As software becomes cheaper, faster and more autonomous, the scarce asset in banking may not be intelligence at all. It may be credible human judgment at the moments when customers most need someone to take responsibility.
For most of the digital era, banks have treated human service as a cost to be compressed. Mobile apps replaced routine branch visits. Straight-through processing removed manual handling. Chatbots absorbed simple questions. Artificial intelligence now promises to push that logic much further: drafting communications, detecting fraud, triaging complaints, assessing documents, personalising offers and, in some cases, contributing directly to decisions that affect customers' access to money. The economic case is obvious. Software scales almost without friction. Experienced human judgment does not.
But the same shift creates an underappreciated scarcity. When every institution can offer instant answers, generate polished explanations and automate routine interactions, speed alone stops differentiating the service. What becomes harder to copy is the ability to make a difficult call, explain it credibly, recognise when the standard process is failing, and put a named person behind the outcome. In that environment, the human element is not a nostalgic defence of branches or call centres. It becomes a premium capability - especially in lending, fraud, wealth management, business banking and vulnerable-customer support, where the cost of being misunderstood can be financial as well as emotional.
The evidence does not prove that customers will pay a measurable price premium for human contact in every banking product. That would be too strong a claim. What it does show is that AI adoption is accelerating faster than trust, explainability and governance can be taken for granted. The 'human premium' is therefore an inference from a widening gap: automated capability is becoming abundant, while accountable judgment remains scarce. For banks, fintechs and regulators, that gap is starting to look less like a soft customer-experience issue and more like a question of operating resilience and franchise value.
Automation is moving from the back office into the judgment layer
AI is no longer a small experimental layer on top of conventional banking. In the Bank of England and FCA's 2024 survey of 118 financial firms, 75% of respondents were already using AI and another 10% planned to use it within three years. More strikingly, 55% of reported AI use cases involved some degree of automated decision-making. Only 2% were fully autonomous, but the direction is clear: automation is moving closer to decisions that customers experience as consequential, not merely to internal efficiency tasks.
The European picture is similarly advanced. The European Banking Authority's 2025 monitoring of AI in banking and payments reported that 92% of EU banks in its sample were deploying AI, with the remainder pilot-testing or discussing use cases. That matters because customer-facing AI is not arriving after banks have solved every problem of data quality, explainability or accountability. It is arriving while those controls are still evolving. In the UK survey, 46% of respondent firms said they had only a partial understanding of the AI technologies they used, compared with 34% reporting complete understanding, with third-party models a significant source of the gap.
The technology itself is also becoming more capable. The Bank of England's July 2026 Financial Stability Report describes agentic AI as systems able to plan and execute multi-step tasks at machine speed using external tools with limited human oversight. That does not mean banks are about to hand the balance sheet to autonomous agents. It does mean the old control idea of putting a human somewhere in a long workflow is becoming less meaningful. The important question is no longer whether a person appears in the process. It is whether that person has the information, authority, time and skill to challenge the machine when challenge actually matters.
Digital convenience is not the same thing as trust
Banking has become dramatically easier to access digitally, yet that does not translate automatically into confidence in the institution. The UK's FCA Financial Lives 2024 findings show the contrast. Only 7% of day-to-day account holders said they did not bank online or use a mobile app, down sharply from 22% in 2017. Yet just 39% of adults said they had confidence in the UK financial services industry, while 36% thought most financial firms were honest and transparent in the way they treated them. The numbers describe one jurisdiction rather than the world, but the tension is instructive: digital adoption can become nearly universal while institutional trust remains much less secure.
This distinction matters because banks often optimise what they can measure most easily: log-in rates, digital containment, average handling time, cost per interaction, approval speed and conversion. Those metrics are useful, but they do not necessarily capture the moment a customer decides whether the bank is acting in their interest. Trust is usually tested at the edge of the standard journey - after a scam, during bereavement, when a payment is frozen, when a loan is declined, when a small business faces a cash-flow shock, or when markets fall and a wealth client is tempted to make an irreversible decision.
In those moments, the customer's problem is rarely a lack of information. It is a lack of certainty about what the institution will do next and who will stand behind the decision. AI can summarise policy, retrieve documents and recommend an action. Trust depends on something more: a belief that the institution can recognise context, correct an error and accept responsibility. That is why the human premium is not simply 'talking to a person'. A poorly trained employee reading a script creates little value. The premium comes from discretion backed by competence and accountability.
The premium appears when the standard journey breaks
Regulatory evidence already points to where that value emerges. In its review of the consumer-support outcome, the FCA highlighted firms that deliberately paired digital tools with human touchpoints when complex or sensitive needs appeared. One example automatically redirected a chatbot conversation about bereavement to a customer-support representative; another used vulnerability-related keywords to trigger a priority handoff. These are small design choices, but they capture a larger principle: automation works best when it knows the limits of automation.
The same FCA work found that in 19% of recent contacts or attempted contacts with a financial-services provider, consumers said finding the right contact information was difficult or impossible. Even when they reached someone, 13% said the provider's response was difficult to understand. Those frictions are not just service annoyances. They are moments when a bank can unintentionally turn operational efficiency into perceived indifference. If AI makes the front door faster but makes escalation harder, the institution may save money on routine contacts while increasing distrust precisely among customers with the most complex needs.
The strongest operating model is therefore asymmetric. Routine questions should become almost invisible: instant, accurate and cheap. Human capacity should be concentrated where ambiguity, vulnerability, conflict or consequence is high. That can mean an experienced fraud specialist who understands how victims behave under pressure; a credit officer who can evaluate an unusual but viable borrower; a relationship manager who can negotiate through a liquidity event; or an adviser who can stop a client from turning short-term fear into a long-term loss. In each case, the human is valuable not because the machine is incapable of producing an answer, but because the institution needs someone who can own the consequences of it.
When the algorithm says no, explanation becomes part of the product
This is most visible in credit. A model may be statistically strong and still feel illegitimate to the person on the wrong side of the decision if the bank cannot explain what happened. The BIS Innovation Hub's Project Noor is built around that problem. Led with the Hong Kong Monetary Authority, the UK's FCA and the Saudi Central Bank, the project is testing explainable-AI methods that can help supervisors interpret models used for decisions such as mortgage approvals, card limits and fraud flags. Its premise is straightforward: human-readable explanations can strengthen confidence, but financial institutions still retain responsibility for explainability.
The legal direction reinforces the same point. In the United States, the current Regulation B notification rule requires creditors taking adverse action to provide specific principal reasons, or to tell applicants how to obtain them. The rule also says that a generic reference to internal policy or failure to achieve a qualifying score is insufficient. That requirement is technology-neutral. A more sophisticated model does not make the customer's right to a meaningful reason less important; if anything, model complexity makes the bank's ability to translate a decision into understandable language more valuable.
Europe is moving in a similar direction through a different legal architecture. Under the EU AI Act's Annex III, systems used to evaluate the creditworthiness of natural persons or establish a credit score are listed as high-risk, subject to the Act's classification rules and exceptions. The Act's human-oversight provisions are designed so that people assigned oversight can understand system capabilities and limitations, detect anomalies and intervene where appropriate. Following the AI Omnibus changes, the high-risk rules for Annex III systems are scheduled to apply from 2 December 2027, while the Act's transparency rules began applying on 2 August 2026. The timetable has moved, but the policy signal has not: consequential automation needs intelligible accountability.
The counterargument: AI can make banking more trustworthy
There is a legitimate counterargument to the human-premium thesis. Humans are not automatically fairer, clearer or more reliable than machines. People tire, improvise, miss patterns, apply policy inconsistently and can carry biases of their own. AI can give customers 24-hour access, translate complex language, search large records instantly and identify suspicious activity that a person would never notice. In the Bank of England and FCA survey, the highest perceived current benefits of AI included analytical insight, anti-money-laundering and fraud work, and cybersecurity. Among firms already using AI, 81% employed some form of explainability method.
The right conclusion is therefore not that human judgment should defeat automation. It is that the two are complements at different points in the risk curve. AI can raise the floor by making routine service more consistent and by giving employees better information. Humans can raise the ceiling by handling exceptions, exercising discretion and taking responsibility for choices that cannot be reduced safely to a standard path. A bank that forces people to do machine work will be expensive. A bank that forces machines to do all human work will eventually encounter cases where the cost of the error is larger than the savings from automation.
That complementarity is consistent with the wider labour evidence. The ILO's 2025 global update on generative AI and jobs estimated that one in four workers were in occupations with some degree of GenAI exposure, while concluding that transformation of jobs was more likely than wholesale replacement because many tasks still require human input. Banking is a particularly clear case of that distinction. The technology can absorb drafting, searching, classification and pattern recognition, allowing fewer people to handle more relationships. The strategic question is whether institutions use that productivity dividend to remove people indiscriminately or to move skilled people toward higher-stakes work.
A dangerous outcome: human service becomes a luxury good
There is a distributional risk hiding inside the premium. Wealthy clients may continue to receive senior bankers, private advisers and direct access to specialists, while mass-market customers are routed through increasingly automated funnels. Economically, that segmentation is easy to understand. Human time is expensive and scarce. But if access to a competent person becomes available only to customers with the largest balances, the industry could create a two-tier model of accountability: bespoke explanation for the affluent and automated finality for everyone else.
That would be a mistake for both banks and regulators. A customer dealing with bereavement, coercion, fraud or financial distress may have a low account balance and a very high need for judgment. A small-business borrower can be commercially important even if a standard scoring model struggles with the pattern of its cash flows. Digital-first fintechs face the same challenge: a branch network is not required to provide meaningful human escalation. What matters is whether the customer can reach someone with authority before the situation becomes a complaint, a loss event or a reputational crisis.
This is where customer-protection rules and competitive strategy begin to overlap. A baseline of accessible human support may be a regulatory expectation in some markets, while faster access to specialised judgment can still be a commercial differentiator above that floor. The mistake is to equate 'human premium' with 'paywall'. The better interpretation is that every institution needs a credible human backstop, while the depth and sophistication of relationship service can vary by product and segment.
From human-in-the-loop to human-on-purpose
Banks will need to redesign roles around that principle. 'Human-in-the-loop' sounds reassuring, but it can describe a weak control if the person merely clicks approve on a machine recommendation. Human-on-purpose is a tougher standard: the institution defines in advance which decisions require judgment, what evidence the employee must see, when they can override the model, how the override is recorded, and how management learns from repeated exceptions. The goal is not to maximise human intervention. It is to make intervention meaningful.
That requires better data as much as better people. The BIS Financial Stability Institute's March 2026 work on AI data use argues that privacy, quality, security, third-party dependency and service-provider concentration become more acute as advanced AI spreads through finance, and that managing those risks is essential to preserve trust, resilience and financial stability. A human decision-maker who sees incomplete or distorted information cannot rescue a weak system. The premium depends on a chain: reliable data, understandable model output, clear authority and employees trained to use both without surrendering judgment.
The workforce implications follow. Banks may need fewer employees whose primary task is to retrieve information, copy data between systems or apply simple rules, but more people who can investigate anomalies, communicate under stress, understand products across silos and challenge automated recommendations. Those are not merely 'soft skills'. They are control skills. In a highly automated bank, the most valuable employee may be the one who can detect that the machine's answer is technically plausible but contextually wrong.
Boards and investors should measure trust as an operating variable
The financial case for AI will often be presented through cost-to-income ratios, faster processing and lower headcount. Those measures matter, but investors should also ask what happens to the cost of failure. Does greater chatbot containment lead to more repeat contacts? Are fraud false positives trapping customers in longer resolution cycles? How often do humans override credit or transaction models, and how often are those overrides later shown to have prevented harm? Are complaints becoming concentrated around automated journeys? Is customer attrition higher after adverse decisions that cannot be explained clearly?
Those questions sit close to the emerging supervisory agenda. The Financial Stability Board's June 2026 consultation on responsible AI adoption proposes 12 sound practices spanning organisation-wide governance, AI development and deployment, cyber risk and third-party risk. It is explicitly a consultation rather than a binding international standard, but its intended audience includes boards and senior management. That is significant. AI governance is moving out of the innovation lab and into the same category as other issues for which leadership is expected to understand materiality, controls and risk ownership.
Supervisors outside Europe and North America are making similar points. In a May 2026 address, South African Reserve Bank Deputy Governor Fundi Tshazibana noted that survey evidence on financial institutions' AI governance showed risk-management frameworks were common but weaknesses remained in areas such as fairness testing, third-party model risk management and board-level oversight. The geography changes; the governance problem does not. Financial institutions can outsource models, infrastructure and data services, but they cannot outsource the accountability customers and supervisors ultimately attach to the regulated entity.
The premium will be earned in moments of doubt
Banking has always sold confidence, even when the product appeared to be a deposit, a loan, a payment or an investment account. AI does not overturn that business model. It makes the underlying promise more visible. If a customer believes the bank can act instantly but cannot explain itself, speed becomes unsettling. If a customer believes a bank has sophisticated fraud controls but no one can help when a legitimate payment is blocked, automation becomes friction. If a wealth client receives endless machine-generated advice but no one is willing to say, 'Do not do this,' information becomes noise.
The human premium is therefore unlikely to show up as a simple fee called 'trust'. It will appear indirectly: in retention after difficult events, in the willingness of customers to consolidate assets, in lower complaint escalation, in better recovery from errors, in a bank's ability to serve complex clients and in the credibility of its explanations when regulators ask why an automated system behaved as it did. Some institutions will capture it through premium relationship models. Others will capture it by making ordinary customers feel that an automated bank still has a responsible person behind it.
The winners will not be the banks that keep the most humans, nor the ones that automate the most aggressively. They will be the institutions that make the boundary between machine and human intelligence feel deliberate. Software should handle what is repeatable. People should handle what is consequential, ambiguous or emotionally charged - and they should be equipped to challenge the software rather than merely endorse it. As AI makes intelligence cheaper, that capacity to exercise accountable judgment may become one of the few banking advantages that cannot be copied by simply buying the same model.
Conclusion
The future bank may be more automated than any institution customers have known, but that does not imply a future without human value. It implies a different kind of human value. Routine service will increasingly be generated, predicted and executed by software. Trust will be won when something unusual happens: a model is wrong, a customer is frightened, a borrower does not fit the template, a fraud case is ambiguous, or a market move makes a technically rational action emotionally difficult. In those moments, the most valuable product a bank can offer may be a simple promise: the technology can act fast, but a competent person remains accountable when it matters.
References
1. Artificial intelligence in UK financial services - 2024. Bank of England and Financial Conduct Authority.
2. Rising application of AI in EU banking and payments sector. European Banking Authority, 2025.
3. More people have bank accounts but one in ten have no cash savings, FCA survey reveals. Financial Conduct Authority; Financial Lives 2024 findings.
4. Consumer support outcome: good practices and areas for improvement. Financial Conduct Authority.
5. In data we trust? Emerging policy and supervisory approaches to AI data use in financial services. BIS Financial Stability Institute, 26 March 2026.
6. Project Noor: explaining AI models for financial supervision. BIS Innovation Hub, 18 August 2025.
7. Sound Practices for Responsible Adoption of Artificial Intelligence (AI): Consultation report. Financial Stability Board, 10 June 2026.
8. Timeline for the Implementation of the EU AI Act. European Commission AI Act Service Desk; current 2026 timeline.
9. Annex III - High-risk AI systems. European Commission AI Act Service Desk.
10. Article 14: Human oversight. European Commission AI Act Service Desk.
11. AI Omnibus enters into force. European Commission, 2026.
12. 12 CFR Part 1002, Section 1002.9 - Notifications. Consumer Financial Protection Bureau; current Regulation B.
13. Generative AI and jobs: A 2025 update. International Labour Organization.
14. Financial Stability Report - July 2026. Bank of England.
15. Regulation and supervision of the financial sector in the age of artificial intelligence. Address by Fundi Tshazibana, South African Reserve Bank, 8 May 2026.