The financial services industry is currently standing on the precipice of its most significant technological shift since the introduction of electronic trading.
For the past two years, the narrative has been dominated by Generative AI – large language models (LLMs) capable of summarising reports, drafting marketing copy, and powering increasingly conversational chatbots.
While impressive, these applications have been mainly assistive. They help humans do existing tasks slightly faster.
However, a more profound transformation is underway, one that moves beyond “generative” to “agentic”.
What’s the meaning of agentic AI?
Agentic AI refers to systems that do not merely generate text but take autonomous action to achieve specific goals. These agents can plan, reason, execute multi-step workflows, and interact with other software systems without constant human supervision.
In the world of finance – a sector defined by rules, protocols, and vast flows of structured data – this shift from passive assistant to active agent is revolutionary.
We are witnessing the birth of the “Agent-to-Agent” (A2A) economy, where the friction of human mediation is systematically removed from operational loops.
From the back offices of corporate treasury departments to the high-stakes crucible of high-frequency trading, the role of the human is evolving from operator to orchestrator.
The operational pivot: beyond the chatbot
The popular imagination of AI in finance often defaults to the “robo-advisor” or the customer service chatbot. While these are visible, they are merely the tip of the iceberg.
The true tectonic shifts are occurring deep within the operational layers of financial institutions – in the murky, unglamorous world of reconciliation, invoicing, and cross-border settlements.
Jared Shulman, Co-founder and CEO of Daylit, a company at the forefront of this operational shift, argues that the industry has been looking in the wrong direction.
“The most transformative application of agentic AI in fintech won’t be the flashy consumer-facing chatbots – it will be the quiet revolution happening in operational communications. Think about the millions of daily exchanges around payment timing, delivery status, invoice disputes, and reconciliation. These interactions are ripe for a fundamental shift.”
Shulman outlines a clear trajectory for this technology, describing a “three-phase evolution” that mirrors the broader maturation of enterprise AI.
“Today, most B2B financial communications remain human-to-human. An accounts receivable specialist emails a customer about a late payment, waits for a response, then manually updates systems. It’s slow, inconsistent, and doesn’t scale. The emerging phase is AI-powered human communication. AI agents now gather full context – invoice history, delivery documentation, prior email threads, ERP data – and draft appropriate responses. The human reviews, adjusts tone if needed, and sends. This isn’t about replacing judgment; it’s about eliminating the friction of context-gathering and response drafting.”
This second phase is already yielding significant dividends. According to a recent McKinsey & Company report on banking, integrating agentic capabilities into operational functions is projected to reduce costs by 15–20% in the most likely adoption scenarios.
By automating the “gathering” phase of work, teams can redirect their cognitive load toward decision-making.
However, the third phase Shulman describes represents the true leap forward.
“But the real transformation comes in phase three: agent-to-agent communication. With protocols like Google’s A2A and Anthropic’s MCP establishing interoperability standards, we’re building the infrastructure to enable agents to communicate directly across organisations. When your agent can query my agent about invoice status, receive a structured response, and update both systems automatically – without either of us touching a keyboard-the entire velocity of B2B financial operations changes.”
Aggressive market forecasts support this vision. Gartner projects that by the end of 2026, 40% of enterprise applications will embed AI agents, a staggering rise from less than 5% in 2025. This suggests that the infrastructure for this “quiet revolution” is being laid at a pace that few organisations have fully priced in.
The US$100 billion social gap
While B2B operations focus on velocity and efficiency, Agentic AI in the public sector focuses on accessibility and equity.
The financial lives of the most vulnerable are often governed by a labyrinth of bureaucratic forms, eligibility criteria, and fragmented state systems that act as barriers rather than gateways.
Shreenath Regunathan, Founder and CEO of Starlight, highlights the immense scale of this inefficiency.
Over $100 billion in government benefits goes unclaimed every year. Many Americans don’t know what they qualify for – from heating assistance to childcare subsidies – and those who are eligible often give up trying to navigate many fragmented systems and confusing forms.”
This “unfulfilled eligibility,” as it is technically termed, is not merely a rounding error; it represents a massive failure of the social safety net’s delivery mechanism.
In the UK, similar patterns emerge, with billions in pension credit and support going unclaimed annually due to complexity.
Regunathan sees Agentic AI as the solution to this “last mile” problem in public finance.
“Agentic AI can create a bridge for people when they are going through tough times, creating a straightforward process for identifying the right benefits for the right people. In the future, it could take the next step of helping people fill out applications so they can review and submit. This can help millions of Americans access the support they need during difficult times, rather than getting lost in bureaucracy.”
The “bridge” metaphor is apt. Unlike a traditional website that presents information passively, an agent can actively navigate the user’s specific context.
It can ingest a user’s financial data, cross-reference it against thousands of changing eligibility rules across federal, state, and local jurisdictions, and autonomously navigate the application portals. This transforms the citizen’s experience from bewildered searching to guided confirmation.
Real-time unit economics
As we move from social impact to the complex mechanics of profitability, the need for precision becomes acute.
Fintech products, particularly in payments and neobanking, often operate on razor-thin margins.
A fraction of a cent increase in fraud-detection costs or a slight rise in infrastructure latency can wipe out the profitability of a single transaction.
Matt Guarini, Executive Director of the TBM Council, emphasises that the traditional methods of managing these economics are dangerously outdated. He advocates for the integration of Technology Business Management (TBM)-a framework for aligning tech costs with business value-with Agentic AI.
“FinTech products operate at extreme scale and thin margins, where small changes in fraud controls, infrastructure performance, or regulatory overhead can materially impact profitability. Yet most organisations still manage product economics through lagging financial reports. TBM provides a unified model of product-level unit economics, connecting infrastructure, platform, security, and compliance costs to individual transactions, customer segments, and geographies.”
The limitation of TBM historically has been the speed of analysis. It often served as a retrospective microscope. Guarini argues that agents turn it into a real-time gyroscope.
“Agentic AI builds on this model by continuously monitoring performance and cost signals, simulating tradeoffs, and recommending or executing bounded actions to protect margin in real time. Together, TBM and Agentic AI transform product economics from retrospective analysis into an adaptive, continuously optimised system.”
This capability is critical as cloud costs spiral. With global data centre electricity consumption projected by Deloitte to double to 1,065 TWh by 2030, the price of the compute underlying these fintech products is volatile.
An agent that can autonomously balance workload distribution to optimise for spot-pricing or energy efficiency, while maintaining compliance, becomes a direct contributor to the bottom line.
The compliance portfolio
The second pillar of Guarini’s framework addresses one of the largest drains on capital in finance: compliance.
Traditionally, investment in Anti-Money Laundering (AML) and Know Your Customer (KYC) technology is viewed as a “grudge purchase “-a necessary cost of doing business that offers no return on investment other than staying out of prison.
Guarini suggests a paradigm shift where compliance is managed as an active investment portfolio. “FinTech firms face constant pressure to invest in fraud prevention, AML, and compliance technologies without clear visibility into how those investments reduce risk or enable growth. The result is often over-investment in overlapping controls or under-investment until regulatory action forces change.”
By utilising Agentic AI, firms can move beyond static rule-sets to dynamic simulation.
“Agentic AI uses this foundation to simulate regulatory change scenarios, identify redundancies, and recommend optimised investment paths that balance cost, risk reduction, and business outcomes. Together, they enable leaders to manage compliance as a strategic value portfolio rather than an unavoidable cost burden.”
This mirrors the findings of a 2025 Grand View Research report, which noted that fraud detection agents already account for the largest revenue share (33.8%) of the AI agent market in financial services.
The industry is already voting with its wallet, recognising that static rules cannot catch dynamic criminals. Only an agent that can reason about intent and behaviour can hope to keep pace.
DeFi and the governed economy
The final piece of the TBM Council’s vision concerns the often-chaotic world of Decentralised Finance (DeFi) and blockchain.
To date, institutional adoption has been hampered by a lack of defensible economic models. It is difficult to explain to a board why a transaction should be moved to a blockchain when “gas fees” (transaction costs) are volatile and unpredictable.
Guarini argues that agents provide the necessary governance layer.
“Despite their promise, blockchain and DeFi initiatives often stall because execfutives lack a clear, defensible view of their true economics, risks, and comparative value versus traditional financial infrastructure… Agentic AI continuously evaluates network conditions, transaction economics, security events, and regulatory signals, simulating when and where decentralised approaches outperform centralised alternatives.”
This creates a hybrid future where the choice between a SWIFT transfer and a stablecoin settlement isn’t ideological – it’s an automated, economic decision made by an agent in real-time, based on the specific cost/speed/risk parameters of that exact moment.
“Together, TBM and Agentic AI turn blockchain and DeFi from speculative experiments into governed, economically rational options that can be scaled with confidence.”
The unforgiving crucible
Perhaps the most intellectually demanding frontier for Agentic AI lies in the financial markets themselves.
Unlike B2B invoicing or benefits applications, which, while complex, obey fixed rules, markets are adversarial. They react.
Michal Prywata, Co-Founder of Vertus, draws a sharp distinction between the “forgiving” environments where most AI operates and the brutal reality of capital markets.
“The interesting thing about agentic AI in finance isn’t autonomy. Perhaps financial markets are the only environment that actually forces AI to think. For the moment, most AI is operating in what I’d call “forgiving environments”.
Wrong answer? Let’s give it another go. Hallucination? User notices something is off and corrects it. Markets are very directly unforgiving in a specific way: every decision is validated instantly. Either the reasoning was sound, or the capital’s gone.”
This “instant validation” creates a training pressure cooker that is absent in other domains. If an LLM writes a mediocre poem, there is no immediate penalty. If a trading agent misinterprets market sentiment, the P&L impact is immediate and irreversible.
Prywata notes that this forces a move away from “probabilistic confidence scores”-the standard metric of LLMs-toward genuine reasoning.
“Our systems transact over a billion dollars daily, and you can’t really hide behind probabilistic confidence scores. It’s just not enough. We aren’t asking AI to craft a witty LinkedIn post. Here’s what we’ve learned: agentic AI gets interesting when the feedback loop is both immediate and irreversible. Training on static data produces excellent pattern matching. But operating and learning from decisions with direct consequences produces something closer to actual reasoning.”
This distinction between pattern matching and reasoning is the holy grail of AI research. Pattern matching looks at the past to predict the future.
Reasoning models the causal mechanics of the system. In a market where every other participant is also using AI, pattern matching fails because the patterns themselves shift the moment they are exploited.
“The agent has to model not just “what usually happens” but “what happens when everyone else knows I’m acting on what usually happens. I suppose that is the real frontier right now. In finance and beyond, of course. Agents that can reason in environments where the environment itself is intelligent. Finance happens to be a very high-stakes lab for this kind of intelligence.”
The interoperability imperative
Underpinning all these developments, from Shulman’s operational agents to Prywata’s trading systems, is the need for interoperability. An agent that lives in a silo is of limited utility.
True exponential gains occur when agents can communicate with one another.
Shulman referenced protocols like Google’s A2A and Anthropic’s MCP (Model Context Protocol). These are effectively the “diplomatic language” of the new AI economy. They allow an agent running in Salesforce CRM to securely query an agent running in SAP ERP, negotiate a timestamp, and execute a reconciliation task.
Without these standards, we risk building a “Tower of Babel” where powerful agents are trapped within walled gardens.
The financial institutions that will win in the next decade are those that architect their systems not just for human usage, but for agent accessibility. This means exposing data via APIs that agents can query, structuring documentation in machine-readable formats, and defining clear “trust boundaries” for automated decision-making.
From operator to orchestrator
What does this mean for the human workforce? The fear of displacement is natural, but experts cite a shift in roles rather than a loss of relevance.
As Shulman noted:
“The human role evolves from operator to orchestrator. We’ll spend less time drafting payment reminder emails and more time setting the policies, escalation thresholds, and trust parameters that govern how our agents interact with the world.”
In the banking context, McKinsey’s analysis suggests that this orchestration role will lead to a 30% increase in productivity.
Bankers will not cease to exist; they will stop doing the “robotic” work that dominated their days-data entry, form chasing, compliance box-ticking-and focus on high-touch relationship management and strategic oversight.
In the trading context, the human becomes the algorithm’s risk manager, defining the “guardrails” within which the reasoning agent operates. The skill set shifts from execution to governance.
The agent-to-agent economy
The trajectory is clear. The Agentic AI market is forecast to grow at a Compound Annual Growth Rate (CAGR) of over 46% through to 2030, reaching a valuation of over $50 billion. But the dollar figure belies the structural change.
We are moving from an era of digitisation, where we turned paper forms into PDFs, to an era of cognition, where software can reason about the data it holds.
As Jared Shulman concluded:
“The organisations that recognise this shift – and build for agent interoperability now – will have significant advantages as the agent-to-agent economy emerges.”
Whether it is closing the $100 billion gap in social benefits, protecting the margins of a neobank, or navigating the volatility of global markets, the agent is no longer coming. It is here. And it is ready to work.
The bottom line
The “quiet revolution” of Agentic AI is rewriting the operating system of global finance. It is stripping away the friction of the past century, resolving the inefficiencies of bureaucracy, and introducing a new form of machine reasoning to the markets.
For leaders in the fintech space, the mandate is no longer to ask “what can AI generate for me?” but “what can AI do for me?” The answer, increasingly, is “almost anything.”