A customer notices an unfamiliar payment and opens the chat window in a banking app.
Today, the chatbot might identify the transaction, explain how card disputes work and direct the customer to a form. A more autonomous system could go several steps further. It might review the account history, ask the customer targeted questions, collect the necessary evidence, prepare the dispute and submit it for approval.
That difference between explaining a process and carrying it out is at the center of growing interest in agentic artificial intelligence.
The financial industry has used AI for years across areas such as credit assessment, customer service, transaction monitoring and the identification of potentially suspicious activity. Generative AI added the ability to produce natural-language answers, summaries and documents. Agentic AI introduces another element: the capacity to pursue an assigned objective, choose among possible steps and interact with other systems with limited supervision.
Its emergence could change how people shop, manage accounts and communicate with financial providers. It could also alter the internal operations of banks, payment companies and fintech platforms.
Yet greater autonomy changes the risk equation. When an AI system stops merely suggesting an action and begins performing it, questions about permission, accountability and consumer protection become harder to avoid.
Beyond the Familiar Chatbot
The term “agentic AI” does not have one universally settled definition. In general, it describes systems designed to work toward a goal rather than respond only to isolated instructions.
A conventional chatbot typically handles one exchange at a time. It answers a question, generates text or retrieves information. An AI agent may instead divide a larger objective into smaller tasks, use software tools, assess the outcome of each step and decide what to do next.
The Bank for International Settlements has said agentic AI systems can support decision-making and automate tasks by coordinating complex workflows. That capability remains an area of active development rather than a finished, uniformly adopted technology.
Consider a small business trying to collect several overdue invoices. A chatbot could draft a payment reminder. An agent could potentially identify the unpaid invoices, check whether any are disputed, prepare individualized messages, schedule follow-ups and update the accounting platform after payment arrives.
The attraction lies not in any single task but in joining those tasks together.
Financial services contain countless processes that cross departments, databases and software platforms. Customers repeatedly enter the same information. Employees move data between systems, review documents and wait for approvals. An agent that can navigate those steps could reduce administrative work and shorten processing times.
That promise explains the industry’s interest. It does not establish that the systems are ready to operate without close controls.
Payments Move to the Forefront
Payments are becoming one of the most visible testing grounds for agentic AI.
AI-assisted shopping already helps consumers research products and compare prices. The next step is enabling an authorized agent to complete a purchase. That requires more than a capable language model. The agent must be able to identify itself, access an approved payment credential, follow the user’s limits and leave an auditable record of what occurred.
Visa introduced Visa Intelligent Commerce in 2025 as an initiative intended to connect AI agents with its payment network. Its developer materials describe agent-specific payment tokens and controls designed to ensure that an agent’s purchasing activity corresponds with authenticated customer instructions. Visa also cautions that the product remains in deployment and that its eventual features may differ from those currently described.
In December 2025, Visa said it and participating partners had completed secure agent-initiated transactions and planned additional pilot activity in several regions. The announcement represented progress toward commercial deployment, although it came from a company with a direct interest in developing the market.
Mastercard introduced Agent Pay in 2025 to support transactions involving verified AI agents. In June 2026, it introduced Agent Pay for Machines, a service that builds on the earlier program and is designed for high-frequency, low-latency and low-value transactions carried out by software agents and machines. Mastercard said the service includes credentialing, permissioning and transaction controls, with settlement across cards, accounts and stablecoins.
What an Agent Could Do with Money
The possibilities extend beyond buying a product online. A consumer might instruct an agent to pay household bills while maintaining a minimum account balance. A traveler could ask one to compare payment methods based on exchange rates and fees. A business could use an agent to reconcile incoming payments, monitor working capital or prepare supplier transfers for authorization.
In payments infrastructure, researchers at the BIS tested generative AI agents in simulated cash-management scenarios. The agents maintained liquidity buffers, prioritized urgent payments and balanced liquidity costs against settlement delays across different scenarios. The authors said the findings indicated potential operational benefits, but they also emphasized the need for safeguards, human oversight and further research.
A research result produced under controlled conditions does not prove that an agent can safely manage real institutional funds. Real financial systems contain incomplete data, cyberattacks, regulatory obligations and unusual events that may be difficult to reproduce in a simulation.
Outside payments, agentic systems could support fraud investigations, customer complaints and compliance reviews.
An agent examining a suspicious transaction might retrieve account records, compare device information, review previous alerts and assemble a case for a human investigator. In customer service, it could follow a complaint across several internal systems instead of directing the customer from one department to another.
These applications may sound less dramatic than an AI agent spending money independently. They could nevertheless have an earlier and broader operational impact because financial institutions already perform such workflows at scale.
Delegation has Limits
The convenience of agentic finance depends on delegation. Its risks do too.
A customer might be comfortable allowing an agent to compare insurance policies but not to purchase one. The same person may authorize automatic grocery orders below $100 but require confirmation before any larger transaction.
Businesses are likely to establish more elaborate restrictions. An agent may be allowed to prepare payments but not release them, move funds only between approved accounts or operate solely during specified hours.
These permissions will need to be precise and understandable.
“Buy the cheapest suitable flight” may appear to be a straightforward instruction. In practice, it leaves several questions unanswered. Does “cheapest” include baggage charges? Can the agent book a flight with a long layover? May it choose a nonrefundable ticket? How should it weigh price against arrival time?
Financial instructions can be equally ambiguous. An agent asked to reduce monthly expenses could cancel a service the customer considered essential. One told to maximize investment returns could take more risk than the user intended.
Greater technical capability does not eliminate the need for clearly defined authority.
Trust Becomes a Product Requirement
Consumers may judge financial agents less by their sophistication than by whether they remain under the user’s control.
Visa-commissioned research involving online shoppers in Australia, New Zealand and the United States found that approximately 85% of respondents wanted explicit control over the data available to an agent. About half were concerned about decisions being made without them, while roughly 43% worried that an agent could select the wrong product. Because Visa funded and published the research, the findings should be treated as company-sponsored evidence rather than a neutral assessment of the entire market.
Those concerns resemble earlier debates over online banking and digital wallets, but agentic systems add another layer. The customer is not only trusting a digital channel to carry an instruction. The customer may be trusting software to interpret the instruction first.
Payment networks are therefore focusing on tokenization, authentication and agent identification. The aim is to establish which agent initiated a transaction, whose authority it was using and whether the payment remained within the customer’s stated limits.
Even strong authentication cannot settle every dispute. An agent may be genuine and properly authorized but still misunderstand the customer.
When an Agent Makes the Wrong Decision
A chatbot that gives an inaccurate answer can confuse a customer. An agent that acts on an inaccurate answer can cause a financial loss.
It might purchase the wrong item, transfer an incorrect amount, miss a suspicious payment or expose information to an unsafe third party. It could also be manipulated by fraudulent instructions embedded in a website, email or document it has been asked to process.
Responsibility may be spread across several organizations. A bank could hold the account, a payment network could route the transaction, one company could supply the AI model and another could operate the application used by the customer.
When something goes wrong, each participant may have influenced the outcome.
This fragmentation creates difficult questions. Was the instruction sufficiently clear? Did the agent operate outside its permission? Should the financial institution have recognized unusual behavior? Did the technology provider adequately test the system? Can the transaction be reversed?
Consumers will need practical answers, not explanations of how a model reached its decision.
Existing Rules Still Apply
Regulators have not waited for a single body of law labeled “agentic finance.”
Financial institutions remain subject to requirements concerning consumer treatment, privacy, operational resilience, outsourcing, financial crime and risk management. Using AI does not necessarily remove those obligations or transfer them to the technology provider.
The United Kingdom’s Financial Conduct Authority says its existing regulatory framework applies to firms using AI. In July 2026, it published the Mills Review, examining how AI could reshape retail financial services through 2030 and recommending how the regulator should respond. The review identified opportunities to improve services and detect harm, alongside risks involving fraud, exclusion, market concentration and failures that could spread more quickly through the financial system.
The BIS has similarly warned that inadequate controls around AI could amplify financial vulnerabilities. Potential concerns include dependence on common technology providers, greater market correlation and the rapid transmission of errors.
Agentic AI may make these issues more urgent because it combines analysis with execution. A flawed recommendation still requires someone to act on it. A flawed agent may take the next step itself.
The Human Role Changes Rather Than Disappears
Agentic finance is often described through the language of autonomy, but its near-term development is likely to involve varying levels of human supervision.
Routine, low-risk processes may be automated more extensively. Sensitive decisions such as denying credit, freezing an account or transferring substantial funds are more likely to retain formal approval requirements.
Human oversight, however, must be meaningful.
An employee who receives hundreds of AI-generated recommendations and approves them automatically is not providing an effective safeguard. Reviewers need sufficient information, authority and time to challenge the system.
Financial companies will also need mechanisms for customers to stop an agent, alter its permissions and dispute its actions. An automated process should not make human assistance harder to reach precisely when something has gone wrong.
The most credible systems may therefore be those that are selective rather than fully autonomous: able to perform clearly defined tasks, transparent about their limits and designed to hand control back to a person when uncertainty becomes significant.
From Assistance to Authority
Agentic finance remains an emerging field. Many products are being piloted, announced or developed rather than used routinely by the public. Some of the most ambitious predictions come from companies seeking to sell the infrastructure behind them.
AI is being connected to payment credentials, banking systems, merchant platforms and financial data. The debate is shifting from what AI can say to what it should be permitted to do.
The transition will probably occur gradually. Agents will research first, then recommend. They will prepare transactions before they execute them. They will operate within narrow limits before receiving broader authority.
Each step will test whether convenience can be added without weakening accountability.
The defining question of agentic finance is therefore not whether an AI system can complete a financial task. Increasingly, it can. The harder question is whether customers, financial institutions and regulators can remain in control once it starts taking action.
