Public support creates public expectations

Banking operates through a public–private compact. Charters confer authority. Deposit insurance protects covered depositors. Central-bank liquidity supports eligible institutions during funding stress. Public and private payment systems underpin commerce. In return, banks accept capital, liquidity, consumer-protection, community, financial-crime and supervisory obligations.

This support should be described accurately. The FDIC’s Deposit Insurance Fund is primarily financed by risk-based assessments on insured institutions and interest on Treasury securities, while the federal guarantee supplies essential public credibility. FDIC, Deposit Insurance Fund Federal Reserve discount-window lending is collateralized and generally priced above comparable market funding; it is a backstop rather than routine free financing. Federal Reserve, Discount Window FAQ Extraordinary crisis support and “too big to fail” expectations have nevertheless been real policy concerns. Treasury invested $245.5 billion through TARP bank programs and had recovered $275.9 billion by September 2023, while GAO found that crisis-era funding advantages declined after reforms but could not be measured with certainty. U.S. Treasury, TARP Bank Investment Programs, GAO, Large Bank Holding Companies

The balanced conclusion is not that banks receive universally easy or free money. It is that banking benefits from public architecture unavailable to ordinary businesses. Society can reasonably expect this architecture to produce resilience, fair competition, wider access and improving customer service.

Profit and customer value can grow together

Healthy earnings help banks absorb losses, invest and raise capital. FDIC-insured institutions earned $295.6 billion in 2025 and $80.5 billion in the first quarter of 2026. FDIC Quarterly Banking Profile, Q1 2026 Profit is not itself evidence of customer harm. The strategic question is whether revenue comes from useful intermediation and trusted service or from avoidable opacity, friction and customer inertia.

Agentic banking can improve both sides: lower the fixed cost of service, prevent losses, create more useful products and give employees capacity to deepen relationships. Some gains should appear as better prices, faster resolution and wider access; otherwise modernization will not strengthen public trust.

Price and deposits

The CFPB reported average 2024 APRs of 25.2 percent on general-purpose cards and 31.3 percent on private-label cards, alongside approximately $160 billion in interest and $31.3 billion in fees. CFPB, The Consumer Credit Card Market 2025 Prices reflect credit risk, rewards, fraud, capital and funding, but customers should see expected dollar cost and a safer alternative before committing. An agent can calculate that path continuously rather than advertise another product.

Deposit competition is similarly uneven. Banks in a Federal Reserve survey reported an average retail-deposit beta near 40 percent during a period when the policy rate rose 525 basis points, compared with higher betas for wholesale funding. Federal Reserve Senior Financial Officer Survey A governed agent can optimize liquidity and yield across approved accounts while respecting insured limits, upcoming obligations and relationship preferences.

Credit and service

The Federal Reserve estimates that approximately 32 million adults are credit invisible or have records too thin to score conventionally. Cash-flow evidence can improve access, but alternative data must be permissioned, relevant, stable and fair. Federal Reserve, Consumer & Community Context, October 2025 A useful agent assembles evidence and explains an achievable pathway; it does not replace an old opaque score with a more invasive opaque score. Creditors using complex algorithms still must provide accurate, specific adverse-action reasons. CFPB Circular 2022-03

Service problems often come from fragmented state rather than uncaring employees. An agent can authenticate once, retrieve the timeline, coordinate the responsible team and verify that a correction reached every system. It should be a bridge toward simpler architecture—not a permanent layer that hides incoherent processes.

Fraud, small firms and inclusion

Twenty percent of adults reported experiencing financial fraud or scams in 2025. Federal Reserve, Economic Well-Being of U.S. Households in 2025 Banks must intervene quickly without turning probabilistic suspicion into an unexplained freeze. Agents can combine context, deliver precise warnings, prepare disputes and coordinate recovery while payment rights and liability remain deterministic.

Small firms frequently have viable operations but fragmented records and no finance staff. The 2026 Small Business Credit Survey found that small-bank applicants were more likely to report full approval, while 60 percent of online-lender borrowers said actual cost was higher than expected. Federal Reserve Banks, 2026 Report on Employer Firms Agentic preparation can preserve relationship judgment while lowering the cost of serving a small account. That matters in a country where 4.2 percent of households remained unbanked in 2023 and useful inclusion still extends beyond account ownership. FDIC Economic Inclusion

The operating loop

Observe → Understand → Reason and plan → Act → Verify → Learn → Escalate to HITL (human in the loop)

This loop adapts the ReAct pattern—reasoning interleaved with tool use—to a regulated environment. ReAct is useful because it translates a disciplined form of human problem-solving into an agent: observe evidence, reason about the next step, act, inspect the result and revise rather than committing to one untested answer. The agent retrieves evidence, chooses the next authorized step, observes the result and updates the case. “Learning” means governed case memory, evaluated feedback and approved system improvement; it does not mean uncontrolled retraining on customer activity.

“Escalate” is an explicit handoff to a human in the loop, not an error message. Uncertainty, policy conflicts, vulnerable-customer signals, exceptions and consequential decisions must arrive with the relevant evidence, agent reasoning and proposed next step so an authorized person can review, override or stop the workflow. The agent can coordinate the case, but people continuously monitor the system and retain control of consequential judgment. Authoritative balances, contractual terms, interest, entitlements and limits remain in deterministic systems; a model should never improvise a ledger entry or legal right.

A compact hardship workflow

A mortgage-hardship agent can recognize an authenticated request, assemble loan and payment context, retrieve current assistance rules, calculate options through approved services and prepare a complete case. It may schedule, request missing evidence and send authorized notices, but a qualified employee owns contested facts, adverse decisions and vulnerability. The agent then monitors deadlines and verifies that an approved change posted correctly.

This is the practical difference between answering and completing. The customer does not have to reconstruct the process at every handoff, and the employee receives a coherent case rather than another unstructured conversation.

A visible permission envelope

Authority should be expressed as purpose, permitted data, allowed tools, monetary and time limits, reversibility and escalation. Four bands are enough for most deployments:

A prompt saying “do not move more than $500” is an instruction. A payment service that technically refuses $500.01 is a guardrail.

Five safety questions

Before production, an owner should answer:

1. What customer or employee outcome does the agent own, and whose interest does it serve?
2. Which facts, policy versions and calculations support each action?
3. What can the agent do, and what technically prevents action outside its mandate?
4. How does it perform on ordinary, adversarial, vulnerable and system-failure cases?
5. Who can stop, correct and appeal the result?

NIST’s voluntary AI Risk Management Framework and generative-AI profile provide useful cross-sector governance language, but banking law and institutional controls still govern the action. NIST Generative AI Profile

Journey one: household financial health

A household agent can:

1. Map income, obligations, rates, fees and available liquidity.
2. Forecast upcoming pressure and warn before a shortfall.
3. Compare repayment, saving and refinancing paths with visible tradeoffs.
4. Execute only approved transfers inside a customer mandate.
5. Monitor results and bring a banker or adviser into consequential decisions.

The customer gains continuous guidance instead of disconnected product dashboards. The bank gains retention, appropriate lending and fewer preventable service events. Compensation and optimization criteria must remain visible so “financial health” does not become disguised cross-selling.

Journey two: the small-business finance desk

For a business facing a seasonal funding gap, an agent can:

1. Reconcile orders, invoices, bank activity, payroll and accounting.
2. Maintain a probability-weighted 13-week cash forecast.
3. Identify the date, size and cause of a financing need.
4. Compare credit, invoice finance, supplier terms and owner capital.
5. Assemble verified evidence and monitor the chosen facility after approval.

This turns fragmented operating data into a finance capability the owner could not otherwise afford. It also allows a relationship manager to serve more companies without reducing judgment. The owner makes commitments and the banker owns underwriting.

Journey three: fraud prevention and recovery

When a customer is manipulated into sending money, an agent can:

1. Join device, payee, transaction and interaction signals.
2. Deliver a contextual warning and verified callback.
3. Apply only policy-authorized delay or confirmation.
4. Capture the customer’s story once and coordinate recovery evidence.
5. Track receiving-bank communication, liability review and customer updates.

Speed matters before settlement; empathy, law and clarity matter afterward. The agent should preserve evidence and coordinate action, while people own contested liability and vulnerability.

The opportunity map

Company inclusion is illustrative, not an endorsement. These companies occupy different layers, and inclusion does not mean every product is already a fully autonomous agent. Public descriptions do not prove production scale, accuracy or regulatory fitness.

Where established platforms and networks fit

It is worth including large providers, but separately. They are not startup wedges: they supply widely deployed customer, workflow, cloud and payment surfaces through which bank agents can retrieve context or take controlled action. A bank may combine one of these platforms with specialist startups rather than choose between them.

These companies can accelerate distribution and integration, but the bank still owns data entitlement, product suitability, customer recourse and consequential decisions. Their platforms also create openings for startups that provide specialized workflows, independent assurance, interoperability and migration rather than another general-purpose agent layer.

Seven practical entry strategies

Startups should not imitate a universal bank. They should select one expensive exception, prove safe completion and earn authority in stages. Banks can use the same framework internally: build where proprietary data, strategic differentiation or core authority matter; partner where a specialist can move faster.

The regulatory moat

The United States divides authority among the OCC, Federal Reserve, FDIC, NCUA, state regulators and product-specific agencies. Nonbanks may also require lending, servicing, collection, advisory or money-transmission licenses. CSBS reported that 31 states had enacted the Money Transmission Modernization Act in full or part by August 2026, improving consistency without eliminating state variation. CSBS, Money Transmission Modernization Act

A bank partnership does not transfer responsibility to the startup or excuse the bank from supervision. Interagency guidance expects lifecycle planning, due diligence, contracts, monitoring and termination. OCC, Interagency Guidance on Third-Party Relationships The four essentials are:

• choose the regulated role before designing the interface;
• encode delegated authority and jurisdiction into the workflow;
• preserve source, decision, complaint and appeal evidence; and
• provide operational continuity and a practical exit path.

These requirements slow a demo but create a durable product. A startup that converts regulation into reliable agent behavior gains a moat that a general-purpose model cannot reproduce.

Partnership is a competitive strategy

Large banks already deploy AI at scale through platforms such as Bank of America’s Erica, Wells Fargo’s Fargo, Morgan Stanley’s adviser assistants and JPMorgan Chase’s enterprise programs. Bank of America, Erica, Wells Fargo, Fargo, Morgan Stanley AI Assistant, JPMorgan Chase Annual Report

Their advantages are distribution, proprietary context, capital and governance. Their disadvantages are integration complexity and slower organizational change. Startups have the opposite profile. Enterprise sales, white-label delivery, minority investment, joint ventures and acquisition can therefore be as disruptive to industry economics as direct competition.

Build, partner, invest or acquire

The decision should follow the source of advantage rather than an institutional preference for building everything internally.

Partnership should not become vendor dependence. Banks need data portability, version notice, audit rights, continuity and the ability to replace the model or company. Startups need clear success criteria, timely bank decisions and access to the employees who understand the real workflow.

Where agentic banking can use digital-asset infrastructure

Agents can compare rails, enforce wallet mandates, screen counterparties, reconcile records and send exceptions to a human. They must never infer investment suitability, expose private keys or bypass AML and sanctions controls.

U.S. regulation is becoming clearer but remains incomplete

The U.S. framework is developing through the enacted GENIUS Act, the pending CLARITY Act, and guidance from the OCC, FDIC, Federal Reserve and SEC. Banks and startups still need product-specific review of licensing, custody, securities, AML, sanctions and consumer-protection obligations.

Innovation and risk should remain in the same frame

Investors see clearer rules as a path to U.S. innovation and capital formation (a16z crypto); the BIS emphasizes liquidity, illicit-finance, reserve and bank-funding risks (BIS). The strongest startup opportunities therefore lie in regulated custody, treasury, identity, compliance, reconciliation, wallet security and settlement—not speculative token promotion. Banks can provide trusted money and regulatory infrastructure; startups can make the new rails usable and safe.

Choose the instrument from the outcome

A payment or treasury agent should compare instant payments, cards, wires, correspondent banking, stablecoins and tokenized deposits by cost, speed, currency, recoverability and compliance. It may find an established rail best for a domestic payment, a regulated stablecoin useful after hours across borders, or a tokenized deposit suitable for institutional settlement.

Bitcoin and other volatile assets remain investments, not equivalents to insured deposits or dollar payments. An agent can explain their risks and prepare a supervised transaction, but explicit human approval and deterministic controls must preserve customer rights, money and recourse.

Zero involuntary job loss should be the program design

Agentic banking should begin with an explicit operating commitment: zero involuntary job losses attributable to the transformation program. The first use of returned capacity should be unresolved customer demand, compliance backlog, fraud investigation, quality improvement, product development and work that institutions never had time or resources to schedule.

This is a design objective, not a claim that technology can never affect employment. BLS projects different trajectories across banking occupations, including a decline in teller employment and modest growth in loan officers. BLS Occupational Projections The ILO concludes that job transformation is more likely than wholesale replacement in aggregate, while clerical occupations have the highest exposure. ILO, Generative AI and Jobs: A 2025 Update Institutions decide whether technology becomes a growth platform or a headcount target.

U.S. credit intermediation and related activities employed approximately 2.54 million people in June 2026. BLS, Credit Intermediation and Related Activities These employees contain the relationship knowledge, legal accountability and judgment that agents lack. The objective is to stop using that expertise to transport data between systems.

Four commitments make augmentation credible:

1. Automate tasks before deleting roles. Measure the work removed and where capacity goes.
2. Protect apprenticeship. Replace lost entry-level repetition with simulation, supervised cases and explicit judgment training.
3. Reward challenge. Employees must be able to stop or override an agent without a productivity penalty.
4. Share gains. Use productivity for better service, lower customer effort, new products, reskilling and career growth.

A concise economic value case

The relevant question is not how many hours an agent appears to save. It is whether the complete program produces value after technology, integration, control and transition cost.

The economic framework in Agentic AI: $16Tn in Global Economic Value Enabled and $8Tn Captured includes an independent 2031 banking and financial-services base scenario:

The arithmetic is $100B + $90B + $80B − $40B = $230B. These are author-modeled planning values—not forecasts endorsed by a bank, regulator or research firm. Payments and fintech’s separately modeled value overlaps banking and should not be added again.

The table also clarifies what creates value. Capacity matters, but nearly two-thirds of the gross captured value in this scenario comes from new products, growth, fraud reduction, credit quality, conduct, compliance and capital—not staff removal.

The net-value test for one bank

A national scenario does not justify an individual investment. Each bank should calculate value at the workflow level:

usable employee capacity + new contribution margin + avoided loss and risk benefit + customer surplus retained by the relationship − full program cost

“Usable capacity” means time that is actually redeployed to demand, quality, advice or product work. “New contribution margin” means revenue after the cost of delivery, not gross transaction value. Avoided loss needs an observed counterfactual rather than more alerts or fewer approvals. Full program cost includes models, data, integration, security, evaluation, employee transition, oversight, failure response and vendor exit.

For example, a service agent that drafts responses faster but leaves repeat contact unchanged has not created customer or operating value. A dispute agent that removes two handoffs, reduces repeat evidence, shortens correct resolution and returns employee time to complex cases has. The business case should specify where the returned capacity goes before the institution books it as a benefit.

At board level, five measures are sufficient: completed outcome, customer effort, employee capacity redeployed, risk-adjusted financial benefit and full cost. Adoption, prompts and generated text are operating diagnostics—not value.

Who benefits

Agentic banking can also improve the allocation of capital. Better cash-flow evidence can help capable borrowers and small firms; faster trade and treasury operations can release working capital; and continuous risk monitoring can identify problems before they become losses. These gains raise productive capacity rather than merely transferring income between a bank and customer.

Countries benefit when they build both the technology and the institutions required to deploy it: trusted identity, interoperable payments, strong consumer rights, skilled workers, competitive startups and regulators capable of inspecting agentic systems. The national prize is not only lower banking cost. It is more businesses formed, more useful products, better-directed investment and a financial system able to recover from shocks faster.

Make compliance a reusable agent service

Instead of embedding regulation separately in every chatbot or workflow, institutions can create shared services for jurisdiction, policy retrieval, permitted data, required notice, calculation, record retention and escalation. Every agent calls the same current control rather than reproducing law in a prompt.

This architecture makes regulation an innovation interface. A startup can demonstrate exactly which obligation is enforced, which evidence is produced and which action remains human. Regulators can inspect traces and outcomes instead of relying only on documents describing intended controls.

1. Select a broken outcome and baseline it

Choose one result customers or employees can recognize: resolve a card dispute without repeat contact, produce an underwriter-ready small-business file or investigate a reconciliation break. Measure time, handoffs, rework, errors, complaints, disparities, employee load, loss and cost before changing the process.

2. Redesign the process and map the rules

Remove duplicate data entry, unjustified approval and contradictory policy. Identify the legal entity, customer right, license, notice, record and responsible decision-maker at each stage. Automating an incoherent process only makes failure faster.

3. Connect context, memory and tools

Build a verified source hierarchy and a time-stamped case state. Connect only the systems needed to complete the outcome. Give the agent deterministic services for money and policy-dependent calculations, plus narrow tools for retrieval, request, scheduling, routing and approved action.

4. Encode guardrails and observability

Put purpose, access, monetary limits, counterparties, authentication, reversibility and escalation outside the model. Record every source, inference, tool call, result, model version and human approval so the complete workflow can be reconstructed.

5. Pilot in shadow and employee-assisted modes

First compare agent work with real cases without allowing external action. Then let employees use, edit and reject recommendations. Test rare, adversarial, multilingual, inaccessible, vulnerable, high-volume and vendor-failure cases. Study disagreement rather than averaging it away.

6. Scale, redeploy and measure net value

Increase authority only for the exact workflow, population and version that earned it. Remove the redundant old step, redeploy employees to named work and track capacity, new revenue, avoided loss, customer value and full program cost. Publish limitations and failures alongside benefits.

Roadmap by actor

Government also benefits directly from capable agentic banking: stronger fraud detection, more efficient supervision, clearer complaint evidence, better delivery of disaster and small-business programs, and greater visibility into systemic technology dependencies. Public procurement and regulatory sandboxes can demonstrate safe practices without implying that a pilot has received blanket approval.

Banking system, households and businesses

• FDIC Quarterly Banking Profile, Q1 2026
• FDIC Economic Inclusion
• Federal Reserve, Economic Well-Being of U.S. Households in 2025
• Federal Reserve Banks, 2026 Report on Employer Firms
• CFPB, The Consumer Credit Card Market 2025

Credit, consumer protection and regulation

• Federal Reserve, Consumer & Community Context, October 2025
• CFPB Circular 2022-03 on Adverse Action and Complex Algorithms
• Federal Reserve SR 26-2
• OCC, Interagency Guidance on Third-Party Relationships
• U.S. Treasury, AI in Financial Services
• GAO, Artificial Intelligence Use and Oversight in Financial Services
• FINRA Regulatory Notice 24-09

Digital assets and tokenization

• White House, GENIUS Act Signed into Law
• Congress.gov, Digital Asset Market CLARITY Act
• OCC Interpretive Letter 1183
• FDIC FIL-7-2025
• Federal Reserve, Crypto-Asset Guidance Withdrawal
• BIS Annual Economic Report 2026

Economic value and workforce

• IndusAgentAI, $16Tn in Global Economic Value Enabled and $8Tn Captured
• BLS, Credit Intermediation and Related Activities
• BLS Occupational Projections and Characteristics
• ILO, Generative AI and Jobs: A 2025 Update