By Nicole Volpe, Contributor at The Financial Brand
Subscribe Now!
Stay on top of all the latest news and trends in the banking industry.
Ever since LLMs made their commercial debut three years ago, bank and credit union strategists evaluating their institutions’ AI readiness have focused mostly on data quality: Is our data sufficiently structured, governed, and reliable to support safe and useful outputs? But as important as it is to avoid the garbage-in, garbage-out problem, by placing such high priority on it, many institutions may be missing a key dimension of AI-readiness.
Taxonomy and data architecture are important first steps. But the question that comes immediately after that is whether an institution’s data can be safely deployed into AI-mediated customer and member relationships: permissioned, revocable, contextualized, and controlled in ways that preserve trust and brand.
This need has come into increasingly sharp relief as chat-based AI use cases give way to agentic use cases. The hard problem now is intelligence, meaning the work of turning data into understanding, prediction, and action reliable enough to help consumers make better financial decisions.
In an interview with The Financial Brand, Crystal Anderson, Chief of Staff at MX, described this shift as a matter of both consumer protection and institutional relevance: As consumers begin asking AI agents to solve financial problems and see them through to execution, financial institutions must be prepared to “ride with them” in these experiences, rather than cede the relationship to third-party interfaces.
To understand what makes a financial institution AI-ready now, it helps to focus on the tension between institutional caution and consumer behavior. Banks and credit unions have good reason to move carefully: they operate on sensitive proprietary data, across siloed systems, within strict governance and regulatory frameworks. But the consumer’s financial life is already becoming more connected, more permissioned, and more portable.
A KPMG survey of U.S. banking executives, including regional and community institutions, captures the institutional side of that tension, noting that “there is still reticence in the industry to release customer-facing GenAI solutions into production.” The survey connected that hesitancy to enterprise data management concerns, including data privacy and risk, data quality, and legacy systems or integration complexity.
Consumers, meanwhile, are already giving providers, financial apps, and even nonfinancial apps access to more of their financial lives. In a 2025 consumer study by MX, 55% of U.S. consumers said they would give their financial provider access to more of their data if it resulted in a better experience, up from 46% in May 2024. Forty-one percent of respondents said they have used digital tools to bring different financial accounts into one view — and as many as 70% may have connected a financial account or shared financial data within a mobile app, the study said.
Once consumers are willing to move financial data into digital experiences, they will increasingly expect those experiences to do more than provide a convenient overview or support specific transactions. They will expect answers to questions about spending, savings, cash flow and affordability that reflect their day-to-day financial needs and lifecycle milestones.
Many of those expectations are already reflected in the value propositions of established fintechs such as Betterment, Credit Karma, or Klarna. AI could give banks and credit unions a way to offer comparable intelligence, and eventually execution, inside relationships where they already hold trust, data and accountability.
Here are three ideas financial institution strategists should keep in mind as they assess the next phase of AI readiness.
1: The difference between deterministic and probabilistic really does matter.
The critical point in an AI-mediated financial experience is the moment data becomes interpretation. A transaction record, account balance or loan payment is one kind of information. A recommendation about what the accountholder can afford, what they should do next, or which financial path is safest is another. That middle step — the inference layer — is where AI-readiness becomes more complicated for banks and credit unions.
Banking systems are built around fixed, auditable outcomes: balances, payments, risk rules, fraud controls, underwriting criteria. AI systems, by contrast, are often used to generate judgments, recommendations or likely next steps. Both can be useful, but they operate differently. The challenge is defining how probabilistic tools can sit alongside deterministic banking systems without blurring the line between displaying data and shaping a financial decision.
A transaction feed can show that a consumer spent $2,500 at a car dealership. A budgeting app can categorize the transaction. But an AI agent may go further. It may infer that the consumer has purchased a car and help answer the question that actually matters to the accountholder: Do future car payments need to be reflected in future budgets? At that point, the experience has moved beyond organizing financial information. It is helping form a financial judgment.
“In ordinary consumer technology, a flawed AI answer may be inconvenient,” Anderson said. “In financial services, a flawed inference can become bad advice.” Once a consumer asks, “Can I afford this?” or “What should I do next?” the system is no longer simply presenting financial data. It is interpreting that data in a way that may influence a real-life decision.
For banks and credit unions, that makes the inference layer a control problem. If their data is going to support AI-mediated experiences, they need to govern more than which data can be shared. They need to understand what the system can infer from that data, what it can recommend, and when an answer needs to be grounded in deterministic validation before it reaches the user.
2: Contextual consent is becoming the control layer for financial AI.
For years, data permissioning in financial services has been treated largely as an access question: Can this app connect to this account? What data can it see? Can the consumer limit or revoke that access? In an AI-mediated environment, that framing is too narrow. Consent must also describe the purpose of the data use: what question the AI is being asked to answer, what information it can use to answer it, and where the consumer’s permission begins and ends.
That shifts permissioning from data movement to data interpretation. A consumer may be comfortable letting a budgeting tool review checking-account transactions to categorize spending. That does not mean the same consumer has agreed to let an AI agent combine deposits, card balances, loan payments and outside-account activity to make a recommendation about a larger financial decision. Each use case carries its own level of sensitivity, consequence, and institutional responsibility.
Consider again the accountholder making a car purchase. In the earlier scenario, the institution or AI system is looking for signals in transaction data. In a contextual-consent model, the end-user makes the purpose explicit: I want help deciding whether I can afford this car. That changes the consent question. The customer or member is no longer granting general access to financial data; they are authorizing a specific analysis using account balances, recurring income, existing debt obligations and other relevant information. Retirement assets might be excluded from the calculation. An external account might be added for this one analysis. Access could be revoked when the question is answered.
For the institution, contextual consent creates a way to set boundaries before its data enters an AI workflow. Those boundaries can reflect product rules, risk tolerance, compliance requirements, and consumer-protection standards. Such boundaries bring clarity to the UX and reduce confusion by tying access to a defined purpose, rather than asking for broad permission and leaving the user to infer what happens next.
MX’s consumer research underscores why that matters: 80% of respondents said it is important to know and manage who has access to their financial data, while 55% said they are worried about which financial providers have access to it. Contextual consent gives banks and credit unions a way to respond to that concern with a more useful form of control — one built into the product architecture, not treated as a single screen or legal acknowledgment.
3: The relationship has to travel with the data.
The AI risk for banks and credit unions is that the new technology could weaken their role in the relationship. Instead of logging into a banking app to gain perspective on their financial situation, an accountholder may start with an AI agent and pull banking data into that environment. Anderson urges institutions to work against that risk by ensuring that their own controls and brand “ride with” the customer or member as financial activity moves into AI-mediated experiences.
This is a different way to think about primacy. Historically, banks and credit unions have defined primacy based on specified criteria like the number of unique products an individual holds, or whether they use direct deposit or bill pay. AI introduces another test: When a customer or member asks a financial question, whose intelligence shapes the answer? If the institution supplies only the underlying account data while another interface explains it, interprets it, and recommends the next step, the relationship begins to shift toward whoever owns that experience.
Anderson offered the example of an MCP-enabled app inside an LLM environment. An accountholder might ask an AI assistant questions about weekly cash flow, travel affordability, or which account to use for a purchase. In one scenario, the AI assistant gets access to account data and generates an answer on its own, with the bank or credit union functioning mainly as the data source. In another, the AI environment prompts the accountholder to connect through the institution’s own MCP-enabled presence. The experience may still happen inside an AI environment, but the answer carries the institution’s data, intelligence, and attribution.
Anderson said: “As AI mediates how people interact with money, institutions can be reduced to back-end utilities while the relationship devolves to whoever controls the interface.”
For bankers, building AI into the authenticated banking experience is useful, but it is only part of the work. They also need to think about AI as a distribution challenge: where an end-user is likely to ask a financial question, what kinds of help they will expect, and how the institution can remain present when the interaction begins outside its own digital walls.
Trust is the thread running through all of this, but in the AI era trust has to become more operational than reputational. A customer may trust their bank in the traditional sense and still hesitate if an AI experience cannot explain why it needs access to certain data, what it is doing with that data, or why its answer should be believed. A credit union may trust the power of its brand and the strength of its compliance posture and still find itself exposed if its data is interpreted in an environment it does not understand or control.
The next phase of AI readiness, then, is about making trust visible in the mechanics of the experience: in the quality of the answer, the clarity of the permission, the source of the intelligence, and the customer or members ability to see that the institution remains accountable for the financial guidance carrying its name.
Join over 2,500 of the brightest minds in banking at the industry’s premier annual event — three days loaded with the big ideas and latest innovations that are transforming banking today.
The Financial Brand is your premier destination for comprehensive insights in the financial services sector. With our in-depth articles, webinars, reports and research, we keep banking executives up-to-date with the latest trends, growth strategies, and technological advancements that are transforming the industry today.
