LOOKING BACK | What Is AI’s Role in Financial Services?

For now, AI’s most successful financial applications are still relatively modest. The technology searches documents, summarizes meetings, writes first drafts, prepares research, reviews transactions, assists programmers and brings relevant information to the attention of employees. It functions as an efficiency aide—a highly capable assistant sitting between a financial professional and the extraordinary trove of data maintained by banks, insurers, asset managers, accounting departments and advisory firms. 

The efficiencies can be substantial. But an assistant that merely organizes information is different from an agent that independently completes a workflow, communicates with customers, moves money or makes a recommendation that affects someone’s financial life. As financial institutions advance from experimentation to deployment, they are confronting a problem that technology alone cannot resolve: deciding where automation should end and human authority should begin. 

Regulators, central bankers and industry leaders increasingly agree that AI should augment human judgment rather than obscure responsibility. Yet no institution has produced a universally accepted division of labor between people and machines. The financial industry appears to know what it wants AI to accomplish—lower costs, faster service, stronger risk detection and more personalized products—but remains much less certain about what human beings will be expected to do afterward. 

AI’s Present Role: An Assistant Surrounded by Data 

At this point, we’ve all heard that financial services is especially well suited to AI because the industry generates, purchases and stores enormous quantities of information. 

A bank may possess years of transaction histories, credit records, correspondence, call-center transcripts and regulatory documents. An asset manager has market prices, portfolio positions, company filings, analyst research and earnings-call transcripts. An insurer has claims files, policy language, medical or property records and fraud indicators. A wealth manager may hold household balance sheets, financial plans, meeting notes, tax information and records of every interaction with a client. 

Earlier generations of automation worked best when this information was structured and a process followed predictable rules. Generative AI expands the addressable workload because it can interpret natural language, retrieve material from unstructured documents and present an answer conversationally. Agentic AI goes further by using that information to carry out multistep tasks. 

The most common uses remain internal and assistive. Federal Reserve Governor Lisa Cook said in a May 27 speech that financial institutions were initially applying the current generation of AI to manual and resource-intensive activities such as compliance, call centers and back-office operations. She also pointed to faster analytics, software development and the integration of legacy systems as areas where AI could improve performance. 

That pattern is visible across accounting, banking and financial advice. Advisers use AI to prepare meeting summaries and emails, brainstorm client communications and support marketing. Earlier industry research cited by PLANADVISER found that advisers most often used generative AI for client engagement, marketing and internal productivity rather than autonomous investment management. 

By July, adoption was beginning to produce measurable optimism. Research reported by PLANADVISER on July 9 found that 82% of surveyed advisers were already using AI tools and 69% believed AI had positively affected the industry. The findings suggest that AI is becoming less of an experimental novelty and more of a regular component of advisory work. 

The practical appeal is straightforward. An adviser who spends less time taking notes, searching account records and drafting routine correspondence can theoretically spend more time talking to clients. An anti-money-laundering analyst who receives a consolidated file rather than manually opening several systems can concentrate on determining whether a transaction is actually suspicious. A finance department that automates data retrieval can devote more attention to forecasting and interpretation. 

AI’s early value therefore comes not only from doing work faster but from reorganizing professional attention. 

Efficiency Gains Are Real, but Transformation Is Uneven 

The financial industry is reporting tangible productivity improvements, although the gains are generally concentrated in defined processes rather than across entire institutions. 

Banks are using AI to accelerate fraud investigations, customer-service responses, compliance reviews and software development. Investment firms are applying it to research and product explanations. Finance departments use it to prepare variance analyses, assemble reports and answer questions about internal data. Insurers are experimenting with automated underwriting, claims intake and document review. 

The results can be dramatic when the task is sufficiently narrow and the underlying data is accessible. AI can reduce the amount of time spent collecting and organizing information from hours to minutes. The employee is then left with the portion of the process that requires interpretation, escalation or communication. 

But the industry’s growing collection of successful use cases can exaggerate how far enterprise transformation has progressed. 

An MIT Sloan Management Review article published June 2 argued that AI had not yet transformed finance because many organizations continued to use it within traditional operating models. Adding a new tool to a process without redesigning the process can produce incremental savings while leaving the department’s role, incentives and decision structure essentially unchanged. 

McKinsey made a similar distinction in its May 28 examination of AI and banking. The technology could do more than cut costs: it could redistribute profits, alter customer relationships and intensify competition. But realizing that potential would require banks to move beyond scattered pilots and rethink their broader strategies. 

The obstacle is frequently organizational rather than technical. Data is fragmented among business units. Employees do not trust the models. Compliance teams are brought into projects too late. New tools are layered on top of old workflows. Firms launch copilots without clarifying when employees should use them or how performance will be measured. 

This helps explain why financial institutions can report hundreds of individual productivity gains while still struggling to identify a comparable improvement in overall institutional performance. 

The value also may not accrue evenly. Online discussions among investors have increasingly questioned who captures the profits created by AI. Banks may save labor, technology vendors may charge for infrastructure and models, customers may receive faster service, and employees may face greater workload expectations. Efficiency does not necessarily determine how the economic benefit is distributed. 

In trading, a similar problem arises. Even where AI provides a temporary informational edge, widespread adoption can erode it. If many market participants purchase similar models and use comparable data, the advantage may migrate to the firms controlling proprietary information, distribution, computing infrastructure or execution rather than to the ordinary trader using a commercial AI tool. 

Has Anyone Found the Right Human–Technology Balance? 

Financial institutions often describe the preferred relationship as “human in the loop.” But the phrase can conceal more than it clarifies. 

A meaningful human role could involve independently evaluating a model’s recommendation, examining competing evidence and taking responsibility for the decision. A nominal human role could consist of approving an AI-generated conclusion because reviewing the underlying reasoning would take too much time. Both arrangements may be described as human oversight, though only one represents substantive judgment. 

The strongest examples of balance are found where AI handles preparation while a qualified professional retains decision-making authority. 

In wealth management, AI can summarize household information, identify changes since the last meeting, prepare educational material and suggest follow-up tasks. The adviser remains responsible for determining whether the recommendation fits the client’s goals, explaining uncertainty and helping the client behave rationally when markets fall. 

Recent surveys suggest clients are relatively comfortable with this supporting role. Janus Henderson research reported before the core period found that 87% of investors would feel either good or neutral about an adviser using AI to create educational resources. That acceptance does not necessarily extend to an AI system independently telling a household when to retire, how much risk to assume or whether to sell during a crisis. 

In insurance, AI can review documents and estimate the probability that a claim requires closer inspection. A human adjuster can investigate disputed facts, apply judgment and communicate with a policyholder. In lending, AI can assemble and analyze credit information while a lender handles exceptions and explains the result. In compliance, AI can prioritize alerts while investigators determine whether the facts warrant regulatory action. 

The common feature is not merely that a person appears somewhere in the process. It is that human involvement increases with the consequence, ambiguity and irreversibility of the decision. 

Yet some recent events illustrate the danger of allowing efficiency to overwhelm the human relationship. Reports of customer and agent dissatisfaction surrounding State Farm’s AI-oriented operational changes demonstrated that automation can degrade service when people perceive it as an obstacle rather than an aid. 

Financial institutions may therefore need a more precise hierarchy. Low-risk, reversible administrative tasks can be automated extensively. Material but reviewable decisions require active supervision. Fiduciary, contested or systemically important decisions require identifiable human authority and an effective means of appeal. 

No company has conclusively solved this problem across all of its operations. The balance is not a fixed percentage of human and machine participation. It changes according to the stakes. 

Regulators Are Defining Principles, Not a Final Answer 

Government officials and international organizations are beginning to describe what responsible financial AI should look like. Their guidance is converging around governance, accountability, resilience and proportionality, but it has not settled the larger philosophical question of AI’s proper role. 

Cook’s May 27 speech offered one of the clearest early frameworks. She described AI as a potential source of economic growth and financial innovation but warned that innovation could amplify existing vulnerabilities if it was not monitored. AI might improve credit access, capital allocation, financial products and risk detection, while also contributing to correlated trading, cyber threats, concentration and market instability. 

The Federal Reserve’s own internal use of AI provided a useful model. Cook described applications that helped researchers classify text, analyze information and run financial-stability scenarios. At the same time, she emphasized that the technology was not being used to set monetary policy. AI expanded the evidence available to decision-makers without being given responsibility for the decision itself. 

At Stanford’s SIEPR Policy Forum, participants similarly called for greater coordination among financial regulators and a stronger model-risk framework. The discussion highlighted how AI, stablecoins, private credit and changing market structure could interact, rather than treating each technology as a separate regulatory problem. 

The World Economic Forum’s June 24 AI Playbook for Financial Services framed the central challenge as integrating AI with both urgency and discipline. It called on institutions moving from experimentation toward scaled deployment to strengthen governance and collaborate with regulators and policymakers on shared risks. 

The Financial Stability Board also entered the discussion. Federal Reserve Vice Chair for Supervision Michelle Bowman said in June that the FSB was preparing sound practices for financial institutions’ use of AI. In July, Bowman again addressed sound AI practices, reinforcing the regulatory emphasis on risk management rather than prescribing particular business models. 

The emerging regulatory position can be summarized as follows: institutions may use AI, but they remain responsible for the outcome. Existing expectations around safety, fairness, consumer protection, cybersecurity and risk management do not disappear merely because a model performed the work. 

That is a sensible starting point, but it does not answer every question. 

Should an AI be permitted to provide personalized financial recommendations? May it execute trades without contemporaneous human approval? Can an autonomous agent switch financial providers on a customer’s behalf? Who owes the duty of care when advice is assembled by a foundation-model developer, customized by a software provider and distributed through a bank? 

The urgency of these questions became clearer in July as consumer use of general-purpose AI for personal-finance advice attracted more attention. PYMNTS noted that an AI system has no fiduciary duty, professional indemnity coverage or inherent obligation to optimize a client’s actual financial outcome. A model is designed to produce a plausible response; a licensed adviser is legally accountable to the client. 

Regulators understand the outcomes they want: accountability, transparency and protection from harm. They have not yet established how those principles should apply when an AI agent behaves less like software and more like a financial intermediary. 

From Copilot to Operator 

The next phase of financial AI will be defined by the movement from assistance to execution. 

McKinsey described this transition in June as the point when AI begins doing the work rather than merely helping an employee do it. Agentic systems can complete multistep processes, interact with software and potentially receive access rights resembling those held by human workers. 

This distinction is more important than the difference between one generation of language models and another. 

A copilot helps an employee write a report. An agent retrieves the data, analyzes it, writes the report, distributes it and creates follow-up tasks. A copilot summarizes a loan application. An agent requests missing documents, checks information, produces an initial risk assessment, proposes terms and routes exceptions to an officer. A copilot explains portfolio performance. An agent monitors the portfolio and prepares or executes rebalancing trades within an authorized range. 

The likely near-term growth areas are operations that are document-heavy, repetitive and governed by identifiable rules: 

• Customer onboarding and know-your-customer reviews; 
• Fraud and anti-money-laundering investigations; 
• Commercial underwriting; 
• Insurance claims processing; 
• Compliance monitoring; 
• Financial planning preparation; 
• Investment research and product comparison; 
• Accounting close and management reporting; 
• Procurement, vendor review and contract analysis; 
• Customer-service resolution. 

Specialized agents are already appearing in advisory technology. In June, YCharts introduced an agent built on Anthropic’s Claude with access to the platform’s financial data, tools and reporting formats. The product’s significance lies in its specialization: rather than asking a general chatbot about markets, advisers can use an agent situated inside an established financial workflow. 

This is likely to become a dominant model. Financial AI will move away from stand-alone chat windows and into the systems where work already takes place. Agents will be connected to approved data, granted specific permissions and evaluated according to their ability to complete tasks. 

The industry’s unit of automation will therefore change. Instead of automating individual steps, firms will automate outcomes. 

Human Connection May Become More Valuable, Not Less 

The transition will inevitably affect employment, but the consequences will not be limited to job losses. 

AI is particularly capable of taking over work traditionally assigned to junior professionals: gathering information, preparing preliminary analysis, checking documents, formatting reports and creating first drafts. If firms automate too much of this work, they may reduce entry-level hiring and weaken the apprenticeship process through which experienced financial professionals are developed. 

There is also a risk that companies will treat expected AI productivity as justification for reducing headcount before the technology is ready. Reports in July indicated that some employers outside and inside technology-intensive industries had reconsidered layoffs undertaken in anticipation of AI replacement, discovering that the systems could not fully absorb the eliminated work. 

The likely outcome is not the immediate disappearance of entire professions but the decomposition and redistribution of their tasks. 

Routine research, documentation and reconciliation will shrink. Exception handling, relationship management, model supervision and accountability will become more prominent. Demand may rise for professionals who combine domain knowledge with the ability to evaluate automated systems: AI-risk officers, model validators, data stewards, cybersecurity specialists, workflow designers and compliance professionals. 

In financial advice, this could strengthen the importance of the human relationship. As AI makes basic financial information and generic planning more abundant, advisers may be valued less for retrieving facts and more for understanding households, resolving conflicting priorities and helping clients act under uncertainty. 

AI can calculate how much a person should save. It may not understand why the person repeatedly fails to do so. It can identify a technically efficient estate strategy. It may not recognize that implementing it could intensify a family conflict. It can recommend staying invested during a decline. It cannot automatically create the trust necessary for a frightened client to follow that recommendation. 

The more analysis becomes automated, the more differentiation may depend on judgment, communication and human connection. 

Does the Industry Know What It Wants AI to Do? 

The answer is both yes and no. 

Financial institutions know what they want in operational terms. They want AI to reduce costs, increase capacity, improve fraud detection, personalize customer service, produce better forecasts and help employees navigate complex data. 

They are increasingly vocal about those objectives. Banks describe AI-first operating models. Advisers discuss technology that gives them more time with clients. Regulators call for controlled experimentation. Technology companies promise agents that can complete entire workflows. 

What remains unresolved is the institutional end state. 

Is AI primarily a tool used by a financial professional? Is it a digital employee supervised by a professional? Is it a new intermediary operating between consumers and financial institutions? Or will it become an autonomous economic participant capable of selecting products, negotiating terms and transacting on behalf of individuals and businesses? 

Different parts of the industry appear to be pursuing all four models simultaneously. 

The sensible near-term principle is not that a human must manually approve every action. That would neutralize many of AI’s benefits and encourage meaningless rubber-stamping. Nor should firms automate every decision for which a model achieves a better average result. 

The appropriate boundary depends on whether the action is explainable, bounded, observable and reversible. 

AI should be permitted broad authority over low-risk administrative work. It may receive conditional authority over financial operations conducted within clear limits. Human beings should retain decisive authority where an action imposes a fiduciary obligation, affects a person’s access to essential financial services, involves disputed facts or creates systemic consequences. 

That model assigns AI a large role without allowing accountability to disappear. 

The Role Ahead: More Than an Assistant, Less Than an Institution 

AI’s immediate role in financial services is to make human workers more productive. Its emerging role is to operate financial workflows. Its eventual role may be to participate directly in economic decisions. 

The transition will not occur evenly. Some firms will remain focused on summarization and employee copilots. Others will deploy specialized agents within underwriting, compliance, research and customer service. A smaller group will attempt to build institutions around automation from the beginning. 

The winners will not necessarily be the organizations that automate the greatest number of tasks. They may be those that determine most clearly which decisions should be automated, which require human judgment and how customers can challenge a machine-generated outcome. 

Financial services has always depended on a combination of information, trust and accountability. AI can expand the information available and reduce the cost of acting on it. It cannot, by itself, decide what deserves trust or who should be accountable. 

That remains the human role—and defining it may become the industry’s most important AI project.

Researched by DWN Staff