, anti-money laundering (AML) efforts use AI to monitor transaction flows and customer profiles, scoring the risk of money laundering activity. AI can cross-reference many data points (locations, peers, history) far better than a manual review of alerts. PayPal has shared that AI models help it block a large volume of fraudulent transactions across its platform, keeping fraud losses to minimal levels even as volumes grow.

In summary, fraud detection is a prime example where AI’s ability to find needles in haystacks pays off. Financial firms feed in streams of transaction data and user behavior, and AI models output risk scores or alerts in milliseconds. This not only saves money (preventing fraud losses) but also builds customer trust that their assets are safe. As fraudsters get more sophisticated, using AI to fight AI (some fraud is AI-assisted too) is becoming the norm. One key point executives should note: these systems continuously learn. Fraud detection AI gets smarter with every confirmed fraud or legit transaction, constantly refining its understanding of what “normal” looks like for each account or card. That adaptability is something traditional software couldn’t offer, which makes AI indispensable in this arena.

Risk Scoring (Credit and Beyond)

Financial services are fundamentally about risk management – assessing and pricing risk whether in lending, investing, or underwriting insurance. AI has supercharged risk scoring by analyzing far more data points than a human typically can. In credit risk, as mentioned, AI-driven credit scoring considers hundreds of variables about a borrower to produce a more nuanced risk assessment. Lenders like Upstart have shown that using AI models (that might include education, employment, spending behavior, etc.) can reduce default rates significantly at the same approval rates – one report noted 53% fewer defaults at the same approval rate compared to traditional models. That means AI can expand lending safely, which is a big win for business growth and financial inclusion.

Beyond consumer credit, enterprise risk uses AI too. Banks deploy ML in market risk to predict how certain portfolios might react to market moves (like stress testing a variety of extreme scenarios), learning from historical correlations. Operational risk – e.g., the risk of a process failure or human error – is harder to quantify, but AI can analyze incident databases and process data to identify where the next breakdown might occur. For instance, a bank could use AI to analyze call center transcripts and operational logs to predict an increased risk of fraud or errors during peak periods, allowing preemptive action.

Insurance firms use AI for risk scoring in underwriting. Instead of relying solely on actuarial tables, insurers feed ML models with granular data (telematics from cars for auto insurance, drone imagery for property insurance, etc.) to more accurately predict risk of claims for each policyholder. This can lead to more personalized pricing – rewarding low-risk customers with better rates and identifying high-risk cases more accurately.

One cutting-edge example: some asset managers are using ML for ESG risk scoring, parsing news and reports with NLP (natural language processing) to score companies on environmental, social, governance risk factors in near-real-time, rather than waiting for infrequent ratings updates. The common theme is more data, analyzed more intelligently. AI systems might crunch not just numeric data but also text and even images (like analyzing satellite images of crops for agri-commodity risk). The output is typically a score or classification – e.g., a numeric risk score for each loan applicant, or a “high/medium/low risk” tag for transactions or clients.

Executives should understand that AI-powered risk scoring continuously learns and recalibrates. If conditions change (say, a pandemic hits, altering consumer behavior drastically), AI models can be retrained on the latest data to update risk assessments. This agility can be a competitive advantage. Of course, with great power comes responsibility: using AI in risk scoring also raises concerns about explainability and fairness. Regulators are asking, “Can you explain why the model gave this credit score?” AI models, especially complex ones, can be black boxes. Financial firms are working on “explainable AI” techniques to translate model outputs into human-understandable reasons (like, “Loan denied due to high debt-to-income and short employment history”). Despite these challenges, the benefit is clear – more accurate and granular risk insight. In fact, many banks report that AI models have improved accuracy of risk predictions and cut losses, which in turn frees up capital (since fewer unexpected losses means less capital needs to be held in reserve). AI is becoming the analytical brain behind risk management, augmenting the experts who ultimately make the decisions.

Chatbots and Virtual Assistants

When you hear “conversational AI” in banking, it’s about chatbots – automated virtual assistants that interact with customers (or employees) via chat or voice. In recent years, many financial institutions have launched AI-powered chatbots to handle front-line service queries. For example, Bank of America’s Erica chatbot has been hugely successful: by 2024 it surpassed 2 billion interactions with 42 million clients. Erica can do things like assist with routine transactions (“Pay my credit card bill”), provide account information (“What’s my checking balance?”), and even offer financial advice insights (“How much did I spend on groceries this month?”). It uses NLP to understand questions and connects to backend systems to fetch answers or execute requests. Notably, Erica leverages AI techniques like language processing and predictive analytics to serve customers quickly. Over 98% of client inquiries to Erica are resolved within sub- minute responses, and only a tiny fraction need escalation to a human rep. This demonstrates how chatbots can massively scale up customer support at low marginal cost.

For financial executives, deploying a chatbot can mean handling thousands of customer chats concurrently without hiring thousands of agents, and improving response times. Chatbots are available 24/7, which meets the modern expectation for immediate service. Beyond basic Q&A, more advanced bots can handle tasks like onboarding new customers (guiding them through form filling, ID verification), troubleshooting common issues (“I can’t log into my app”), or even cross-selling (“You’re eligible for a credit line increase – would you like to explore that?”). In wealth management, some firms have virtual assistants for advisors – an AI chat interface where an advisor can ask, “Summarize this 10-page stock research report” and get a quick digest. JPMorgan and others have internal bots to help employees navigate procedures or find information quickly, improving productivity.

Early bank chatbots had mixed reviews – often they were too scripted and could frustrate customers with limited understanding. But the latest generation, powered by large language models, are far more conversational and capable. They remember context within a chat, handle slang or typos, and can cover a broader range of queries. We’ve also seen industry-specific tuning; for instance, a fintech startup might train a chatbot on its product FAQs and support logs so the bot becomes an expert in that domain. Personalization is another trend: bots that know if you’re a new customer versus long-time customer, or what products you have, and tailor the conversation accordingly.

One real-world case: Capital One’s Eno (a text-based assistant) alerts customers to possible double charges or unusually large tips, not just responding to queries but proactively messaging when it detects something odd. That’s an AI assistant acting almost like a guardian. Another: Morgan Stanley’s AI assistant for advisors uses generative AI to retrieve information from the firm’s knowledge base so advisors can answer client questions faster. These examples show the range from customer self-service to employee support. The bottom line: chatbots in finance reduce call center volume (cutting costs), improve response time and consistency, and can even drive engagement (some customers prefer the quick chat interaction). With generative AI, we’re now seeing chatbots that can handle more complex requests – like “Explain why my investment account went down this quarter” – by synthesizing data and narrative.

Importantly, these bots are continuously learning. They improve as they handle more conversations (learning from successes and failures, usually with human supervisors reviewing tough cases). And if implemented well, they escalate to humans at the right time – ensuring that high-value or high-stakes interactions get a personal touch. For a non-technical executive, the key takeaway is: Chatbots aren’t just a novelty; they’ve matured into reliable digital team members that can enhance customer service capacity and even quality. Bank of America’s Erica, for instance, began with basic tasks and over six years has expanded to support not just retail banking but also areas like Merrill Lynch (wealth) and corporate cash management, essentially scaling AI assistance across the enterprise. This success story underscores that when done right, AI assistants can become a core part of a financial firm’s service strategy.

(In summary, key AI concepts in finance include: NLP for understanding text (used in document processing and chatbots), Predictive analytics for forecasting outcomes (credit scoring, trading), Fraud detection with ML to catch anomalies in transactions, Risk scoring with AI models for credit and other risks, and Chatbots/virtual assistants providing automated conversational service. Each of these is already in real-world use, making financial operations faster, smarter, and often more cost-effective.)

Section 4: Fintech Case Study – AI-Focused Company Profile

To ground these concepts, let’s look at a high-profile fintech company that has AI at its core: Upstart. Upstart is an AI-driven lending platform founded in 2012 by Dave Girouard (former President of Google Enterprise) along with Anna Counselman and Paul Gu. The founding team’s goal was to improve access to affordable credit using modern technology and data beyond traditional credit scores. Interestingly, Upstart’s very first product idea was an income share agreement platform, but by 2014 they pivoted to personal loans – a move that set the stage for their AI focus in underwriting.

Core offering: Upstart’s platform uses machine learning models to underwrite consumer loans (like personal loans and auto refinance loans) for banking partners. Instead of judging a borrower only by FICO score and a few financial ratios, Upstart’s AI models consider non-traditional variables such as education (where you went to school, what degree you earned), job history, residence, and even how you interact with the application (some models subtly consider user behavior patterns). The AI analyzes over 1,000 data points to assess the likelihood that a borrower will repay. This approach expands credit access – for example, a young borrower with a short credit history but a good job and degree might be deemed creditworthy by Upstart’s model even if a traditional score is mediocre. According to Upstart, their AI-driven underwriting allows bank partners to approve almost twice as many borrowers with fewer defaults compared to conventional lending models. In fact, the CFPB (Consumer Financial Protection Bureau) did a no-action letter study and found Upstart’s model approved 27% more applicants than traditional methods at similar loss rates – an impressive boost in inclusivity.

How is AI embedded in their service? Essentially, Upstart is a fintech that presents itself as a better risk model. Banks and credit unions use Upstart’s platform (either referrals or a white-label solution) to evaluate loan applicants. When an applicant applies, Upstart’s cloud-based AI crunches the data and instantly returns a credit decision and interest rate. This is all automated – more than two-thirds of Upstart-funded loans are fully automated, with no human underwriter intervention. That automation is powered by the continuous learning of their models. Each month, as loans repay or default, the AI retrains to refine its predictions. Upstart also uses AI in other parts of the process: for example, fraud detection during application (flagging if an identity might be stolen) and income verification (using document recognition AI to read paystubs or bank statements). But underwriting risk is their secret sauce.

Founding story & growth: Upstart’s founders came from tech (Google), not banking, which gave them a fresh perspective on lending. Their thesis was that many people are creditworthy but get overlooked by traditional scoring. By 2014, they partnered with a few forward-thinking lenders to test their AI models. The results were strong, leading to more partnerships. Upstart raised significant venture funding (including from folks like Eric Schmidt and Mark Cuban) and eventually went public in December 2020. The IPO and subsequent performance showed investor enthusiasm for AI in finance – Upstart’s revenue jumped as more banks signed on to use their platform, reaching $842 million in revenue in 2024 (serving over 3.3 million customers as of last year). Their success has also been a proof point that AI can reduce bias in lending: Upstart has published findings that their model improved access to credit for minority groups while still reducing default rates, an important aspect as regulators watch AI fairness closely.

Beyond personal loans, Upstart is now expanding into auto loans (using AI for instant approval of car financing) and small dollar loans. Their strategy is effectively “AI-as-a-service” for lenders. A traditional bank might be risk-averse about certain segments, but by using Upstart’s AI, they can safely lend to more people. Upstart’s value proposition to banks is increased approval rates, lower loss rates, and a fully digital lending experience to attract tech-savvy customers. For borrowers, it often means a fast online application and potentially getting approved even if they lack a top-tier credit score.

AI in practice: From an executive standpoint, what Upstart does with AI is a great illustration of letting the data speak. They identified that many borrowers were being lumped as “average risk” by legacy models, but in reality some were much safer. By ingesting a richer dataset (academic credentials, job history, etc.), their AI discovered patterns – for example, someone with a stable engineering job and good grades might have a low chance of default, even if their credit history is short. These signals aren’t obvious to a rules-based system, but the AI can quantify them. Upstart then continuously validates its model in production. The result has been strong performance: Upstart-powered loans often have lower default rates at the same approval level, meaning it can make lending both more inclusive and less risky. That caught the attention of major banks and even regulators as a potential win-win application of AI.

In founding stories, culture matters too. Upstart highlights that it considers itself more of a “tech company” than a finance company in terms of ethos. They have AI researchers, data scientists, and even had their co-founder Paul Gu testify to Congress about AI in credit underwriting – playing a role in shaping how policymakers view AI fairness.

To summarize Upstart: It’s a fintech unicorn built on the premise that better data and better algorithms can drive better lending decisions. The company’s journey from a startup to an AI lending platform working with dozens of banks (and becoming publicly traded) exemplifies how AI isn’t just theoretical in finance – it’s delivering results. Upstart’s story also shows the importance of continual learning and improvement. Their competitive moat is the years of loan performance data they’ve accumulated, which keeps training their models, hopefully keeping them ahead of would-be competitors. For financial services executives, Upstart is a case study in leveraging AI to transform a core business process (credit underwriting) and create value by reducing risk and unlocking new customer segments. As of now, they’ve facilitated over $40 billion in loans and their success has inspired many other fintechs and incumbents to invest in AI for lending.

(Why it matters: Upstart demonstrates how an AI-focused fintech can challenge traditional methods. By using machine learning for risk modeling, they achieved higher loan approvals with lower defaults, attracting both customers and bank partners. The company’s rise – from founding by ex-Googlers to $800M+ revenue – exemplifies the power of AI-driven strategy in financial services.)

Section 5: Mini-Exercise – Identifying AI Use Cases in Your Organization

Now let’s turn the lens inward. The goal of this mini-exercise is to help you identify opportunities for AI in your own organization’s workflows. Financial services executives often ask, “Where can we apply AI effectively?” This exercise provides a simple, structured way to brainstorm that. It doesn’t require technical knowledge – just knowledge of your business and an open mind about improvement.

Step 1: List Key Processes or Workflows. Start by jotting down 3–5 core processes in your team or department. Preferably choose processes that are fairly repetitive, involve a lot of data processing, or currently consume significant employee time. Examples might be: processing mortgage applications, reconciling transaction data at month-end, reviewing compliance reports, responding to customer FAQs, generating risk reports, etc. Basically, any workflow where you sense inefficiency or a heavy manual burden is a good candidate. If you’re in insurance, maybe underwriting evaluation or claims triage. In wealth management, maybe portfolio rebalancing or preparing client review decks. List these out.

Step 2: For each process, ask – What are the pain points? Why did this come to mind? Perhaps loan processing is slow because analysts sift through documents manually. Or compliance report review is tedious because someone reads lengthy texts to find issues. Write a sentence or two about the challenges or bottlenecks in each workflow. For example: “Mortgage application review – pain point: verifying income and documents takes 3 days and multiple follow-ups.” Or “Customer service – pain point: reps answer the same 10 questions over and over, tying up phone lines.” Identifying the pain helps clarify where AI might add value (usually by speeding things up or improving accuracy).

Step 3: Envision a solution with AI. This is the creative part. Ask yourself, if I had a magic AI tool, what could it do here? Don’t worry about feasibility just yet – think conceptually. For the mortgage example, maybe an AI could automatically read and verify documents (using NLP for text and even computer vision for document images) and flag any issues to the human underwriter, thus cutting review time to same-day. For the customer service example, maybe a chatbot or AI assistant could handle those top 10 questions across chat or voice, reducing call volume by 50%. Write down one idea per use case. It could be as simple as “Use NLP to analyze compliance documents and highlight key deviations, instead of an analyst reading line-by-line” or “Use ML to predict which transactions are likely errors and prioritize those for review”. The idea is to connect a pain point with an AI capability. Common patterns include: automation of a manual data task, prediction to inform a decision, or insight extraction from data. If you recall the earlier sections: NLP could apply to any text-heavy task, predictive analytics to any forecasting or scoring task, computer vision to any image-heavy task (e.g., inspecting property damage photos), and chatbots to any high-volume Q&A context.

Step 4: Consider data and feasibility briefly. Next to each idea, note what data or inputs would be needed and if those are available. For instance, an AI document reviewer would need digital copies of the documents – do you have them in a consistent format? A fraud prediction AI would need historical transaction and fraud outcome data – do you log confirmed fraud cases to train on? You don’t need a detailed plan, just a reality check. Also consider any compliance or privacy constraints (customer data used in AI must be handled carefully, etc.). The point is to ensure the idea has the raw ingredients (data and permission) to be explored further. Many AI projects fail not because the algorithm doesn’t work, but because the necessary data was inaccessible or of poor quality. So if one idea would require data you simply don’t have, you might mark that as a lower priority.

Step 5: Prioritize use cases. Look at your list and rank which one or two ideas seem most promising. Criteria for priority: potential impact (time/cost savings or revenue upside), and feasibility (doable with available data/tech within a reasonable time). Often, starting with a “low-hanging fruit” – a process that is relatively self-contained and where AI can make a measurable difference quickly – is a smart strategy. For example, deploying a chatbot to handle routine inquiries might be easier and faster to implement than a complex trading algorithm overhaul. On the other hand, a high-impact area like fraud prevention might be worth tackling first if fraud losses are a big pain point, even if it’s a bit more complex to implement. Circle the top 1–2 use cases on your list that you want to explore further.

This exercise is provided as a downloadable worksheet (see: AI Use Case Brainstorm Worksheet). Feel free to print it out for your team or use it as a template in a workshop. The worksheet guides you to fill in: Process name → Pain points → AI ideas → Data needed → Priority. By completing it, you effectively create a mini-inventory of AI opportunities in your domain.

Why do this? Because successful AI adoption starts with identifying the right problems to solve. It’s easy to be wowed by AI capabilities and try to shove them in anywhere; it’s better to start from real business needs. In fact, industry research shows much of the early AI adoption in finance has focused on efficiency gains in internal processes – precisely those pain points where AI can save time and money. By reflecting on your workflows, you tap into that same principle: find where AI can drive efficiency or insight, and you have a strong use-case to pursue.

Take a few minutes (or better, gather your team for 30 minutes) to run through this exercise. It often sparks valuable discussion and ideas. In Section 6, we’ll talk about sharing and mapping these ideas in a group setting, which can further refine your strategy.

(Mini-exercise recap: list your key workflows, note pain points, brainstorm how AI/GenAI could help, consider needed data, then prioritize. Use the downloadable AI Use Case Worksheet to structure your thoughts. This sets the stage for deeper discussions on integrating GenAI into your operations.)

Section 6: Worksheet Activity – Mapping GenAI Applications to Workflows (Group Discussion)

Building on your individual brainstorming from Section 5, the next step is a collaborative worksheet-based activity. The idea here is to take the potential AI use cases identified by different team members and map them onto your existing workflows in a group setting. This helps in two ways: it surfaces a variety of ideas, and it ensures cross-functional perspectives (since one department might not realize a workflow in another department could also benefit from AI). Group discussion can refine ideas and catch blind spots.

Activity setup: Schedule a small workshop with colleagues (could be within your leadership team or include reps from various units – risk, operations, IT, etc.). Aim for a group of 5–10 people so it’s manageable. Everyone should bring their completed “AI use case” worksheet or notes from Section 5. Prepare a larger version of the worksheet or a whiteboard divided into sections (one section per major workflow, for example).

Step 1: Share identified use cases. Have each participant briefly share one or two top use cases they wrote down. As each speaks, note the use case on a board or shared document. For instance, someone from compliance might say, “I identified regulatory change monitoring – AI could track new regulations and summarize what we need to do.” Another from customer service might offer, “AI chatbot to handle routine inquiries about card activation.” List all these. You’ll likely notice clusters or related ideas. It’s okay if there are duplicates – that shows consensus on pain points! Group similar ones together. For example, multiple mentions of “automating document review” could be grouped under that theme, even if for different document types.

Step 2: Map to workflows or departments. Now align the ideas with where they apply in your organization. You might make columns for Operations, Risk, IT, Customer Service, etc., and place each idea under the relevant one. Or, if you prefer, map by process categories like “Customer Onboarding,” “Transaction Processing,” “Compliance Monitoring,” etc. The goal is to see which parts of the business would be touched by GenAI solutions. This mapping helps ensure you involve the right stakeholders later for each idea. For instance, an idea for an AI trading assistant clearly falls under the Trading desk workflows, whereas an AI that writes management reports might fall under Finance or Strategy.

Step 3: Discussion prompts: Use prompts to stimulate discussion on each cluster of ideas. Ask questions like: “Which of these use cases do we think could be implemented with current technology (low hanging fruit)?” – this gets people talking about feasibility. Another: “Which idea, if successful, would yield the greatest benefit for us (in cost savings, revenue, risk reduction, or customer satisfaction)?” – this focuses on impact. You can also discuss potential challenges: “What data would we need for this, and do we have it?” or “Are there regulatory hurdles to using AI here?” Encourage honest input – maybe IT points out that some data is siloed, or Legal raises a flag about customer consent. This is good; better to surface concerns early.

One effective prompt is to have each participant vote on the top 1 use case they believe in (outside of their own area). This can quickly highlight a frontrunner that the whole group feels is valuable. For example, you might find that almost everyone votes for “AI for fraud detection” because it cuts across units (risk, IT, ops all see value in it). Or maybe “customer service chatbot” gets a lot of support because it directly improves client experience which is a strategic goal.

Step 4: Prioritize and assign next steps. Based on the discussion, narrow the list to a few high-priority GenAI opportunities. Often, groups will settle on 1–3 ideas that stand out as both impactful and feasible. These become your short list. For each of these, it’s useful to assign a small task force or point person to investigate further (a next step might be doing a quick proof-of-concept or gathering more data on ROI). For example, if “AI document processing for loan applications” is chosen, you might assign someone from Operations to work with someone from IT to explore vendors or internal capabilities for OCR/NLP solutions, and report back. Essentially, convert ideas into action items.

Using a worksheet for group work: We provided a worksheet that can also be used in group format – think of it like a canvas. There’s also an available GenAI brainstorming worksheet (link provided) which includes sections for listing workflows, AI ideas, data needs, and expected benefits. In a group meeting, you could fill this out collaboratively on a screen or printed large format. For instance, under “Workflow: Customer Onboarding,” you collectively fill “AI idea: ID verification with facial recognition and NLP to read documents – Benefit: reduce onboarding time from 5 days to 1 day – Data: Need access to govt ID database, good quality scanner data.” Doing this live gets everyone on the same page regarding what the solution entails.

Small group discussions on findings: The benefit of group discussion is peer feedback – someone might have already tried a similar AI tool and can share lessons, or someone might see a risk others didn’t. For example, the marketing team could mention they implemented a basic chatbot last year and have usage stats, informing the customer service chatbot idea. Or Risk Management might caution that any AI for credit decisions will need model governance and sign-off from regulators, adjusting how you approach that use case. These insights ensure that when you proceed, you’re doing so with eyes open.

One prompt to spur conversation: “How might these AI use cases change our employees’ roles or required skills?” This gets into change management considerations. If AI handles certain tasks, employees might focus on higher-level oversight. Are there training needs? This could surface in the discussion and is valuable for planning. Another prompt: “Do we build or buy?” – some IT folks might have views on whether to develop an in-house solution or consider third-party AI vendors for each idea. While you won’t resolve that in this brainstorming, it’s good to note preferences or constraints (perhaps a strategy is to try an off-the-shelf solution for chatbots, but build in-house for proprietary trading models, etc.).

By the end of this worksheet activity, you should have: (a) a refined list of GenAI application ideas mapped to where they fit in your business, (b) a sense of priority among them, and (c) initial thoughts on resources/data needed. This output is essentially a mini “AI roadmap” for your organization. It’s the bridge between theoretical potential and practical implementation plan. Many organizations find that doing this exercise reveals quick wins they hadn’t initially considered, as well as aligning leadership on pursuing AI strategically rather than ad hoc.

As an example, let’s say in your bank’s workshop, the top three ideas that emerged are: 1) an AI assistant for loan officers that summarizes customer financial info and suggests loan options (to speed up loan meetings), 2) a generative AI tool to draft and personalize marketing outreach (using customer data to tailor product offers), and 3) a regulatory compliance analyzer that uses NLP to monitor new regulations and compare against your policies. You’d then allocate these to respective managers to scope out. This targeted approach is much more effective than just saying “we should do something with AI” in abstract. It also creates internal buy-in, since the ideas came from the team itself.

Finally, encourage the group to keep the dialogue going. Maybe set up a periodic check-in (an “AI task force” meeting) to track progress on the ideas. The technology is evolving fast – what wasn’t possible a year ago might be doable now – so maintaining a collaborative, cross-functional forum for AI opportunities will keep your organization nimble and informed.

(Group activity recap: share and compile individual AI use case ideas, map them to business workflows, discuss feasibility and impact using prompts, then prioritize. Use the provided worksheet as a collaborative tool. The outcome is a shortlist of promising GenAI applications in your organization and concrete next steps for exploration. This ensures diverse input and organizational alignment on where to pilot AI solutions.)

Section 7: Prompt Engineering Fundamentals (Hands-on Intro)

We’ve identified what we want to do with AI – now let’s talk about how to talk to AI. This is where prompt engineering comes in. In plain language, prompt engineering is the craft of writing effective inputs (prompts) to guide generative AI models to produce useful outputs. If you’ve ever asked ChatGPT a question and then rephrased it to get a better answer, you’ve done prompt engineering on a basic level. It’s about knowing how to ask.

Think of a prompt as your instructions or question to the AI. Because generative AI (like GPT-4) can do so many things, the prompt is critical in telling it what you want. A poor prompt might yield irrelevant or confusing output; a well-crafted prompt can yield amazingly precise and helpful results. For example, you could prompt an AI writing assistant, “Write a summary of this 10-page policy document.” But if you just say “Summarize this,” you might get too generic a response. If you say, “Summarize the attached 10-page policy document, focusing on the compliance requirements and written in bullet points for a non-technical audience,” you’re giving much clearer guidance – the output will likely be a concise bulleted summary highlighting compliance points, as you intended.

At its core, prompt engineering is about designing inputs that produce optimal outputs. McKinsey defines it succinctly: it’s “designing inputs for AI tools that will produce the best possible outputs.” In practice, that means choosing your words, format, and level of detail in the prompt intentionally. Generative AI is powerful but somewhat like a genie – it will take what you say quite literally, so you must phrase your wishes carefully!

Let’s do a quick hands-on illustration. Below are a few example prompts and how a generative AI (like a GPT-3.5/4 model) would respond:

• Example 1 (general prompt): “Explain what a balance sheet is.”
  AI’s response: “A balance sheet is a financial statement that provides a snapshot of a company’s financial condition at a specific moment in time. It lists assets, liabilities, and equity… etc.”
  This is fine, but pretty general.
• Example 2 (more detailed prompt): “Explain what a balance sheet is to a high school student, and give an analogy.”
  AI’s response: “Imagine a company’s finances like a weighing scale… (analogy follows). A balance sheet is a financial report that shows on one side what the company owns (assets) and on the other side what it owes or is worth (liabilities and equity)… etc.”
  Notice how adding the audience (high school student) and asking for an analogy guided the AI to produce a more accessible and creative explanation. This is prompt engineering in action – giving context and style direction.
• Example 3 (financial-specific prompt): “You are a financial advisor. Summarize the key points of the attached quarterly market outlook in 3 bullet points, and do so in a neutral tone.”
  AI’s response (imaginary): “- Global equities saw moderate gains amid easing inflation.
  • Central banks signaled a pause in rate hikes, boosting investor confidence.
  • Economic growth forecasts remain modest, with potential headwinds from geopolitics.”
    Here, the prompt sets a role (“you are a financial advisor,” which can influence tone and jargon), specifies the format (3 bullet points), and tone (neutral). The result is concise and tailored. By contrast, if we hadn’t specified, the AI might have given a verbose paragraph with too much detail or an overly optimistic/pessimistic tone.

As you can see, how you ask matters a lot. Prompt engineering fundamentals include being clear about what you want (clarity), providing context (like specifying the role or audience), and sometimes giving examples. It’s a bit of an art and science: art in the creative way you might coax the best answer, and science in that there are some known techniques and best practices.

Important: Prompt engineering is iterative. Rarely will your first prompt be the perfect one (even AI experts try multiple phrasings). You might start broad, see the output, and then refine. For instance, you ask the model to draft a customer email – the first draft is too long. So you refine the prompt: “Keep it under 100 words.” Then maybe it’s concise but too generic – so you refine further: “Mention our 24/7 helpline in the email.” Through these iterations, you engineer the prompt to steer the AI closer to your desired output.

For non-coding professionals, think of prompt engineering like giving instructions to a very literal intern. The intern is super smart (it has read the entire internet), but you need to tell it exactly what output format and content you need. If you were delegating a memo to a junior employee, you’d specify the key points to cover and the style expected. Same with AI: if you need a summary, tell it how long, if you need an analysis, tell it what to focus on, if you need a draft letter, tell it the key messages to include.

A great starting tip is to state your intent or task clearly in the prompt’s first sentence. For example: “Summarize…”, “Draft an email…”, “Provide a list of…”, “Explain in simple terms…”, “Translate the following into Spanish…”. Models respond well to clear directives. Also, providing context like who the output is for or what your role is can influence the nuance of the answer. You might say, “As a risk analyst, list the top 5 risks in this scenario,” versus just “list top 5 risks” – the former may trigger the model to answer in a more analytical tone, because you set the persona.

Let’s do one more example tailored to financial services:

Prompt: “The following is a draft of an investment policy. Your task is to identify any unclear or potentially problematic statements. Respond with a numbered list of issues, and for each, suggest an improvement. The tone should be professional and helpful.”

Imagine we feed the AI a policy draft after this instruction. AI’s response might be:

“1. Issue: The policy says ‘invest in high-risk assets liberally’ which is vague and could conflict with risk management guidelines. Suggestion: Clarify the criteria for high-risk investments or rephrase to specify the allowable limit or conditions.
2. Issue: No definition is given for ‘long-term’ in the context of holding periods. Suggestion: Define what ‘long-term’ means (e.g., more than 5 years) to ensure consistency.
3. Issue: …”

This output is structured (numbered list), identifies problems, and suggests improvements – exactly what the prompt asked for. If the first attempt wasn’t great, we could refine further, but chances are a detailed prompt like that yields a useful result.

In summary, prompt engineering fundamentals are about being specific, providing context, and iterating. The beauty is that as a financial services executive, you don’t need to code to leverage AI – you just need to communicate your request clearly to the AI. And prompt engineering is the skill of doing just that. It’s almost like learning to ask questions in a new language – once you get a feel for it, you can unlock incredibly powerful outcomes from generative AI tools, whether it’s drafting reports, analyzing data, or creating new content. In the next section, we’ll introduce some key principles and techniques (like “zero-shot” vs “few-shot” prompting, using examples in prompts, etc.) to level up your prompt engineering game.

(Key point: Prompt engineering = writing effective instructions for AI. Always start with a clear request, include any necessary context (role, audience, format), and don’t be afraid to refine and try again. The examples above show how adding details or constraints to prompts leads to more useful outputs. It’s a hands-on skill – the more you practice with real AI models, the better you get at coaxing high-quality results.)

Section 8: Introduction to Prompt Engineering Principles

Now that we’ve seen the basics, let’s outline some core principles of prompt engineering, especially as they apply to financial contexts. Mastering these will help you communicate with generative AI tools more effectively and reliably.

Clarity is King

Be clear and explicit about what you want. Ambiguity is the enemy. If you need a summary in 100 words, say “100 words” (or a paragraph of a certain length). If you want a specific format (bullet points, table, step-by-step), mention that. In finance, where precision matters, clarity in prompts ensures the AI doesn’t guess or assume incorrectly. For example, instead of asking “How did our portfolio perform?”, a clearer prompt would be “Provide a performance summary of our portfolio for Q3 2023, including total return percentage and mention the best and worst performing asset classes.” This leaves little doubt about what’s expected.

Tip: You can even include what not to do in the prompt if relevant. E.g., “Draft a client-friendly explanation of VAR (Value at Risk) – do not use any equations or jargon, and keep it under 4 sentences.” The model then knows to avoid certain things.

Provide Structure in the Prompt

Models often mirror the structure you use in your prompt. If you ask a question in a long, run-on way, the answer might be rambly. Instead, you can structure your prompt to guide the answer’s structure. Using bullet points or numbers in your prompt can make the output more organized. For instance, try prompting:

“List the following as bullet points: reasons a bank would increase its loan loss reserves.”

Chances are, the AI will respond in bullet points because you signaled that structure. Similarly, starting a prompt with “1.” might lead it to enumerate answers (though some models do that by default for lists). In a financial report context, if you want an analysis broken into sections (e.g., Market Overview, Portfolio Impact, Recommendation), you can actually frame the prompt like:

“Generate an analysis with the following sections: Market Overview, Portfolio Impact, Recommendation. The content for each section should be based on the data below… [then maybe you provide data].”

This structured prompt will yield a nicely sectioned output.

Context and Roles

We touched on this: giving context about who is speaking or for whom the answer is can dramatically change the tone and detail. This is often called setting the “persona” or role. For instance, starting a prompt with “You are an experienced financial advisor…” can make the response more measured and authoritative. Alternatively, “Explain as if I’m a beginner” yields a simplified explanation. Use this to your advantage: for internal technical analysis, you might set a persona like “You are a risk model validation expert,” whereas for client communications you might say, “Respond in a friendly, reassuring tone as a customer service rep.”

Generative models are surprisingly good at adopting roles. Just by stating it, you shape the style. It’s almost like instructing an actor on which role to play. And if the response isn’t quite right, you can adjust (“be more formal,” “use a confident tone,” etc.). For financial services, maintaining an appropriate tone is crucial (e.g., empathetic in customer service, concise and factual in investor memos, cautionary in risk analysis), so don’t hesitate to specify that.

Iterative Refinement

Rarely is the first prompt the best. An important principle is iterative testing: try a prompt, review the output, then refine. Think of it as a dialogue. If the answer was off track, clarify in a follow-up prompt. For example, you might get an answer that’s correct but too verbose. Your next prompt can be, “Great, now make it more concise.” The model will then shorten its response. Or if it included something you didn’t want (say it gave an opinion but you wanted just factual), you can say, “Please provide only factual statements without personal opinions.” This back-and-forth is normal and expected. Even pros rarely hit the bullseye on the first try for complex prompts.

One approach is called prompt stepping: break a complex task into parts. For instance, if you want the AI to produce a detailed financial analysis, you might first prompt it to gather relevant points, then prompt it again to organize those points into a narrative. E.g.,

1. “List key economic events affecting Q3 2023 markets.” (AI lists events)
2. “Using those events, write a summary of how our balanced portfolio was impacted, in two paragraphs.”

This stepwise prompting can lead to more coherent outputs for complex tasks. It’s akin to guiding the model through your thought process.

Zero-shot vs Few-shot prompting

These terms sound technical but are simple concepts:

• Zero-shot prompting means you just ask the question or give the instruction without any examples. The model has to answer based solely on its general knowledge. For instance, “Explain compound interest in one sentence.” This is zero-shot (no examples given, just the request). Modern large models are very good at zero-shot for many tasks, especially with clear prompts.
• Few-shot prompting means you provide examples in your prompt to show the model what you expect. You essentially “prime” it with a couple of Q&A pairs or input-output examples, then ask it to do a similar task on new input. For example, a few-shot prompt for a chatbot might look like:
  • Prompt:
    Q: What is 2FA in banking?
    A: 2FA stands for two-factor authentication, a security process in which users provide two different authentication factors to verify themselves.
  • Q: How does a credit score affect a loan?
    A: A credit score is a numerical representation of a borrower’s creditworthiness. Lenders use it to decide loan approval and interest rates – higher scores generally lead to better loan terms.
  • Q: What is an APR?
    A: … (the model should continue in the same style).

In this example, by giving two Q&A pairs, we set the style and depth of answer. Few-shot is powerful if the model tends to give answers that are too shallow or too detailed – you can guide it by example. In finance, you might show it a formatted output example, like a sample of a properly formatted report section, so it imitates that format for your query.

Few-shot prompting basically leverages the model’s pattern recognition: it sees the pattern in the examples and continues it. Use cases: if you have a very specific style or format in mind and the model isn’t naturally producing it, show an example or two. Another case: when dealing with calculations or logic, giving a worked example can help (though these models aren’t always consistent in math).

However, note that few-shot prompts can be long (because you’re including examples), and some tools have input length limits. But within reason, it’s a great trick. For non-technical folks: it’s like saying, “Here’s an example of a good answer, now please answer my actual question in the same way.” Humans do this when learning as well, so it’s intuitive.

Best Practices Summary

1. Be Specific – Specify the task, format, length, audience, etc. Don’t assume the AI will fill in details correctly; spell them out.
2. Use Constraints – If you need the output in 3 bullet points or you need it to include a particular term, say so explicitly. AI will follow instructions like “include the word ‘annualized’ in the answer” or “output should be a JSON object” (for the more technical use cases).
3. Avoid Ambiguous References – For example, if your prompt says “Review this report,” make sure the AI knows what “this” refers to (if you’ve provided the report text, say “the report text below”). Models don’t have persistent memory of your files unless you include them in the prompt or they were trained on them in general knowledge.
4. Quality over Politeness – Unlike humans, AI doesn’t require polite fluff. You can drop words like “please” without offending it (though including them doesn’t hurt either). More importantly, extra polite phrasing can sometimes confuse a model if it obscures the instruction. “Could you maybe sort of give me an answer about X, please?” is less clear than “Provide an answer about X.” Feel free to be direct.
5. Iterate and Refine – We’ve emphasized this: treat the AI as a collaborator. If the output isn’t right, refine the prompt and try again. You can even feed back to the model its own output with instructions. For example: “Here is your last answer. Now, based on that, do XYZ.” This iterative approach often converges on a great answer.

Prompt engineering in financial services might also involve adding relevant context that the model wouldn’t otherwise know. For instance, if you’re asking it to draft a letter to clients about a merger, you might first feed it a summary of the merger details, then ask it to draft the letter. The principle here: supply the right information in your prompt if the AI might not have it. (Models like ChatGPT have cut-off knowledge in 2021 for GPT-3.5, and even for updated ones, they might not know specifics of your company – you can include those specifics in the prompt.)

A quick example of context inclusion: “Our bank, ACME Bank, is merging with XYZ Bank. Include in the letter that the merger will bring more ATM locations and improved mobile app features. Now, draft a letter to customers explaining the merger benefits.” The output will then incorporate those facts because you provided them.

By following these principles, you’ll get far better results from AI tools. You don’t have to memorize them all – the key is to remember it’s an interactive process and that you have control over guiding the AI. As you practice, you’ll develop an intuition for what phrasing gets the best outcome.

(Principles recap: Clarity – be explicit about what you want; Structure – guide format with your prompt style; Context – specify roles/audience for tone and correctness; Iterate – refine prompts based on output; Zero-shot vs Few-shot – know when to give examples to steer the model. By applying these, your interactions with GenAI will be more productive and reliable. Next, we’ll apply these specifically to financial tasks with some prompt examples.)

Section 9: Financial Services-Specific Prompting Techniques

Let’s tailor our prompt engineering know-how to common financial services scenarios. Different domains have their own lingo and requirements, and prompts can be adapted accordingly. We’ll walk through a few examples of prompt techniques for finance tasks – including policy summarization, document review, and data interpretation – to illustrate how to get the most out of GenAI in these contexts.

Example 1: Summarizing Policies or Reports

Use case: You have a lengthy policy document or market research report and you want a concise summary. This is common for compliance policies, research notes, or lengthy emails.

Prompt strategy: Clearly instruct for a summary and specify the focus and format. Financial documents often have a lot of detail, so decide what’s most important: key points, action items, compliance requirements, etc., and mention those. Also, set a length (like a few bullet points or a short paragraph) to avoid the model giving back something nearly as long as the original!

Example Prompt:
“Summarize the following investment policy in 5 bullet points, focusing on the compliance requirements and any limitations on asset allocation. Use clear, plain language (imagine explaining to a new analyst).”
(Then you would include the text of the policy or at least the relevant sections.)

Why this works: We’ve told the model exactly: give 5 bullet points, and specifically to concentrate on compliance rules and allocation limits – which are likely the crucial parts of an investment policy. We also specified audience (“new analyst” implies we want straightforward language, not heavy legalese). The model will parse the policy and aim to extract those points. For instance, the output might be:

• “Must maintain a diversified portfolio with no more than 10% in any single stock (compliance requirement).
• Cannot invest in commodities or derivatives for this client portfolio (prohibited asset classes).
• Requires quarterly reporting to the compliance committee on portfolio risk metrics.
• …etc.”

This type of summarized output is extremely useful and saves a ton of reading time.

If the first attempt misses something important (say it didn’t mention an allocation limit that you know is in the policy), you can refine: “Include the portfolio allocation limit for equities stated in the policy.” The model will then add that if it’s present in the text.

Tip: For regulatory or legal docs, consider adding “in bullet points” or “in plain English” to ensure clarity. Also, if you suspect the model might hallucinate or guess, you can add “If information is not in the text, do not speculate.” This keeps it anchored to the content.

Example 2: Document or Contract Review with AI

Use case: You want AI to review a document (like a contract, term sheet, or lengthy email thread) and pull out specific items – e.g., key terms, potential risks, or required actions. This isn’t just summarizing; it’s identifying particular content.

Prompt strategy: Ask for extraction of details. For example, in a contract, you might ask for a list of obligations of each party, or any mention of fees or dates. The prompt should direct the AI exactly what to look for.

Example Prompt:
“You are an assistant helping to review a loan agreement. Read the agreement text and extract the key financial terms: loan amount, interest rate, maturity date, collateral, and any covenants. Present them in a labeled list. If a term is not found, state ‘Not specified.’”
(Followed by the contract text.)

Why this works: We gave the AI a clear mission: find these specific terms. By labeling what we want, the output will likely be something like:

• Loan Amount: $5,000,000
• Interest Rate: 6.5% per annum (fixed)
• Maturity Date: Dec 31, 2028
• Collateral: 10,000 shares of Company X stock
• Covenants: Borrower must maintain Debt/EBITDA below 3.0; No additional senior debt issuance without consent.

This is extremely helpful for quickly gleaning contract details. Essentially, the AI did a pseudo-OCR and analysis in one step. Note we told it how to handle missing info (“Not specified”), preventing it from making something up if, say, collateral wasn’t mentioned.

For a long email chain scenario, you might prompt: “Summarize the attached email thread. List any questions that were asked and whether they were answered, and highlight any agreed-upon next steps.” The AI would comb through and possibly output:

• Question: Can we extend the deadline to Nov 15? – Answer: Yes, agreed by John in email 3.
• Question: Will budget increase by 10%? – Answer: Not directly answered in the thread.
• Next Steps: Jane will draft the updated contract; team to meet on Oct 30 to review changes.

By structuring the prompt, we got an organized output focusing on questions and next steps, which is often what a manager wants to know without reading every email.

Example 3: Data-to-Insight Transformation

Use case: You have raw financial data or metrics and want an insight or narrative. For example, turning a table of quarterly results into an explanation of trends, or interpreting financial ratios. Generative AI can analyze and articulate patterns – a huge time-saver for report writing.

Prompt strategy: Provide the data in the prompt (if it’s small enough) or at least summarize it, then ask for analysis or commentary. Be specific on what kind of insight you want: comparison to past, identification of anomalies, suggestion of causes, etc. Also specify format if needed (paragraph vs bullet insight, etc.)

Example Prompt:
“Our company’s quarterly revenue for the last 4 quarters were: Q1 $10M, Q2 $12M, Q3 $9M, Q4 $15M. Operating profit margins for those quarters were: 20%, 22%, 18%, 25%. Analyze this data and provide two insights: one about revenue trend and one about profit margin trend, in 2-3 sentences each.”

Why this works: The AI sees the numbers and will likely notice: revenue dipped in Q3 then spiked in Q4, margins followed a similar pattern (down then up). The prompt asks for exactly two insights, guiding the output. For example, it might say:

1. Revenue Trend: “Revenue shows a significant dip in Q3 ($9M down from $12M in Q2) followed by a strong rebound in Q4 to $15M. This suggests a one-time issue in Q3 or seasonality, with Q4 more than making up for the earlier shortfall.”
2. Margin Trend: “Operating profit margins dropped in Q3 (18% from 22% in Q2) then hit their highest in Q4 at 25%. This mirrors the revenue recovery, indicating improved cost efficiency or higher operating leverage in Q4 contributing to better margins.”

This kind of analysis is what an analyst might write, and the AI did it in seconds. Notice how the prompt explicitly asked for two insights, preventing the model from writing a long essay or going off-topic.

For more complex data, you might need to pre-process or summarize it for the model if it’s large. But an interesting technique is you can feed it CSV-like data or JSON and prompt it to analyze. For instance:

“Data: Product A sales: 100, 120, 130 (Q1-Q3); Product B sales: 80, 85, 90. Task: Compare the sales growth of Product A and Product B and indicate which is growing faster in percentage terms.”

It should compute or at least correctly deduce that Product A grew 30% from Q1 to Q3 (100 to 130), Product B grew ~12.5% (80 to 90), and then respond accordingly (“Product A’s sales growth (~30%) outpaced Product B’s (~13%) over the first three quarters, indicating much faster expansion for A”).

However, caution: models can make arithmetic mistakes, especially older ones. For reliable calculations, it’s often better to provide the result of calculation or cross-verify. But for qualitative insight like “faster” or “slower,” they usually get it right if numbers are clear.

Example 4: Compliance and Regulatory QA
A specialized scenario in financial services is interpreting regulations or policies. For instance, you might want to ask, “According to Policy X, are employees allowed to trade stocks of clients?” If you have the policy text, you can prompt the AI:

“Based on the following excerpt from our Code of Conduct, answer the question: Can employees trade stocks of client companies? Excerpt: ‘…Employees are prohibited from using non-public information for trading. Trading of securities of clients is subject to pre-clearance and must not conflict with insider trading laws…’”

The AI might answer: “Employees are not outright banned from trading client company stocks, but they must get pre-clearance and cannot trade based on any confidential information. Essentially, they can trade only if it doesn’t conflict with insider trading rules and they follow approval procedures.”

This shows how to focus the AI on a specific rule. By asking a pointed question and giving the relevant text, you get a clear interpretation. Without prompt guidance, the AI might waffle or talk generally about insider trading, but the way we framed it yields a direct answer.

General Prompting Tips for Finance:

• If asking for definitions or explanations of finance terms, you can specify the depth: “explain like I’m 5” vs “provide a technical definition as per IFRS.”
• For strategy or advisory outputs (like “What should our bank consider when adopting AI?”), you can ask for pros/cons or a SWOT analysis format. The model can produce structured strategic thinking, which can be a great starting point for planning.
• When generating content like emails, reports, or presentations, always specify the key points or data that must be included. E.g., “Draft a client letter about our new savings product (interest rate 2.5%, no monthly fees, FDIC insured). Emphasize the high interest rate and no fees in a friendly tone.” The specifics ensure the letter contains the correct facts.
• Sanity-check numeric outputs: If you ask the AI to calculate or use numbers, double-check. It’s good at logic but not infallible with math.

Finally, remember to keep sensitive data considerations in mind. If using a public AI service, avoid putting truly confidential details verbatim (paraphrase or use placeholders that you later replace). There are enterprise AI solutions that keep data in-house which are safer for that.

The techniques above demonstrate that with the right prompts, generative AI can become a versatile assistant for finance professionals: summarizing lengthy documents, extracting critical info, analyzing data for insights, and answering domain-specific questions. By crafting your prompts thoughtfully, you essentially program the AI on the fly to perform tasks that would normally eat up a lot of human time.

(Finance prompting techniques recap: For summaries, specify focus and format (e.g., bullet points, key requirements). For document/contract review, instruct the AI on exactly which fields or issues to extract. For data-driven insights, feed the data and ask for specific comparative or trend analysis, or explanations of drivers. For policy Q&A, supply relevant excerpt and ask direct questions. In all cases, clarity and context yield the best results – guiding AI to be a powerful ally in financial analysis and communication.)

Section 10: Hands-on Workshop – Creating Your First Financial Services Prompts

It’s time to roll up our sleeves and practice. In this hands-on workshop, we’ll go through three exercises that let you craft prompts and see the kind of output GenAI can produce for typical financial services tasks. These exercises are designed to be done with any generative AI tool (like ChatGPT, an internal LLM-based assistant, etc.). Try these prompts out and then fine-tune them based on the results – that iterative practice is how you build skill.

Exercise 1: Information Extraction from Financial Documents

Scenario: You have a snippet of a financial document – say, a bank statement or a transaction log – and you want to extract structured information from it.

Task: Write a prompt that gets the AI to extract specific data points.

Example Document Snippet: (for the sake of exercise, we include it)
“Date: 2025-01-15; Description: Wire Transfer from ACME Corp; Amount: +$50,000
Date: 2025-01-20; Description: ACH Payment – Invoice 12345; Amount: -$8,000
Date: 2025-01-22; Description: ATM Withdrawal; Amount: -$500”

Your Prompt: You might craft something like:
“Extract all transactions from the following bank statement snippet and present them as a JSON array of objects, with fields: date, description, amount (as number, positive for credit, negative for debit). If description contains an invoice number, include a field ‘invoice_number’.

Text: <document text>” (where <document text> is the snippet above).

Expected Output: The AI should output something like:

[
  {
    "date": "2025-01-15",
    "description": "Wire Transfer from ACME Corp",
    "amount": 50000
  },
  {
    "date": "2025-01-20",
    "description": "ACH Payment – Invoice 12345",
    "amount": -8000,
    "invoice_number": "12345"
  },
  {
    "date": "2025-01-22",
    "description": "ATM Withdrawal",
    "amount": -500
  }
]

This exercise flexes your ability to specify an output format (JSON) and conditional extraction (invoice number if present). Once you run it, examine the output. Did it capture everything correctly? If the first run missed something (maybe it didn’t catch the invoice number because of formatting), adjust your prompt or formatting of input and try again. In a real scenario, you could do this with any structured output you need – CSV, bullet points, etc.

Discussion: This simulates an AI reading a statement and outputting data in a structured way, which could then be fed into another system. It shows how AI can bridge unstructured and structured data. Think about other documents you could do this for (e.g., parsing an email for key fields, extracting terms from a term sheet as we did earlier). Share with the group (if doing a workshop together) what prompt you wrote and how the output looked. Did adding an example in the prompt help? Did you need to tweak any wording?

Exercise 2: Generating Compliance-Focused Summaries

Scenario: You have a chunk of a regulatory document or lengthy compliance guideline. You need to summarize it for a meeting, focusing on action items the bank needs to take to comply.

Task: Write a prompt that produces a concise summary highlighting obligations or actions required by the regulation.

Example Excerpt: (Hypothetical regulation text)
“According to Section 4.3 of the Fair Lending Act 2025 amendments, financial institutions must implement quarterly training for all lending officers on anti-discrimination practices. Institutions are also required to submit an annual fair lending report to the regulator detailing outreach programs for underserved communities and any corrective actions taken in response to fair lending complaints.”

Your Prompt:
“Summarize the following regulatory excerpt into a brief paragraph of compliance requirements for our bank. Emphasize what actions we must take (e.g., training, reporting). Use clear bullet points for each required action:

Excerpt: “””

Expected Output: Possibly:

• “Conduct quarterly anti-discrimination training for all lending officers (mandated by Sec 4.3).
• Prepare and submit an annual Fair Lending report to regulators, including outreach efforts to underserved communities.
• Document any fair lending complaints and corrective actions as part of this report.”

The output highlights specific actions (training, reporting, documentation of complaints). That’s exactly what you as an executive need to know. If the output came as a paragraph, you might iterate: “Please use bullet points.” If it missed an item, you’d refine accordingly.

Discussion: In practice, you might plug in a section of a law or internal policy and get an easy checklist of to-dos. That’s gold for compliance meetings. Try variations: maybe ask for it as a “checklist” or “executive summary for compliance team.” You’ll notice tone and detail level shift. Share how adding certain keywords (“checklist”, “executive summary”) changed the response. This exercise trains you to pull out the operational essence of legalese with AI’s help.

Exercise 3: Transforming Financial Data into Insights

Scenario: You have some raw data or metrics (could be portfolio returns, sales figures, etc.), and you want an insightful narrative or recommendation drawn from it. This is like getting an analyst’s commentary on data.

Task: Provide the data and ask for an analysis or recommendation.

Example Data:
“Fund A returned 8% YTD, Fund B returned 5% YTD. Fund A’s volatility (std dev) was 12%, Fund B’s was 6%. Both funds have similar asset allocations (60% equity, 40% bonds).”

Your Prompt:
“Given the performance and volatility data below for Fund A and Fund B, advise which fund performed better risk-adjusted, and which you’d recommend for a conservative investor, and explain why in 2-3 sentences.

Data: Fund A +8% YTD, volatility 12%; Fund B +5% YTD, volatility 6%. Allocation: both ~60/40 equity/bonds.”

Expected Output: The AI should reason that Fund A had higher return but also higher volatility, while Fund B had lower return and lower volatility. It might calculate or intuit that risk-adjusted (maybe via Sharpe-like thinking), Fund A’s return per unit volatility is roughly 0.67, Fund B’s is ~0.83, thus Fund B is actually better risk-adjusted. And for a conservative investor, likely Fund B is preferable due to lower volatility.

So an answer might be:
“On a risk-adjusted basis, Fund B performed better – its return relative to its volatility is higher (5% with half the volatility of Fund A). For a conservative investor, Fund B is recommended because it exhibits much lower volatility, even though its raw return is lower; the steadier performance aligns with a conservative risk profile.”

This demonstrates how the AI can make a comparative judgment and justification. If it just stated facts without the insight, you might tweak: “explicitly state which has better return per unit of risk.” If it gave too much detail, you could say “keep it brief.”

Discussion: Did the AI correctly interpret “risk-adjusted” and the conservative angle? These models generally understand such concepts, but if it didn’t, maybe you provide the hint of calculating return/volatility. The beauty here is you got a bite-sized advisory insight – something an investment analyst might produce – in seconds. Imagine feeding it more data points or more funds; it can still follow the logic if well prompted.

During a workshop, have participants try different data sets (one might do revenue growth vs profit margin trade-off, another does credit score vs default rate example, etc.) and see how the AI handles different scenarios. This solidifies that formulating the question clearly yields useful analytic commentary.

---

Wrap-up: You’ve now written prompts for extraction, summarization, and analysis – three common uses. In each case, review the outputs. Are they accurate? If something was off or not quite how you want, this is where iterative refinement (from Section 8) comes in: adjust the wording and try again. Prompt engineering is like a dialogue: use the output as feedback to refine your next prompt.

As a final part of the workshop, consider these reflection prompts:

• Which prompt that you wrote gave the best result on first try, and why do you think that is? (Often, it’s the one where you were most specific.)
• Which prompt required the most tweaking? What did you learn in the process? (E.g., maybe you learned the AI misunderstood a term until you clarified it.)
• Can you think of a workflow in your day job where a similar prompt could save you time? Perhaps preparing a draft of a client email, or analyzing a set of monthly KPIs.

The more you practice, the more prompt patterns you’ll recognize. For instance, you might discover that starting with “Act as an X and do Y” works well for certain outputs (like “Act as a CFO and assess these financial ratios…”). Or that giving one example in the prompt drastically improved the output’s format. These are your personal prompt engineering heuristics.

By completing these exercises, you’ve taken theoretical knowledge and applied it. You should feel more comfortable with guiding AI models. Remember, you don’t have to get it perfect on the first go – use the AI as an interactive partner. Unlike delegating to a human, you can keep refining your instructions without anyone getting frustrated!

Conclusion of Workshop: With these hands-on trials, you’ve seen firsthand how a well-crafted prompt leads to useful results in seconds, whether it’s pulling insights from dense text or turning raw numbers into narrative. The key takeaway is that you are in control of the output – the AI will do exactly (well, usually) what you ask, so learning to ask in the “AI’s language” is the secret sauce. Consider integrating these techniques into your routine tasks and continue to experiment. As generative AI becomes more integrated in financial tools, your prompt engineering skills will be an asset, enabling you and your team to leverage AI effectively and responsibly.

(In summary, the exercises had us: extract data from a document via a structured prompt, summarize a compliance excerpt focusing on required actions, and analyze performance data to make a recommendation. By trying these, you practice how to instruct AI for practical finance tasks. Continue exploring with your own examples – the more you practice writing and refining prompts, the more value you’ll get from generative AI in your work.))