The Future of AI in Financial Services

The question for banks and fintechs is no longer whether to use AI. It is whether the next system of record will take actions on its own. Between 2026 and 2028, the future of AI in financial services shifts from models that answer questions to agents that execute multi-step work: opening accounts, screening transactions, drafting credit memos, and closing the books with limited human input.

That shift is already priced in by the institutions that move first. McKinsey estimates that global banking profit pools could shrink by as much as 10% over the next five to ten years for banks that fail to reinvent their operating models, while AI pioneers open a gap of roughly four percentage points of return on tangible equity over slow movers. The UAE is pushing harder than most markets, with a federal plan to run half of government services on autonomous AI by 2028. This is a forward look at where the technology, the regulators, and the talent question are heading, and what a finance or technology leader should build now.

Agentic AI: AI systems that plan, decide, use tools, and complete multi-step workflows with limited human input, rather than producing a single response to a single prompt. An agent can move work forward on its own; a chatbot can only answer.

From assistive AI to autonomous operations

The AI most banks deployed between 2023 and 2025 was assistive. A relationship manager asked a model to summarize a client, a developer used it to write code, a fraud analyst used it to draft a case note. The human stayed in the driver’s seat and the model handed back text. That phase produced real efficiency, but the bottleneck stayed human. Someone still had to read the summary, make the call, and key the action into a core system.

The 2026 to 2028 phase removes that bottleneck for defined tasks. McKinsey describes agents that act autonomously and inherit the same access rights as the people they work alongside, executing multi-step processes end to end, and suggests an AI agent could become the channel of choice for customer interactions within three to five years.

The adoption data backs the direction. Wolters Kluwer surveyed finance leaders and found that while only 6% currently use agentic AI, a further 38% plan to adopt it within twelve months, putting 44% of finance teams on agentic AI in 2026, an increase of over 600%. Gartner expects the broader software layer to follow, predicting that 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5% the year before.

Where autonomous operations land first in finance

The early wins concentrate where work is repetitive, rule-bound, and high-volume. Transaction monitoring and financial-crime screening are obvious candidates, because the work is pattern matching at a scale humans cannot staff. Account onboarding and KYC refresh are another, since most of the steps are data retrieval and validation. Back-office finance is a third: McKinsey reports a US bank that rebuilt its credit-risk memo process around AI agents and saw a 20% to 60% productivity gain with a 30% improvement in credit turnaround, and notes agentic deployments can reduce manual workloads by 30% to 50%.

The pattern is consistent. Agents do not replace the judgment at the top of a credit decision or a financial-crime escalation. They compress the hours of retrieval, formatting, and reconciliation around it. For a UAE bank pushing acquisition through digital channels, where mobile banking adoption has crossed 83%, the constraint is rarely demand. It is the capacity to onboard, monitor, and service at speed without expanding headcount linearly. That is the gap autonomous operations close.

The UAE is building the rails autonomous finance needs

Forward outlook in this market cannot be separated from what the regulators and the state are building. The UAE has made AI a sovereign priority, not a sector trend. In April 2026 the federal government announced a framework to move 50% of government services to autonomous AI systems by 2028, positioning the country to be the first government to operate at that scale. That sets the tone for what regulators will expect from the private sector and what customers will treat as normal.

Underneath that ambition sits financial infrastructure. The Central Bank of the UAE launched its FIT Programme (Financial Infrastructure Transformation) in 2023, and its components are going live: the Aani instant payments platform is operational, and the Digital Dirham central bank digital currency is in piloting. Federal Decree-Law No. 6 of 2025 unified banking, fintech, and insurance under a single regulatory framework. These are the rails that AI-driven finance will run on. An agent that moves money, settles a payment, or issues a token needs a real-time, programmable infrastructure underneath it, and the UAE is building exactly that.

CBUAE FIT Programme: the Central Bank of the UAE’s Financial Infrastructure Transformation programme, launched in 2023, which delivers the country’s instant payments platform (Aani), the Digital Dirham, and supporting open-finance and digital-identity infrastructure.

For an institution planning to 2028, the implication is concrete. The AI roadmap and the payments-and-core roadmap are the same roadmap. Autonomous agents will not work on a batch-processed legacy core that settles overnight when the surrounding infrastructure runs in real time. This is why fintech app development in the UAE and core modernization are now AI questions, not just engineering ones. The work Kentro does across core banking modernization, fintech product development, and AI/ML converges on a single requirement: a system of record that an agent can read from and write to safely.

The regulatory response is arriving with the technology

Autonomous systems that take financial actions will not get a quiet rollout. The harder the agent acts, the more a regulator wants to know how it decided. Deloitte’s 2026 outlook for financial institutions in the region names transparency and explainability as the leading adoption hurdle, with more than half of surveyed institutions citing it, precisely because the methods are moving from simple machine learning to entire agentic workforces. The same outlook reports that only around one in five organizations has a mature governance model for autonomous AI, which is the gap regulators will press on.

Data protection is part of this. The UAE’s PDPL (Personal Data Protection Law, Federal Decree-Law No. 45 of 2021) governs how personal data is processed, stored, and moved across borders. An AI agent that pulls a customer’s transaction history to make a lending decision is processing personal data at machine speed, and the lawful basis, residency, and auditability of that processing have to hold up. Build the agent without that, and the compliance failure surfaces later, under examination, at the worst possible time.

Why most agentic projects will fail, and how to be in the minority

The forward outlook is not uniformly bullish, and pretending otherwise would be dishonest. Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. That number is the most useful figure in this entire piece, because it tells you the failure mode in advance.

The cancelled projects share a profile. They start from the model, pick an impressive demo, and only later find that the underlying data is inconsistent, the audit trail is incomplete, or no one owns the risk when the agent gets it wrong. A single inconsistent customer record, multiplied across thousands of automated decisions a day, produces error volume no oversight team can absorb. The technology was rarely the problem. The data foundation and the governance were. The surviving 60% invert the order: fix identity, master data, and audit logging first, scope the agent to a bounded task with a clear owner, keep a human in the loop on consequential decisions, and measure value against a baseline before scaling.

Talent and data foundations decide the outcome

The constraint on autonomous finance through 2028 is not access to models. Frontier models are a commodity any bank can buy. The constraint is the foundation underneath them and the people who can build and supervise it.

On data, the requirement is unglamorous and decisive. Agents act on the system of record, so it has to be clean, consistent, and queryable in real time. In the Wolters Kluwer survey, 44% of finance leaders named data readiness as the single biggest driver of AI adoption, ahead of any model or tooling concern. The bottleneck is data, not intelligence. An institution that spends 2026 cleaning master data and instrumenting its core for real-time access can deploy agents in 2027. One that skips that step will be in the cancelled 40%.

On talent, the shift is from people who use AI to people who supervise it. The roles that matter can specify what an agent should and should not do, design the human-in-the-loop checkpoints, and read the agent’s reasoning when a regulator asks. In the region, Deloitte expects Arabic-optimized agents to proliferate, which adds a localization layer most global tooling does not handle out of the box. This is where a regional implementation partner earns its place: not in buying the model, but in building the data foundation, the controls, and the integration into local infrastructure that turns a model into a system a bank can run. Kentro’s delivery across banking, fintech, and enterprise is built around that integration problem rather than the model itself.

What to do in the next twelve months

The honest near-term plan is narrow. Pick one bounded, high-volume, rule-heavy workflow where an error is recoverable, not catastrophic, such as KYC refresh or a defined back-office reconciliation. Audit the data that workflow depends on and fix it before touching an agent. Define who owns the risk, what the human checkpoint is, and how you will measure value against today’s baseline. Then deploy, measure, and only scale what clears the bar. The institutions that compound an advantage by 2028 are the ones that do this quietly in 2026, not the ones that announce a flagship project and kill it a year later.

The future of AI in financial services is autonomous, regulated, and grounded in data foundations that most institutions have not yet built. The technology is moving faster than the governance, the UAE is moving faster than most markets, and the gap between leaders and laggards is widening into a return-on-equity difference that is hard to close later. The work to be in the leading group is available now, and it starts with the foundation, not the model.

Frequently asked questions

What is the difference between agentic AI and the AI banks already use?

Most AI deployed in banking through 2025 was assistive: it produced text or analysis and handed it to a human, who made the decision and took the action. Agentic AI plans and executes multi-step work on its own, such as completing a KYC refresh or drafting and filing a credit memo, with a human supervising rather than driving. The difference is autonomy. An assistant answers; an agent acts. That is why governance, data quality, and audit trails matter far more for agentic systems than they did for assistive ones.

How much does an agentic AI or autonomous finance project cost?

There is no fixed number, because the cost is driven by the state of your data and core systems, not by the AI itself. A project on a clean, real-time system of record with one bounded workflow is a different scale of engagement from one that first has to remediate master data and modernize a legacy core, and model licensing is usually the smallest line item. The right way to get a real figure is to scope the specific workflow and the data foundation it depends on, which is what a strategy call is for.

How long does it take to deploy an AI agent in a regulated financial workflow?

For a single, well-scoped workflow on data that is already clean, a first production deployment with human-in-the-loop controls is typically a matter of a few months. When the data needs remediation or the core needs real-time access first, that foundation work usually takes longer than the agent itself, pushing a realistic timeline across two phases over six to twelve months. The honest answer depends on how ready your data and infrastructure are.

What return should we expect from agentic AI in financial services?

Returns depend on the workflow and your baseline, so a fixed multiple would be misleading. Public examples are instructive rather than guaranteed: McKinsey reports a US bank that saw a 20% to 60% productivity gain on credit-risk memos, and notes agentic deployments can cut manual workloads by 30% to 50%. The realistic way to forecast your own return is to measure the current cost of a target workflow, scope the agent against it, and compare. Returns are real where the workflow is high-volume and rule-bound, and thin where it is low-volume or judgment-heavy.

Why do so many agentic AI projects fail?

Gartner predicts more than 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. The common pattern is starting from an impressive demo and discovering too late that the underlying data is inconsistent, the audit trail is incomplete, or no one owns the risk when the agent errs. The projects that survive fix data, identity, and governance first, scope the agent to a bounded task, and keep a human in the loop on consequential decisions.

How does the UAE regulatory environment affect autonomous AI in finance?

The UAE is moving faster than most markets, with a federal target to run 50% of government services on autonomous AI by 2028 and the CBUAE FIT Programme building real-time rails such as Aani and the Digital Dirham. That raises expectations for the private sector. At the same time, the PDPL governs how agents process personal data, and regulators increasingly expect transparency and explainability for autonomous decisions. The practical effect is that autonomous finance in the UAE has to be built with auditability, data residency, and human oversight designed in from the start, not added after an examination.