The next phase of digital banking may not be a better app. It may be a market in which software agents represent customers and banks, compare terms, make offers, execute transactions and continuously optimise financial relationships.
Imagine giving a personal AI a standing instruction: keep my emergency savings inside insured limits, earn the best available return, avoid teaser rates, never charge me an overdraft fee, and ask before moving more than 10% of my cash. The agent checks your accounts, scans the market and approaches banks. On the other side, the bank's own software evaluates your eligibility, relationship value, funding needs and retention risk, then decides what it is prepared to offer. The two systems settle on terms. You see the result, not the negotiation.
That is not yet how mainstream retail banking works. But it is no longer science fiction either. The UK's Financial Conduct Authority said in July 2026 that one in five consumers in its research would consider AI that acts autonomously within pre-set goals, while its Mills Review explicitly examined a future of rapid switching, ongoing financial management and AI interfaces that become new routes to market. (FCA - Landmark review into the impact of AI on retail financial services) (FCA - The Mills Review: AI and the future of retail financial services)
Meanwhile, the execution layer is moving faster than the negotiation layer. Santander and Mastercard have completed a live end-to-end AI-agent payment in Europe; BBVA and Visa have tested an agent-initiated payment using real card credentials and an active merchant; DBS has moved its corporate virtual assistant from answering questions toward completing authenticated banking tasks; and ING has piloted an agentic mortgage assistant while keeping the final lending decision with a human employee. (Mastercard - Santander and Mastercard complete Europe's first live end-to-end payment executed by an AI agent) (BBVA - First AI agent-initiated transaction with Visa) (DBS - Gen AI-enabled virtual assistants go agentic) (ING - How ING uses agentic AI to speed up mortgage decisions)
Evidence versus inference: Live agentic payments and authenticated banking task execution are already happening. Continuous AI-to-AI negotiation over savings rates, loan pricing or fees is still an emerging scenario. This article treats that negotiation layer as a plausible next step, not as a description of a mature market today.
The key change is delegated authority, not a smarter chatbot
Generative AI made banking interfaces conversational. Agentic AI changes something deeper: software can be given a goal, permission to use tools and a bounded authority to act. JPMorganChase describes the security shift in similar terms: agents do not merely produce outputs; they operate with delegated authority, invoke external services and execute actions, making runtime authorisation and auditability central controls. (JPMorganChase - Securing the next generation of AI agents)
The technology is also beginning to demonstrate decision capability inside financial workflows. A BIS working paper tested a generative AI agent in simulated wholesale payment scenarios and found that it could replicate key cash-management practices, including maintaining liquidity buffers, prioritising urgent payments and balancing settlement speed against liquidity cost. The authors explicitly called for safeguards, human oversight and further research before wider adoption. (BIS - AI agents for cash management in payment systems)
Agentic banking therefore should not be understood as 'ChatGPT inside a banking app'. The more consequential model is a software representative that can observe a financial position, reason against a mandate, communicate with other systems and execute permitted actions. Once both the customer and the bank have such agents, negotiation becomes a logical extension of automation - although not an inevitable one.
What an AI-to-AI negotiation would actually look like
In practice, machine negotiation would probably look less like two chatbots haggling and more like a structured market. A consumer agent could broadcast a request for terms: a six-month savings home for GBP 50,000, instant access, no introductory cliff, deposit protection required. Bank agents could respond with machine-readable offers derived from product rules, liquidity needs, customer eligibility and pricing guardrails. The consumer agent could rank those offers, ask for a better rate or fee waiver, and execute only if the result stays inside its mandate. (FCA - The Mills Review: AI and the future of retail financial services)
A possible agent-to-agent banking workflow
| Stage | Customer agent | Bank agent | Required control |
| Mandate | Defines goals, limits and exclusions | Accepts only valid, verifiable instructions | Identity, consent, revocation |
| Discovery | Requests eligible products and terms | Returns machine-readable product data | Open access and transparent ranking |
| Offer | States amount, duration and preferences | Prices within policy, risk and balance-sheet limits | Fair treatment and approved pricing rules |
| Counteroffer | Asks for improved rate, fee or feature | Accepts, declines or revises terms | No hidden paid prioritisation |
| Execution | Approves automatically or seeks human sign-off | Books product or initiates payment | Strong authentication and transaction limits |
| Aftercare | Monitors outcome and triggers switching rules | Services product and records rationale | Audit trail, complaints and redress |
The point is not that banks will expose unrestricted bargaining engines. The likely near-term architecture is bounded: products, eligibility criteria and pricing ranges remain controlled by the institution, while agents automate search, comparison, requests and execution. The 'negotiation' is therefore less about improvisation and more about reducing the friction that currently protects margins and slows switching.
Payments are becoming the proving ground for delegated action
Payments matter because they force the industry to solve the hardest practical question early: how can a machine spend money on somebody's behalf without making consent invisible? Mastercard's March 2026 test with Santander used predefined limits and permissions. BBVA's July test with Visa used tokenisation, real-time fraud monitoring and biometric payment authentication. Visa's Agentic Ready programme has been expanded across multiple regions so issuing banks can test enrolment, tokenisation, authentication and transaction authorisation before agent-led commerce scales. (Mastercard - Santander and Mastercard complete Europe's first live end-to-end payment executed by an AI agent) (BBVA - First AI agent-initiated transaction with Visa) (Visa - Global expansion of Agentic Ready)
The UK government is now treating this as a regulatory design issue rather than a distant technology forecast. Its July 2026 consultation on modernising payment-services regulation says agentic AI could analyse, initiate, approve and execute payments for consumers or firms, and asks whether rules on authentication, consent and liability for unauthorised transactions need to change. That is an important signal: the problem is moving from 'can an agent pay?' to 'what does valid authority look like when a person is not clicking every step?'. (HM Treasury - Modernising Payment Services Regulation Consultation)
These payment experiments do not prove that AI agents are ready to negotiate mortgages or investment products. They do prove that the financial system is building the identity, permissioning and execution rails on which more ambitious forms of agentic banking would depend.
The bank side is moving from assistance toward action
Bank-side agents are developing along a different path. ING's mortgage pilot analyses applications that would otherwise require manual assessment, explains possible outcomes and suggests next steps, but an employee remains responsible for the assessment and final decision. That is a useful marker of where the industry currently draws the line around high-stakes judgement. (ING - How ING uses agentic AI to speed up mortgage decisions)
DBS provides a more customer-facing example. In July, DBS said its Joy assistant for corporate and SME customers in Singapore had become agentic, allowing authenticated users to move from asking questions toward completing simple banking tasks in one conversation. Its retail digibot is scheduled to add selected agentic capabilities later in 2026, including card servicing and fee-waiver requests, with escalation paths to human support. (DBS - Gen AI-enabled virtual assistants go agentic)
The next economic step is easy to imagine. A bank already has systems for eligibility, credit policy, pricing, fraud, product inventory and customer retention. An agent that can coordinate those systems could decide, within approved boundaries, whether to waive a fee, improve a savings rate, offer a retention incentive or prepare a refinancing option. What is not yet established is whether banks will allow those agents to negotiate dynamically with external consumer agents at scale.
Customer inertia could become a much smaller source of profit
This is where agentic banking becomes strategically important. Retail finance contains a large amount of friction: consumers do not continuously compare savings rates, refinance every time pricing moves, optimise card rewards each week or shift cash the moment another provider offers a better deal. Time, complexity and attention all create inertia.
The Mills Review argues that AI interfaces could materially reduce those barriers. It describes 'hyper-switching' in which agents continuously manage finances and execute rapid switching, potentially increasing competition across retail banking, investment platforms and insurance. The same report warns that effortless movement of savings could erode traditional margins and change bank funding dynamics. (FCA - The Mills Review: AI and the future of retail financial services)
For customers, that could be powerful. An agent can monitor hundreds of conditions continuously in a way a human cannot. For banks, it could compress the value of being the default provider. Deposit rates may have to react faster. Retention offers may become more systematic. Product features will need to be legible to machines, not merely persuasive to people. The economic advantage could shift from owning the customer interface to being the institution that can produce the best risk-adjusted offer in real time.
There is also a stability counterargument. If thousands or millions of agents share similar objectives - maximise yield, minimise perceived risk, switch instantly - they could make deposits more mobile precisely when banks are under stress. South African Reserve Bank Deputy Governor Fundi Tshazibana warned in May that AI agents could move funds from a bank perceived as weak at a speed that makes today's digital bank-run dynamics look slow. The Bank of England's Financial Policy Committee has similarly said agentic AI could create externalities in payments and financial markets even though there is currently little evidence of advanced AI being used for core financial decisions at a systemic scale. (BIS - Fundi Tshazibana on regulation and supervision in the age of AI) (Bank of England - Financial Policy Committee Record, April 2026)
The new gatekeeper may be your agent, not your bank
The optimistic story is that a personal agent becomes a tireless fiduciary-like shopper: it searches the whole market, understands the user's priorities and forces providers to compete. The less comfortable possibility is that the agent becomes another distribution gatekeeper.
The FCA's Mills Review specifically warns that AI interfaces could narrow the effective choice set if access depends on technical integrations, subscription tiers, commercial arrangements or paid prioritisation. A service can appear to search the market while in practice presenting only products it can access or has incentives to surface. In that world, the most important consumer-protection question is not simply whether the bank's offer is fair. It is whether the customer's agent is truly acting for the customer. (FCA - The Mills Review: AI and the future of retail financial services)
That creates a new competitive battleground. Banks may compete for preferential agent integrations. Fintechs and model providers could sit upstream of traditional financial distribution. Smaller banks may gain access to customers if open standards make products easy for agents to compare - or lose visibility if dominant AI interfaces favour the biggest and easiest-to-integrate providers. The same technology can therefore increase contestability or create a new form of concentration, depending on market design.
Identity, authority and liability become financial infrastructure
An agent cannot safely negotiate with a bank merely because it knows the customer's password. Financial services need a stronger concept of delegated authority: who the agent represents, what it is permitted to do, how much it can spend or move, which products it may enter, when it must ask for approval, how authority can be revoked, and how every action can be reconstructed afterwards.
This is why the FCA's Mills Review recommends a trusted framework for AI-agent participation built around identity, valid instructions, structured and verifiable mandates, accountability, auditability and interoperability across data and payments. JPMorganChase makes a parallel security argument: machine-to-machine interactions need authentication, agent actions need to be traceable, and higher-risk operations require stronger safeguards and the ability to stop an agent that deviates from expectations. (FCA - The Mills Review: AI and the future of retail financial services) (JPMorganChase - Securing the next generation of AI agents)
Liability is the harder edge. Suppose a consumer agent misunderstands a mandate and moves a deposit. The bank's agent accepts the instruction because the credentials are valid. The transaction then causes a loss. Was the payment authorised? Was the mandate defective? Did the bank have a duty to recognise abnormal behaviour? Did the general-purpose AI provider provide a regulated service, or merely a tool? The UK's payment consultation is already asking versions of these questions for agentic transactions, but a globally consistent answer does not yet exist. (HM Treasury - Modernising Payment Services Regulation Consultation)
Model risk becomes relationship risk between machines
Traditional model-risk management asks whether a model is accurate, stable, explainable and appropriately governed. Agent-to-agent finance adds a second layer: what happens when individually controlled systems interact repeatedly and adapt to each other's behaviour? A bank may understand its own pricing agent but not the optimisation logic of thousands of customer agents approaching it at machine speed.
The immediate risks are practical rather than exotic. Agents can be manipulated by untrusted inputs, tricked into using the wrong tool, granted excessive authority or compromised through third-party integrations. JPMorganChase argues that agents combining untrusted inputs, sensitive data and external authority require continuous runtime controls because even individually authorised actions can create unintended outcomes when sequenced together. (JPMorganChase - Securing the next generation of AI agents)
At system level, the concern is correlation. If many agents use similar models, data and objectives, they may react in similar ways to the same signal. Bank of England Deputy Governor Sarah Breeden has argued that a more agentic financial system will operate with greater autonomy, scale and speed and will require central banks to rethink how they monitor risk. The FSB's 2026 consultation on responsible AI adoption likewise focuses on governance across the AI lifecycle, cyber risk and third-party dependencies, including newer forms of AI such as agentic systems. (Bank of England - Agents of change, speech by Sarah Breeden) (Financial Stability Board - Sound Practices for Responsible Adoption of Artificial Intelligence)
The commercial winners may be the banks that are easiest for machines to trust
If agentic banking develops, the winning capabilities may look different from today's digital-banking scorecard. A beautiful app still matters to humans, but an AI representative needs structured product data, reliable APIs, clear eligibility logic, digital identity, machine-readable disclosures, fast execution and evidence that a counterparty will honour the mandate and provide redress when something goes wrong.
Banks with strong data architecture and real-time pricing could turn agents into a new distribution channel. Fintechs could specialise in independent consumer representation. Payment networks can provide trusted transaction credentials and agent recognition. Core-banking and software vendors can supply orchestration and controls. The value chain may become more modular at exactly the moment customers see less of it.
But there is a strong counterargument to the most autonomous vision. Consumers remain cautious. The FCA survey underpinning the Mills Review found willingness falling as autonomy increased: 36% said they would use assistive AI, 30% AI that acts with permission and 20% fully autonomous AI. Respondents also emphasised protection, accuracy, the ability to opt out and access to human support. In finance, convenience is not the only objective; people care about responsibility when money is lost. (FCA - The Mills Review: AI and the future of retail financial services)
Regulators will have to supervise the mandate, not just the model
The emerging regulatory principle is that accountability does not disappear because a machine acted. Tshazibana put the point plainly in her May speech: the AI tool does not own the risk; the humans holding the financial-services licence do. The FCA says its current approach continues to rely on existing outcomes-focused frameworks such as the Consumer Duty and Senior Managers Regime while it considers how those frameworks should adapt as autonomy increases. (BIS - Fundi Tshazibana on regulation and supervision in the age of AI) (FCA - Landmark review into the impact of AI on retail financial services)
That suggests the unit of supervision will gradually expand. Regulators will still assess firms and models, but they may also need visibility into mandates, agent identities, delegated permissions, audit logs, external integrations and the behaviour produced when multiple agents interact. The FCA has already recommended foundations for agentic finance and is live-testing use cases that include agentic payments, while the FSB is developing international sound practices for responsible AI adoption. (FCA - The Mills Review: AI and the future of retail financial services) (FCA - Second cohort for AI Live Testing) (Financial Stability Board - Sound Practices for Responsible Adoption of Artificial Intelligence)
Open standards will matter as much as model rules. If agents cannot access product information or customer-authorised data on comparable terms, the market risks becoming closed and vertically integrated. If access is too permissive, fraud and data leakage become easier. Agentic banking therefore sits at the intersection of AI governance, Open Finance, payments regulation, digital identity, competition policy and operational resilience - a combination that makes single-regulator answers unlikely.
The human role may move from clicking to setting policy
The most plausible near-term future is not a fully autonomous financial life. It is bounded delegation. People set objectives, exclusions and thresholds; agents handle research, monitoring and low-risk execution; humans approve high-impact actions and can intervene when circumstances change. Banks use their own agents to assemble offers and service requests, but keep hard limits around underwriting, suitability, fraud, liquidity and conduct risk.
That may still be a profound change. The first era of digital banking moved the customer from the branch to the screen. Agentic banking could move much of the customer journey off the bank's screen altogether. A financial institution may increasingly have to win business by convincing software acting for the customer, not merely by designing a persuasive app or marketing campaign.
The decisive question is therefore not whether AI agents will become capable of negotiation. The building blocks suggest they will. The harder question is whether the industry can create a system in which the agent's authority is limited, its incentives are visible, its actions are auditable and responsibility remains clear when two machines make a financial decision faster than either human party could have done alone.
Conclusion
Agentic banking is emerging first through practical, bounded tasks: payments, servicing, workflow automation and decision support. True AI-to-AI negotiation over financial products is still ahead of the market, but the infrastructure for delegated action is being built now. If that next layer arrives, it could reduce customer inertia, intensify price competition, shift distribution power toward AI interfaces and make bank funding more responsive - and potentially more fragile.
For banks, the strategic challenge is to become agent-ready without giving autonomous software more authority than their governance can safely absorb. For fintechs and technology providers, the opportunity is to become trusted representatives and infrastructure providers rather than just better interfaces. For regulators, the task is to preserve consumer control and market contestability while ensuring that responsibility never disappears into the gap between one agent and another.
References
1. Financial Conduct Authority. FCA publishes landmark review into impact of AI on retail financial services. 6 July 2026.
2. Financial Conduct Authority. The Mills Review: AI and the future of retail financial services. July 2026.
3. HM Treasury. Modernising Payment Services Regulation Consultation. 14 July 2026.
4. Bank for International Settlements. AI agents for cash management in payment systems. 2025/2026 working paper.
5. Mastercard / Santander. Santander and Mastercard complete Europe's first live end-to-end payment executed by an AI agent. 2 March 2026.
6. BBVA / Visa. BBVA Completes Its First AI Agent-Initiated Transaction Together with Visa. 2 July 2026.
7. ING. How ING uses agentic AI to speed up mortgage decisions. 16 June 2026.
8. DBS. DBS Gen AI-enabled virtual assistants reach 10 million customers and go agentic. 28 July 2026.
9. JPMorganChase. Securing the next generation of AI agents. 23 March 2026.
10. Financial Stability Board. Sound Practices for Responsible Adoption of Artificial Intelligence: Consultation report. 10 June 2026.
11. Bank of England. Financial Policy Committee Record - April 2026. April 2026.
12. Bank of England. Agents of change - speech by Sarah Breeden. June 2026.
13. South African Reserve Bank / BIS. Regulation and supervision of the financial sector in the age of artificial intelligence - Fundi Tshazibana. 20 May 2026.
14. Visa. Visa Announces Global Expansion of Agentic Ready Program. 29 April 2026.
15. Financial Conduct Authority. FCA announces second cohort for AI Live Testing. 21 April 2026.