AI in Financial Services: Capabilities and Limitations

The $40 Million Marketing Claim

In March 2024, the SEC charged two investment advisers — Delphia (USA) Inc. and Global Predictions Inc. — for making false and misleading statements about their use of artificial intelligence. Delphia claimed its AI could "predict" which companies were about to grow using "machine learning combined with social media." Global Predictions told clients its platform provided the "first regulated AI financial advisor." Neither firm could substantiate its claims. The SEC fined them a combined $400,000.

These were not obscure bucket shops. They were registered investment advisers marketing to real clients. And the enforcement action signaled something important: the SEC will scrutinize AI claims the same way it scrutinizes any other performance claim. The hype is not harmless. It creates liability.

This matters to you because AI is not optional anymore. Your competitors use it. Your clients ask about it. Your broker-dealer or carrier may require it. But the gap between what AI marketing promises and what AI technology delivers is wide enough to build an entire enforcement docket inside. Understanding that gap — precisely, technically, without either dismissing AI or overselling it — is now a core professional competency.

What AI Actually Means in Financial Services

The term "artificial intelligence" covers a spectrum of technologies, and conflating them creates real compliance risk. When your client asks whether you "use AI," your answer should depend on what kind.

Rule-based automation is the oldest layer. These are systems that follow explicit if-then logic: if a client's income exceeds a threshold, flag the application for additional review. If a portfolio drifts more than 5% from target allocation, trigger a rebalance. This is not intelligence in any meaningful sense — it is software doing exactly what a programmer told it to do. But vendors frequently label it "AI" in marketing materials.

Machine learning (ML) is the layer that changed the game. ML systems learn patterns from data without being explicitly programmed for each scenario. A machine learning model trained on ten years of claims data can identify patterns that predict which auto insurance policies are likely to generate claims — patterns no human underwriter would spot in a spreadsheet. The model does not understand cars or accidents. It finds statistical correlations in data. This distinction matters enormously for compliance: the model can be right for the wrong reasons, or right for reasons that encode prohibited discrimination.

Large language models (LLMs) and generative AI — the ChatGPT layer — are what most clients mean when they say "AI" today. These models generate human-like text by predicting the next word in a sequence, drawing on patterns learned from vast training datasets. They can draft client emails, summarize research, generate meeting notes, and answer questions about tax law or insurance policy provisions. They can also fabricate case citations, invent regulatory guidance that does not exist, and produce confidently wrong answers indistinguishable from correct ones.

Where AI Delivers Real Value Today

Financial professionals who use AI effectively tend to deploy it in four areas.

Document processing and data extraction. AI can read a 200-page annuity contract and extract the surrender schedule, death benefit provisions, and fee structure in seconds. It can scan a client's tax return and populate a financial planning intake form. This is high-volume, low-judgment work where AI excels because the answers exist explicitly in the source documents.

Research acceleration. An advisor researching the tax treatment of a Roth conversion ladder can use AI to quickly survey IRS guidance, Revenue Rulings, and Tax Court decisions. The AI does not replace the advisor's judgment about which authority controls — but it compresses a two-hour research task into fifteen minutes. The SEC and FINRA have not prohibited this use, but both have emphasized that the professional remains responsible for the accuracy of any advice derived from AI-assisted research.

Client communication drafting. AI can generate first drafts of annual review letters, policy renewal notices, and financial plan summaries. Lemonade, the insurtech carrier, uses AI to handle routine customer communications including claims status updates and policy change confirmations. For independent agents and advisors, generative AI can produce a draft quarterly newsletter in minutes rather than hours.

Pattern detection in compliance. Large broker-dealers and insurance carriers use ML to flag potentially suspicious transactions for anti-money laundering review, identify complaint patterns across branch offices, and detect advertising claims that may violate FINRA Rule 2210 or state insurance advertising regulations. These systems do not make compliance decisions — they surface items for human review.

Where AI Fails — and Why It Matters

The failures are not bugs. They are structural features of how current AI works, and they have direct regulatory consequences.

Hallucination. Generative AI models produce plausible-sounding text that is factually wrong. In June 2023, an attorney in Mata v. Avianca submitted a brief containing six case citations fabricated by ChatGPT. The cases did not exist. The court sanctioned the attorney. For financial professionals, the equivalent risk is an AI tool that cites a nonexistent IRS Revenue Ruling, invents a state insurance regulation, or fabricates the terms of a financial product. You are liable for what you tell clients, regardless of where you got the information.

Opacity. Most ML models — particularly neural networks and ensemble methods — cannot explain why they reached a particular conclusion. If your underwriting system denies an application, the applicant (and their state insurance department) may be entitled to an explanation. "The algorithm said so" is not an explanation. The NAIC's Model Bulletin on the Use of Artificial Intelligence Systems by Insurers (December 2023) specifically addresses this: insurers must be able to explain adverse decisions to consumers, even when AI was involved in making them.

Data dependency and bias. AI models are only as good as their training data. If historical lending data reflects decades of redlining, an ML model trained on that data will learn to replicate those patterns. The CFPB has warned that AI-driven underwriting that produces disparate impact on protected classes violates the Equal Credit Opportunity Act regardless of the model's intent. In insurance, similar concerns arise under state unfair trade practices statutes and the NAIC's Unfair Discrimination principles.

Temporal limitations. Most LLMs have a knowledge cutoff date. A model trained through April 2024 does not know about regulatory guidance issued in July 2024. It does not know that a tax provision expired, a state updated its insurance code, or a court overturned precedent. For professionals who must provide current advice, this is not a minor limitation — it is a fundamental constraint.

The Compliance Line: AI-Assisted vs. AI-Dependent

The most important distinction for regulatory purposes is between AI-assisted decision-making and AI-dependent decision-making.

AI-assisted means the professional uses AI as a tool — to gather information, draft documents, or identify patterns — but exercises independent judgment before acting. The professional reviews the AI output, verifies key facts, and takes responsibility for the final work product. This is the model regulators have generally accepted.

AI-dependent means the professional delegates the decision itself to the AI. The model recommends a portfolio allocation, and the advisor implements it without independent analysis. The algorithm approves an insurance application, and the underwriter rubber-stamps it. This is where regulators draw the line. FINRA Regulatory Notice 21-06 emphasized that firms cannot "outsource their compliance obligations" through technology. The SEC's examination priorities for 2024 and 2025 specifically flag "AI-related advisory services" for enhanced scrutiny.

The practical test is simple: if the AI disappeared tomorrow, could you explain and defend every recommendation you made to a client? If the answer is no, you have crossed the line from AI-assisted to AI-dependent — and you have a compliance problem.

What This Means for Your Practice

AI is a tool with genuine utility and genuine risk. The professionals who will thrive are those who understand both. The ones who will face enforcement actions are those who either ignore AI entirely — ceding ground to competitors who use it effectively — or adopt it uncritically, treating AI output as authoritative without the verification that professional responsibility demands.

The remaining units in this course will examine specific applications: robo-advisors and algorithmic underwriting (Unit 2), AI for compliance and client communication (Unit 3), the evolving regulatory response from the SEC, NAIC, and state regulators (Unit 4), ethical considerations including bias and fiduciary duty (Unit 5), and strategic adoption frameworks for your practice (Unit 6). Each unit builds on the foundation established here: AI is powerful, AI is limited, and the professional using it bears full responsibility for the outcome.