Artificial intelligence in finance has passed through three phases. Rule-based automation. Generative AI that drafts and summarizes on command. And now agentic AI: systems that plan, decide, and act with little human oversight.

The AI in finance industry is no longer experimenting at the edges. A 2026 Cambridge survey of 628 institutions across 151 jurisdictions found 42% are already using or assessing agentic AI, while only 21% have put it into production. That gap is the story.

Ask how AI has been used in finance until now and the honest answer is: in the back office. Four of the top five live use cases are internal. Agentic AI is what changes that.

What is agentic AI?

Two independent sources, a Moody’s analysis and an academic survey, land on the same definition. Five traits:

• Autonomy, reasoning and planning. The system decides and acts without step-by-step instruction. It also deliberates before acting. A chatbot answers a question. An agent decides what to do next.
• Adaptability. Learns from feedback and shifts strategy without retraining.
• Collaboration and tool use. Coordinates with other agents, APIs, and systems to run a workflow end-to-end.
• Long-term goals. Pursues an objective across a whole process, not a single transaction.

Moody’s own numbers show the effect. Its Research Assistant users consumed 60% more research and cut task time by 30%. Over 90% of interactions shifted to analytical work. The prize is not speed. It is human time moved toward judgments machines cannot make.

AI applications in finance

Here artificial intelligence applications in finance stop being theoretical.

Governance and risk management

McKinsey’s 2026 research found nearly two-thirds of organizations name security and risk (not regulation, not technical limits) as the top barrier to scaling agentic AI. The counterpoint from the same work: firms investing $25 million or more in responsible AI show higher maturity and are far likelier to see EBIT impact. Governance is not a brake, but a precondition.

Our experience at Vstorm shows that firms address this with explainable AI models, majority-voting across models so no single blind spot rules an outcome, and traceability that logs an agent’s reasoning, not just its answer.

Artificial Intelligence in financial markets: multi-agent trading

Agentic AI treats trading as a game between learning agents, not one optimization. In backtests on 100 Shenzhen Stock Exchange stocks from 2018 to 2021, a system pairing multi-agent reinforcement learning with portfolio-insurance constraints returned 7.76% a year, a 2.18 Sharpe ratio, a 6.60% maximum drawdown.

This beats a plain reinforcement-learning baseline on every measure. At Vstorm, we build the risk control into how the agent acts, not into a check after the fact, and the numbers improve.

Portfolio management

Portfolio management is moving from fixed rebalancing rules to agents that read regime changes and adjust, sometimes across agents with different risk appetites.

Position sizing follows: newer frameworks train agents on Sharpe ratio, profit and loss, drawdown, and turnover cost directly, so the system learns a tradable position rather than an elegant one it cannot execute.

Compliance

The clearest AI use cases in finance and accounting are in compliance. Regulatory logic now sits inside the workflow. Agents watch transactions in real time, flag suspicious items, escalate the rest, all within guardrails, per IMF research on agentic AI in payments.

The same research names a coming problem: Know-Your-Agent, an extension of Know-Your-Customer that verifies the entity behind an autonomous agent. Old identity checks assume the presence of a human on the other end.

Agentic transformation of institutions

Some firms restructure around “agentic crews,” specialized agents handling modeling, validation, and compliance at once, all under human supervision. Moody’s runs multi-agent copilots that pre-screen credit applications and flag anomalies while keeping the audit trail intact.

Customer service with AI agents

Relationship manager churn runs 15% to 35% at many banks, driven by CRM admin, not client work. McKinsey found deal-scoring agents, recommending price and discounts in real time, cut prep time by about 25% and delivered around 10% margin gains in early use.

Allianz Partners, using a tool built with Taktile, cut claims processing from days to minutes and kept people in the loop. Agentic does not have to mean unsupervised.

Personalized guidance and AI financial management

Here artificial intelligence in financial management turns proactive. McKinsey describes a near-term case: an agent notices, unprompted, that a customer could clear a credit card balance with idle cash in another account, and it acts, though the last click stays with the customer for regulatory reasons.

The same capability cuts both ways. If the agent becomes the customer’s default interface, the bank beneath it becomes a backend. The automation that deepens personalization can sever the relationship.

Benefits of AI in finance

The benefits of AI in finance come down to four things.

Speed and scale: monitoring, stress testing, and fraud detection at a volume no human team matches.

Cost: real cuts in processing time without matching headcount.

Better risk-adjusted decisions: when the risk control lives inside the agent’s actions, backtests improve.

Freed judgment: over 90% of Moody’s interactions moved to high-value work. That is the real prize.

How can AI be used in finance?

The pattern is consistent. Start with a bounded process that already has rules and audit requirements; compliance, claims, credit pre-screening; not open-ended autonomous trading. Build in explainability and oversight from day one. Measure value before scaling.

The same Cambridge research found 55% of firms and 63% of regulators struggle to measure AI’s value, rising to 76% in the largest institutions. Scaling what you cannot measure is how a pilot becomes permanent.

The pitfalls

The upside is real. But so are the failures. Several are specific to agentic AI, not generic AI risk:

Goal misalignment. Agents optimizing their own reward can, together, produce what no one intended. As the density of learning agents rises, volatility rises with it, liquidity thins, and markets recover from shocks more slowly. No agent was built to destabilize anything. Individual reason, collective fragility.

Data quality. Agentic systems amplify the data they are given. Data quality, legacy infrastructure, and unclear ownership are the bottleneck behind slow scaling.

Accountability and interpretability. This is where the numbers bite. Making a system more interpretable typically costs 15% to 30% of its performance. In anti-money-laundering detection, the best systems hit 95% to 99% accuracy, yet fewer than 20% can explain themselves well enough to satisfy regulators. The gap between what works best and what can be audited is the hardest problem here.

Compliance and trust. Human-in-the-loop assumes a person reviews each decision. That breaks when an agent makes thousands of decisions a second. The SEC Market Access Rule and MiFID II were written for firms running fixed strategies under supervision, not for systems that rewrite their own strategy as they watch. You need a partner who understands compliance is not just an afterthought.

Summary: machine learning in financial services, what is new?

Machine learning in financial services is not new. Credit scoring and fraud detection have used it for over a decade. What changed is the move from a model that predicts to a system that acts.

An old model: flags a transaction as likely fraud.

An agentic system: flags it, chooses a response within its guardrails, executes, and adjusts its own future behavior, no human required to approve each step.

That is why at Vstorm we think that the governance above matters more than it would for ordinary machine learning banking. Different risks, greater rewards.

Frequently asked questions

How has AI been used in finance so far?

Mostly prediction and automation — credit scoring, fraud flagging, rule-based trading, report generation. Agentic AI is the shift to systems that plan and act across multistep workflows, not just produce output for a human to use.

What are the main AI use cases in finance now?

The mature ones are internal: compliance monitoring, credit pre-screening, claims processing, portfolio rebalancing. Customer-facing uses — personalized guidance, conversational banking — are earlier but growing fast.

What are the benefits of AI in finance?

Speed and scale in monitoring and risk detection, lower cost in processing-heavy work, better risk-adjusted outcomes when risk controls are built in, and skilled staff freed for judgment over data collection.

How can AI be used in finance without excessive risk?

Start with bounded, auditable processes. Build in explainability and oversight from the start. Measure value before scaling — most institutions admit they cannot yet.

Conclusion

Artificial intelligence in finance has cleared the pilot phase in the back office. Now it reaches the hard ground: autonomous decisions in trading, compliance, and the customer’s own account. The firms getting value are not the ones with the best models. They are the ones treating governance, explainability, and measurement as part of the build. The distance between the 42% experimenting and the 21% in production is where that gets tested.