Wednesday, 16 September 2026
Artificial intelligence is rapidly entering banking and payments, but financial institutions are discovering that deploying the technology is considerably more complicated than demonstrating that it works. Fraud detection, credit assessment, merchant services, customer support and payment operations are all becoming more intelligent, yet some of the biggest barriers to adoption increasingly concern accountability, regulation, data quality and trust rather than the capabilities of the underlying AI models. That tension emerged during a discussion at AI4 2026 in Las Vegas involving executives from banking, B2B payments and merchant payment technology. Their experiences illustrated how differently AI adoption can develop depending on whether an organisation is an established regulated bank, a payments group or a smaller financial technology company.
Across these businesses, the potential applications are substantial. AI can identify suspicious transactions, accelerate credit decisions, automate administrative work for merchants and provide increasingly personalised financial services. Finance, however, operates under constraints that do not exist to the same degree in many other industries. A retailer can tolerate an imperfect product recommendation. A bank has much less room for error when an automated system influences a credit decision, payment or piece of financial guidance. Financial AI therefore increasingly has to prove not only that it works, but that its decisions can be controlled, reconstructed and assigned to an accountable organisation or individual. The industry is consequently being pulled in two directions: financial companies are under pressure to adopt AI more quickly while regulators, boards and customers expect increasingly rigorous controls over how it operates.
One of the most immediate opportunities is merchant technology. Small and medium-sized businesses frequently spend considerable time managing menus, changing prices, configuring promotions and navigating administrative systems. AI-assisted tools can simplify these processes, allowing merchants to interact more naturally with their operating technology instead of moving repeatedly through conventional software menus. Payment devices themselves are also becoming more intelligent. Smaller AI models can increasingly operate locally on handheld terminals and other edge devices, potentially allowing merchants to make operational changes through natural-language instructions. The same hardware could eventually support local fraud detection and other functions without sending every task to a remote cloud environment. This represents an important direction for financial AI because many of the most commercially valuable applications may remain largely invisible to customers. The objective is not necessarily to create another chatbot. It is to remove administrative work from businesses whose primary objective is serving customers rather than operating payment software.
A similar development is taking place in B2B payments. Receiving money is only one part of a transaction. Payments must still be identified, reconciled, recorded and connected with accounting and other operational systems. AI agents could complete more of that work automatically, moving the industry towards a model in which a payment is effectively finished only when the business no longer needs to perform additional administrative tasks around it. Banking presents a more complicated opportunity. AI-powered credit assessment can accelerate underwriting and potentially help institutions evaluate customers who have limited conventional borrowing histories. Historical lending datasets can contain gaps because they reflect previous lending behaviour. If particular demographic or economic groups historically borrowed less frequently, traditional models may have less information with which to assess them. More advanced modelling and carefully governed synthetic data could help institutions broaden the information used in credit assessment. The potential benefit is not simply faster lending decisions but the possibility of identifying economically viable borrowers who might otherwise struggle to qualify through traditional scoring methods.
The consequences require rigorous supervision. Credit decisions directly influence access to capital, making explainability, testing and governance essential. AI may discover relationships that conventional rules overlook, but institutions still need to demonstrate that decisions are fair, defensible and compliant with applicable regulation. Fraud provides perhaps the clearest example of AI’s double-edged impact on finance. Machine-learning systems can examine transaction histories, devices, merchant behaviour and numerous other signals simultaneously. This can improve the detection of suspicious activity while reducing false positives that unnecessarily interrupt legitimate transactions. The ideal fraud system operates almost invisibly. Genuine customers complete transactions without additional friction while genuinely suspicious behaviour is escalated for further investigation. AI can improve that process because it can identify relationships across much larger and more complicated datasets than traditional static rules.
The same technological progress, however, is increasing the capabilities available to criminals. Generative AI has reduced the effort required to produce convincing phishing communications, impersonation attempts and other forms of social engineering. Synthetic voices, images and video can make fraudulent communications considerably more persuasive than the poorly written scams consumers and employees previously learned to recognise. Financial institutions are therefore deploying AI against adversaries who increasingly possess similar technological capabilities. This contest between AI-enabled fraud and AI-enabled detection could become one of the defining cybersecurity challenges facing banking and payments. The weakest point may continue to be the person at the end of the process. An organisation can deploy sophisticated fraud detection, but an employee or customer who believes an apparently authentic communication can still authorise an action that bypasses those protections. AI consequently increases the importance of combining technological detection with stronger identity verification and human awareness.
Governance becomes even more complicated inside financial institutions themselves. Employees are already using public generative AI tools for drafting emails, analysing information and preparing documents, sometimes without formal corporate approval. That creates a significant problem for banks because seemingly ordinary business communications can contain commercially sensitive information even when they include no customer names or account details. Attempting to prevent employees from using AI altogether is increasingly unrealistic. The productivity advantages are becoming too visible and consumer tools remain easily accessible. Financial institutions therefore face a more practical challenge: providing controlled alternatives that allow employees to obtain the benefits without exposing confidential information outside approved systems.
This requires understanding an organisation’s complete AI exposure. A bank may control applications used directly by employees while having considerably less visibility over technology embedded inside third-party products. Vendors can use subcontractors, which may themselves depend on AI services supplied by other companies. The financial institution can nevertheless remain responsible for what ultimately reaches its customers. AI risk therefore does not stop at the corporate boundary. Banks increasingly need to understand where artificial intelligence exists throughout their technology and supplier chains, what information those systems process and how their outputs are controlled. The cost of implementing AI consequently extends far beyond model subscriptions or computing resources. Institutions need governance frameworks, security controls, data infrastructure, testing, monitoring, audit systems, employee training and potentially substantial changes to existing workflows. These costs can fundamentally change the return on investment.
For highly regulated organisations, this can create a rational incentive to move cautiously. If a process can continue being performed manually at an acceptable cost, management may prefer additional human resources to an AI system requiring expensive governance while exposing executives and directors to uncertain regulatory consequences. This is one reason AI adoption in finance may progress differently from industries where experimentation carries fewer consequences. A technology can be demonstrably more capable and still fail the institution’s overall risk-adjusted investment test.
Compliance therefore needs to move earlier into product development. Historically, technology teams could design a product before asking risk and compliance departments how it should be controlled. AI makes that sequence increasingly inefficient. Governance needs to be incorporated while systems are being designed rather than added after development. One approach is an organisation-wide AI charter defining the boundaries within which employees and technology teams can experiment. Instead of approving every idea independently from the beginning, companies can establish common principles covering data, accountability, human oversight, security and acceptable use. This can create controlled room for innovation without removing responsibility. It also gives technology, product, legal, risk and compliance teams a common framework rather than forcing every project to renegotiate the same fundamental questions.
Another potentially effective strategy is to begin AI adoption with risk and compliance functions themselves. If AI can demonstrate measurable benefits to the executives responsible for creating organisational guardrails, those executives may become more comfortable supporting carefully controlled deployment elsewhere in the business. The approach effectively turns compliance from a potential obstacle into an internal participant in innovation. Instead of technology teams attempting to persuade risk departments to tolerate new systems, both sides develop experience with the technology together.
Accountability nevertheless remains difficult. If an AI system makes an incorrect decision, organisations need to know who is responsible. Simply attaching an executive’s name to the system does not solve the underlying problem if the organisation cannot reconstruct why the decision occurred. Financial AI consequently requires data lineage, monitoring and auditability. Institutions need records showing which information entered a system, which model or process was used, what action was recommended or executed and where human approval occurred. The importance of these controls increases as AI progresses from providing information towards taking actions. An assistant answering a question creates one level of risk. An autonomous agent transferring money or modifying a financial account creates another.
Financial AI adoption may therefore develop through a gradual increase in autonomy. Initially, an AI system receives permission to read information. Once users become comfortable with its understanding of that information, it can provide recommendations. The next stage allows the system to prepare actions while requiring explicit approval. Only after sufficient confidence has been established does the AI receive authority to execute predefined actions independently. Such a progression could prove particularly important in consumer banking. Customers may become comfortable allowing an AI system to analyse their finances long before they are willing to let it transfer money or make investments without approval.
Financial advice presents an even greater challenge. Consumers increasingly expect digital services to understand their circumstances and provide personalised recommendations, but banks face considerable liability when automated systems provide inaccurate information. The widely reported Air Canada chatbot dispute demonstrated the wider principle. A Canadian tribunal held the airline responsible after its automated service supplied incorrect information about bereavement fares. For financial institutions, the implication is significant: companies cannot necessarily distance themselves from statements generated by automated systems simply because no employee personally produced them.
For banks, insurers and payment providers, accuracy therefore becomes a commercial requirement as well as a technical measure. An incorrect restaurant recommendation may inconvenience a consumer. Incorrect information concerning a mortgage, investment or payment can create direct financial losses and potential legal exposure. This helps explain why financial institutions remain cautious about highly autonomous customer-facing AI despite rapid technological improvements. Trust needs to be earned progressively, while tolerance for mistakes becomes smaller as the system receives greater authority.
Customer expectations, however, are moving in the opposite direction. Consumers and merchants increasingly interact with AI systems capable of understanding natural language and responding almost immediately. They consequently expect financial technology to become more personalised and intuitive as well. Traditional financial services have largely operated on a one-to-many model in which one product, interface or process serves thousands or millions of customers. AI makes something closer to individualised service economically possible. Systems can potentially adapt recommendations, interfaces and workflows according to individual circumstances.
For merchants, this might involve recommendations concerning pricing, staffing, promotions or operating schedules based on transaction patterns. For consumers, it could eventually mean financial services that respond dynamically to spending behaviour, income, savings objectives and risk preferences. Data becomes the strategic asset underpinning these services. Payment processing itself is becoming increasingly commoditised. The greater commercial opportunity may lie in the intelligence that can be built around the transaction. Companies with visibility across more of the payment journey can potentially understand merchants and customers more deeply. Transaction information can support fraud prevention, operational recommendations, lending decisions and numerous other value-added services.
This creates an important competitive question for the payments industry. The winner may not necessarily be the company capable of moving money at the lowest cost. It could instead be the organisation capable of extracting the greatest useful intelligence from transactions while maintaining customer trust and regulatory control. Payments themselves could eventually become increasingly agent-driven. As AI agents begin initiating transactions on behalf of consumers and companies, financial infrastructure will need to determine who authorised each payment, which agent initiated it and what limits were placed on that authority.
Existing card networks may support part of this activity, while new protocols and blockchain-based payment systems could address other requirements. Rather than completely replacing today’s payment infrastructure, an intelligent layer could develop above existing rails to determine how, when and through which mechanism transactions occur. This places identity and authorisation at the centre of agentic finance. An autonomous agent needs more than access to money. It needs clearly defined permission to spend it.
AI is also beginning to change internal productivity, although not always in the ways companies expected. Employees can now generate detailed reports, proposals and presentations extremely quickly. This reduces the cost of producing information while potentially increasing the cost of reviewing it. The result can be a new organisational bottleneck: large volumes of highly polished AI-generated material still need senior executives to determine whether the underlying analysis is useful. Instead of eliminating bottlenecks, AI can simply move them further up the organisation.
This could become an important management challenge. When producing a lengthy report requires minutes rather than days, the scarce resource becomes the attention of the people responsible for evaluating it. Organisations will increasingly need to distinguish between greater content production and genuinely better decision-making. Other productivity improvements are more straightforward. AI systems connected securely to approved corporate information can prepare meeting summaries, management updates and email digests using work employees have already completed. Administrative tasks that previously consumed hours can potentially be reduced to several minutes of review.
The difference is access to organisational context. A generic model knows relatively little about what an employee accomplished during the week. An enterprise AI system securely connected to meetings, messages, documents and workflows can reconstruct much of that information automatically. This points towards another potential transformation in financial organisations: the development of a shared intelligence layer across the company. Instead of employees individually using disconnected AI applications, approved systems could draw on authorised corporate information and make relevant knowledge available across teams.
For banks, achieving this will be considerably harder because information cannot simply become universally accessible. Customer confidentiality, cybersecurity, regulation and internal controls require strict boundaries around who can access particular information. The direction nevertheless appears increasingly clear. AI is moving from being an application employees occasionally open towards becoming an intelligence layer embedded throughout financial operations.
Competitive advantage is therefore unlikely to come solely from possessing the most advanced AI model. Financial institutions generally have access to many of the same underlying technologies. Differentiation is more likely to come from proprietary data, the architecture connecting it, governance surrounding the system and the organisation’s ability to convert AI outputs into reliable actions. This makes trust more than a regulatory requirement. It becomes a competitive asset.
Consumers must trust that an AI system understands their financial circumstances. Merchants must trust automated recommendations affecting their businesses. Regulators must trust that institutions can reconstruct automated decisions. Boards must trust that management understands the risks being accepted. Employees must trust that AI is improving their work rather than simply creating another layer of monitoring or bureaucracy. The companies that solve these problems could move significantly faster than competitors because each successful deployment increases confidence in the next level of autonomy.
The future of AI in finance may therefore develop less as a sudden transition towards autonomous banking and more as a gradual expansion of delegated authority. Systems will first observe, then recommend, then prepare actions and eventually execute selected decisions within clearly defined limits. That progression may appear slower than some of the more dramatic predictions surrounding artificial intelligence, but in financial services it is likely to prove more durable. Money ultimately depends on confidence.
AI can make finance faster, more personalised and potentially safer. But as financial institutions give machines increasing influence over credit, payments and customer decisions, the most valuable capability may not be intelligence itself. It may be proving that the intelligence can be trusted.
Source: CIJ.World Research & Analysis Team
