TL;DR
Agentic AI in financial services means AI agents that plan and carry out multistep work across banking, insurance, capital markets, payments and wealth systems. Today, though, most of that work is decision support, where the agent recommends and a person decides. Agents with broad autonomy over money, credit or trades are still rare. Rules set the limit, from US model risk guidance and the EU AI Act to DORA and conduct rules. So the safest path starts with internal, read-only tasks and adds bounded actions behind human approval. Autonomy over financial decisions should grow only when the evidence from your own controls supports it.
Key Takeaways These agents plan, call approved tools and act inside regulated workflows, not just chat. Production use is real but cautious, with most firms running agents that recommend while staff decide and execute. Banking, insurance, capital markets, payments and wealth adopt at different speeds because their financial exposure differs. US model risk guidance now leaves agentic AI outside its scope, so firms must set their own controls for agents. The EU AI Act treats credit scoring and life and health insurance pricing as high-risk by use, with fraud detection carved out. A phased roadmap moves from internal support to bounded execution, with approval gates before any money-moving action. Watch on YouTube
Why AI Agents Are the Future of Finance
Kanerika explains where AI agents fit in financial work today and why controls, not raw autonomy, decide how far they go.
The Rulebook Changed Twice in One Year
Before 2026, a bank’s model risk team reviewing a new AI agent had a familiar yardstick. SR 11-7 had shaped US model validation since 2011. Across the Atlantic, the EU AI Act’s high-risk duties for credit scoring were due to apply in August 2026.
Then the clock moved twice. On April 17, 2026, the Federal Reserve, OCC and FDIC replaced SR 11-7 with SR 26-2, the revised guidance on model risk management . Its attachment also states that generative and agentic AI models are not within its scope. Then in July, Regulation (EU) 2026/1744 pushed the Annex III high-risk date to December 2, 2027.
So agents now sit between old model rules and new AI rules, and firms have filled that gap with caution. In the Bank of England and FCA survey of AI in UK financial services , 80% of respondents reported agents in production for decision support. By comparison, only 2% reported agents with broad autonomy. That before-and-after picture is the real starting point for this guide.
What Is Agentic AI in Financial Services?
Agentic AI in financial services refers to AI systems that take a goal, break it into steps and complete those steps through approved tools. A claims agent might read a loss notice, pull the policy, request missing documents and draft a settlement note. Because it acts on systems of record, it carries more risk than a model that only scores or writes.
The agent combines a language model with access to core platforms, such as loan origination, policy administration, trading or CRM systems. It also works only through APIs and document stores that the firm has approved. When a step needs judgment or touches money, a well-built agent stops and hands the case to a person.
This post looks across the whole sector. For finance-team work such as close, FP&A and payables, see our guide to AI agents in finance . For a cross-industry list, the companion piece on agentic AI use cases covers HR, supply chain and more.
How Agents Differ From Predictive Models and GenAI Assistants
Financial firms have used predictive models for decades, and most added generative assistants after 2023. Agents build on both, so the line between them can blur. The difference shows up in what each one is allowed to do once it has an answer.
Table 1: Predictive Models vs GenAI Assistants vs AI Agents in Financial Services
Dimension Predictive model GenAI assistant AI agent
Output A score or forecast Text, summaries, drafts A completed multistep task Acts on systems No, people or rules act on the score Rarely, through the user Yes, through approved tools and APIs Typical example Credit risk or fraud score Policy or research summary Dispute investigation from intake to draft resolution Main failure mode Wrong prediction Invented or incorrect answer Wrong action on a real account Main control Model validation Output review Permissions, approvals, audit trail and stop control
The last row is the one risk teams care about. Many products sold as agents are really assistants. That is why our explainer on AI copilots vs AI agents is worth a read before procurement.
Four Levels of Agent Autonomy
Autonomy, in practice, is not a yes-or-no setting. The Bank of England survey uses four levels, and they map well to how risk teams already think about delegation.
Decision support. The agent recommends, and a person decides and executes.Constrained execution. The agent acts within predefined limits and rules.Conditional autonomy. The agent acts alone but needs human approval for certain actions.Broad autonomy. The agent starts and completes actions with little real-time human involvement.In general, financial impact decides which level a workflow can reach. For example, a research summary can run at constrained execution early on. By contrast, a credit decline or a payment release usually stays at decision support until controls are proven.
Where Agentic AI in Financial Services Stands Today
Adoption is broad but shallow in autonomy. The UK survey covered 196 firms, and 80% of respondents reported agents already deployed, with a further 14% planning deployment within a year. Enterprise and coding agents were the most common types, while client-facing agents were less widespread.
What Has Reached Production
Internal work leads so far. Agents that search policies, summarize documents, prepare research packs and triage service tickets are live at many firms. Operations and IT is the busiest area, and 53% of respondents reported decision-support agents there.
Customer contact, by comparison, is growing more slowly. Only 16% of respondents reported agents handling automated customer communications, though that rises to 34% once planned deployments are counted. Even here, most firms keep a person on review for anything that commits the institution.
What Still Sits Behind Approval Gates
Money-moving decisions, in contrast, remain the exception. Even with 12-month plans included, no more than 9% of respondents expected agents to approve or decline customer outcomes. The same cap applied to executing trades, initiating or blocking payments and changing limits or controls.
Autonomous credit approval, claims settlement and portfolio changes therefore remain pilots or tightly gated workflows. We think that caution is right, at least for now. A wrong summary costs time, but a wrong payment costs money, triggers complaints and can draw a supervisor’s attention.
How to Test a Production Claim
Vendors and internal teams both like to say an agent is in production. Before accepting that, ask for evidence from your own environment rather than a demo. Our guide to AI agent evaluation also covers the testing side.
Real users on live data, with a named business owner. Live integrations to systems of record, not copies or sandboxes. A list of actions the agent is authorized to take, and what it may never do. Logs that show tool calls, approvals, overrides and errors for each run. Measured outcomes against a baseline, including the time people spend reviewing. If any of these are missing, the agent is still a pilot, whatever the slide says.
AI Assessment
How Ready Is Your Organization for AI Agents?
Score your data foundations, generative AI and AI agent readiness in a few minutes and get recommendations for your next step.
Start Your AI Assessment → How Agents Work Across Each Financial Subsector
Each part of the sector has its own workflows, data and regulators. So the same agent pattern lands differently in a bank branch, a claims office or a trading floor. The sections below follow the money, from deposits and loans to policies, trades, payments and portfolios.
Retail and Commercial Banking
Banks use agents for tasks such as account servicing, dispute investigation, KYC refresh and loan file preparation. In commercial lending, an agent can assemble a credit memo from financial statements, check covenants and flag gaps for the relationship manager. The credit officer still signs the decision.
Our dedicated post on agentic AI in banking covers these workflows in detail, so this section stays brief. One theme carries across the sector, though. Data quality decides how far banking agents can go, which is why data governance in banking usually comes before the agent itself. Clean customer and account records also make an agent’s decisions easier to explain later.
Insurance
So far, insurers have moved fastest on intake and triage. Agents read first notice of loss reports, sort claims by complexity, gather evidence and route fraud signals to investigators. On the underwriting side, they also pre-fill submissions and check documents against appetite rules.
For an insurer covering health, travel and accident claims, Kanerika added an AI fraud check before payout. Each claim was screened before payment instructions reached the bank. As a result, claim processing time fell 20% and cost savings reached 36%.
Case Study
36% Cost Savings with AI/ML-Powered RPA in Insurance
An insurer covering health, travel and accident claims cut claim processing time by 20% with Kanerika-built fraud checks that run before payment instructions reach the bank.
Read the Case Study → Pricing and coverage decisions, however, are different. In the EU, the AI Act treats life and health insurance pricing as a high-risk use. So agents there usually prepare the file rather than decide. Our posts on AI agents for insurance claims processing and generative AI in insurance also go deeper on each workflow.
Capital Markets
Capital markets firms use agents for research preparation, trade surveillance triage, reconciliation and settlement exceptions. For instance, an analyst can ask an agent to pull filings, compare them with internal notes and draft a briefing. A surveillance agent can gather the context behind an alert so the reviewer starts with a full picture.
Order execution is a separate capability with far higher stakes. That is why FINRA’s 2026 oversight report section on generative AI lists the risks plainly. Agents may act without human validation, go beyond their intended authority, or leave multistep reasoning that is hard to audit.
Post-trade work, on the other hand, is a safer early target. Agents can match breaks between internal books and custodian records, propose fixes and queue them for an operations analyst. Because the agent never touches the order book, a mistake costs time rather than market exposure.
For one investment bank, Kanerika deployed KlarityIQ for documents and Karl for database queries behind role-based access. Staff got answers through one chat interface while controls stayed inside the retrieval layer. As a result, information retrieval became 43% faster, with 100% role-based compliance.
Case Study
43% Faster Document Retrieval for an Investment Bank
Kanerika deployed KlarityIQ and Karl behind role-based access so analysts could query documents and databases in one chat, with 35% higher workforce efficiency.
Read the Case Study → Payments and Fraud Operations
Payments teams and fintechs use agents to investigate fraud alerts, chase failed payments and work chargeback cases. The agent collects device data, transaction history and merchant records, then writes a case summary. An analyst then decides whether to block, release or escalate.
Fraud work also has a notable regulatory position. The EU AI Act’s credit-scoring category expressly excludes AI systems used to detect financial fraud, which gives fraud teams more room to automate triage. Our guide to AI in fraud detection also covers the models that sit underneath these agents.
Blocking a legitimate customer payment still harms that customer, though. So most firms let agents queue and recommend holds, while a person confirms anything that stops money moving.
Wealth and Asset Management
Advisers use agents to prepare client meetings, summarize portfolio changes, draft follow-up notes and monitor accounts against mandates. Similarly, asset managers apply the same pattern to research and reporting. These tasks free adviser time without handing over investment judgment.
Even so, suitability and best-interest duties keep recommendations with licensed people. In the UK survey, 23% of respondents reported decision-support agents in asset management work, and none reported broad autonomy there. Our post on AI in wealth management covers the adviser-side tools in detail.
Table 2: Where Agents Work by Financial Subsector
Subsector Where agents work now What stays with people Main constraint
Banking Servicing, disputes, KYC refresh, credit memo prep Credit approval, account closure Fair lending, consumer protection, model risk Insurance Claims intake and triage, underwriting prep, fraud referral Coverage, pricing, claim settlement Conduct rules, AI Act for life and health pricing Capital markets Research packs, surveillance triage, reconciliation Order execution, trading limits Market conduct, supervision, auditability Payments and fintech Fraud alert investigation, failed payment repair, chargebacks Blocking or releasing payments Payment rules, customer harm, operational resilience Wealth and asset management Meeting prep, reporting, mandate monitoring Investment advice, portfolio trades Suitability and best-interest duties
Across every row, agents do the gathering and preparing, while people own the decision that moves money or changes a customer’s position.
The Rules That Decide How Much an Agent May Do
Regulation is the real ceiling on agent autonomy in this sector. No single rulebook covers AI agents, so firms must read several regimes together. In practice, the four below shape most design choices.
US Model Risk After SR 26-2
Until 2026, SR 11-7 set the standard for model development, validation and ongoing monitoring at US banks for 15 years. SR 26-2 now supersedes it and takes a more risk-based approach, aimed mainly at banking organizations with over $30 billion in assets. However, its guidance attachment states that generative and agentic AI models are outside its scope.
That exclusion is not a free pass, though. Examiners will still expect a clear answer on how agents are governed. The same footnote says a bank’s own risk management and governance practices should still set controls for tools the guidance does not cover. In practice, models an agent calls still need model risk review, while the agent itself needs a separate control framework.
EU AI Act High-Risk Uses
The AI Act classifies systems by intended use, not by industry. Annex III of Regulation (EU) 2024/1689 names two financial uses as high-risk in particular. The first is creditworthiness assessment and credit scoring of natural persons, with fraud detection excluded. The second is risk assessment and pricing in life and health insurance, such as setting health cover premiums.
After the 2026 amendment, those high-risk duties apply from December 2, 2027. So an agent that summarizes documents is not high-risk on that basis alone. But if the same agent decides or materially shapes a consumer credit outcome, the high-risk duties attach to that use.
DORA and Third-Party Dependence
The Digital Operational Resilience Act has applied since January 17, 2025, according to the EIOPA summary of DORA . It covers banks and insurers as well as investment firms and other financial entities. It requires them to withstand, respond to and recover from ICT disruption, including failures at critical technology providers.
That matters because agents lean heavily on outside models, clouds and APIs. The UK survey found third-party systems made up 53% of high-materiality AI systems, up from 35% in 2024. Under DORA, that dependence needs resilience testing, incident processes and an exit plan for each critical provider. Concentration adds to the risk, because the top three cloud and compute providers accounted for 79% of reported relationships in the same survey.
Conduct Rules in Lending, Securities and Insurance
Sector conduct rules apply regardless of whether a person or an agent does the work. In US lending, Regulation B requires creditors to give specific reasons for adverse action. The CFPB has also said complex algorithms do not remove that duty . Broker-dealers face FINRA supervision expectations, while insurers face state conduct rules.
These duties therefore push agents toward explainable steps and clear handoffs. Our post on explainable AI covers techniques that help, and our overview of AI compliance maps the wider obligations.
Watch on YouTube
Klara- Compliance Agent I the AI Agent to Enforce Your Governance Playbook I Kanerika
See how Kanerika’s Klara agent checks documents against a compliance playbook, flags risks and keeps a person in charge of the final call.
Table 3: What Each Rule Means for Agent Design
Rule or regime Main concern Agent design implication
SR 26-2 (US model risk) Validity of models used in decisions Validate the models an agent calls; govern the agent under your own framework EU AI Act Annex III Credit scoring and life and health pricing Classify each agent by use; add documentation, oversight and logging for high-risk uses DORA ICT resilience and third-party dependence Resilience tests, incident runbooks, vendor oversight and exit plans Regulation B and conduct rules Customer harm and unexplained decisions Specific reasons for outcomes, approval gates and accountable staff FINRA supervision Scope, autonomy and auditability of agents Scoped permissions, action tracking and human-in-the-loop review
In short, the rules rarely ban an agent outright. Instead, they decide which steps need a human signature and what evidence the firm must keep.
What a Controlled Agent Deployment Looks Like
A controlled deployment turns those rules into working limits. The pattern is the same across subsectors, even when the systems differ. Four controls carry most of the weight, and our guide to AI agent architecture shows where each one sits in the stack.
Least-Privilege Access to Core Systems
Give each agent its own identity, then grant only the permissions its task needs. Separate read-only tools from tools that change records, and grant write access one action at a time. For example, an agent that drafts dispute letters has no reason to reach the payments ledger.
Sensitive data also needs the same care. When a task does not need account numbers or personal details, mask them before they reach a model. A tool such as Susan, Kanerika’s PII redaction agent , can do this step automatically. Our post on agentic AI security also covers prompt injection and scope abuse.
Human Approval at Money Decision Points
Place approval steps before any action that changes a customer’s position or moves funds. Typical gates sit before credit decisions, claim payments, trades, limit changes and customer commitments. The agent prepares the case, then a named person approves or rejects it.
Good approval design keeps reviewers fast without making them rubber stamps. So show the evidence the agent used, the action it proposes and its confidence. Our guide to human-in-the-loop AI agents explains how to set those thresholds in practice.
A Record That Rebuilds Every Run
Sooner or later, supervisors and auditors will ask what the agent saw, what it did and who approved it. So log every request, data source, tool call, model version, approval, override and final result against one run ID. Without that record, an agent’s decision cannot be explained to a customer or an examiner, even if it was correct.
Monitoring should then watch those logs in production. For instance, alerts on error rates, override spikes and unusual tool use catch drift early. Our post on AI agent observability covers the tooling in detail.
Stop, Roll Back and Escalate
Every agent needs a working stop control that operations staff can use whenever something goes wrong. When a connected system fails mid-task, the agent should halt, reverse what it can and route the case to a person. Assign ownership of these failures before launch, not after the first incident.
Test the stop control the way you test disaster recovery. Run a drill, confirm the agent halts within the agreed time and check that queued work lands with a person. Then keep the result as evidence for the next risk review.
A Phased Adoption Roadmap for Financial Institutions
Firms that scale agents safely tend to add autonomy in stages rather than all at once. Each phase has to earn the next one with evidence. Each phase widens either the scope of the agent’s work or its authority to act.
Phase 1: Internal Support With No Write Access
Start with tasks such as document classification, policy search, research preparation and case summaries. Staff check every output, while the agent cannot change financial records. This phase builds the data access, logging and evaluation habits the later phases depend on.
Phase 2: Bounded Workflow Execution
Next, let agents complete defined steps in claims intake, reconciliation, onboarding and service tickets. Use restricted tools, action limits and exception queues so that errors stay contained. A monthly statement run, for example, fits this phase well.
Kanerika automated statement generation for a bank that produces about 70,000 consolidated statements in a single day each month. Data from three source systems now flows into one template without manual handling. Processing time dropped 80%, and the solution was proven at twice the current volume.
Phase 3: Decision Support for Customer and Specialist Work
Once accuracy and monitoring meet agreed standards, extend agents into customer servicing, underwriting assistance and adviser support. The agent recommends, but specialists decide. Track how often people accept, edit or reject its proposals.
Kanerika built a cash-flow forecasting model for a bank’s treasury team inside its daily funding portal. The model flags surplus cash, while treasury staff decide what to release. Workforce productivity rose 20% and processing delays fell 35%.
Phase 4: Narrow Execution After Risk Approval
Finally, consider letting agents execute specific financial actions within hard limits. This step needs independent review, transaction caps, real-time monitoring and a tested fallback. Most firms will keep some workflows at Phase 3 for good, and that is a sound outcome.
Table 4: Phased Agent Roadmap for Financial Institutions
Phase Autonomy level Example workflows Gate to the next phase
1. Internal support Decision support, read-only Research packs, document search, case summaries Accuracy targets met; data access approved 2. Bounded execution Constrained execution Claims intake, reconciliation, statement runs Exception rates within limits; logs complete 3. Decision support Decision support, people approve key actions Servicing, underwriting prep, adviser support Risk review passed; override rates stable 4. Narrow execution Conditional autonomy within hard limits Defined, low-value financial actions Independent validation, caps and live monitoring
The same phasing also works outside finance. Our post on agentic AI in HR applies it to employee data, where privacy stakes are also high.
Checklist
Enterprise Agentic AI Checklist
A practical checklist for taking agentic AI from pilot to production, covering readiness, governance, controls and rollout.
Get the Checklist → Measuring Whether Agents Pay Off
Time saved is the easy number, and it often misleads. So a fair scorecard counts review effort, exceptions and incidents alongside speed. In general, three groups of measures cover most needs.
Efficiency. Processing time, cases completed, manual effort removed and cost per completed task.Quality and risk. Wrong recommendations, reversed decisions, reconciliation errors and blocked unauthorized actions.Oversight. Override rates, escalation volume, review time and audit log completeness.Then tie each measure to a business outcome the subsector cares about. For a bank, that may be servicing cost per account, and for an insurer, claims turnaround. Capital markets teams track exception backlogs, while wealth firms watch adviser time with clients.
Mistakes That Stall Agentic AI Programs
Agent programs in financial firms tend to stall for the same few reasons, even when the technology works. Each one is avoidable once teams plan for it early.
Calling an assistant an agent. Chat tools that only draft text get counted as agents, so maturity and ROI claims inflate.Starting with high-stakes decisions. Autonomous lending, trading or payouts make poor first projects because controls are unproven.Validating the model but not the actions. Reviewing outputs alone misses tool misuse, workflow errors and unintended actions, especially across systems.Ignoring legacy data. Inconsistent records and brittle interfaces break agents even though they test well on clean samples.Trusting vendor claims over internal evidence. A demo or benchmark is not proof of safe behavior on your systems until you test it there.The legacy data problem deserves extra attention because it sits under every other mistake. Our guide to data migration in banking covers how to fix the foundations before agents depend on them.
How Kanerika Helps Financial Institutions Deploy Agents
Kanerika works with banks, insurers and investment firms to move agents from pilot to controlled production. Our agentic AI services build agents one workflow at a time. We pick a workflow with a measurable baseline, then design the agent’s tools and approval points. Next, we connect it with scoped access and test it on historical cases. After launch, monitoring, audit logs and a review loop keep it accurate.
Governance runs through every stage through our AI governance services , which cover AI inventories, access controls and audit trails. We also bring our own agents where they fit. KlarityIQ handles document intelligence, Karl answers data questions, and Klara checks contracts against a compliance playbook. Each agent ships with its permissions, approval points and logs documented for your risk team.
Delivery also runs under ISO 27001 and SOC 2 Type II controls. Our teams watch for three pitfalls in finance work. Agents get write access before read-only behavior is proven, logs miss tool outputs, and nobody owns the approval queue once volumes grow. We also push for a named owner for every agent before it touches production data. See our banking and financial services and insurance pages for sector work.
Every engagement starts small. We pick one workflow, prove it with your data and controls, and widen scope only when the evidence holds.
How Far Agentic AI in Financial Services Can Go Today
So where does the sector stand, and how much can agents do? Agents are already working across banking, insurance, capital markets, payments and wealth. They mostly gather, prepare and recommend, while people approve anything that moves money or changes a customer’s position.
That limit comes from the rules and from the cost of a wrong action on a real account. SR 26-2 leaves agents to each firm’s own governance. The AI Act also marks credit scoring and life and health pricing as high-risk, while DORA governs the vendors underneath. Firms that start with internal tasks, log every run and add approval gates can then widen autonomy step by step. For now, the right answer for most financial decisions is an agent that does the work and a person who signs it.
Frequently Asked Questions
What is agentic AI in financial services?
Agentic AI in financial services means AI agents that take a goal, plan the steps and complete them through approved tools in banking, insurance, capital markets, payments and wealth systems. Unlike a credit score or a chat assistant, an agent can act on systems of record, so it needs permissions, approval steps and an audit trail.
How is agentic AI used across banking, insurance and capital markets?
Banks use agents for servicing, disputes, KYC refresh and credit memo preparation. Insurers use them for claims intake, triage and underwriting preparation. Capital markets firms use them for research packs, surveillance triage and post-trade reconciliation, while people keep credit, coverage and trading decisions.
Are financial institutions using autonomous AI agents in production?
Yes, but mostly with limited autonomy. In the Bank of England and FCA 2026 survey, 80% of respondents reported agents in production for decision support, where agents recommend and people decide. Only 2% reported agents with broad autonomy, and few firms plan to let agents approve outcomes or move money.
Does SR 11-7 still apply to AI agents at US banks?
No. SR 26-2, issued in April 2026 by the Federal Reserve, OCC and FDIC, superseded SR 11-7. Its attachment states that generative and agentic AI models are outside its scope, so banks must govern agents through their own risk management practices while still validating the traditional models agents call.
Does the EU AI Act classify all financial AI agents as high-risk?
No. The AI Act classifies by use. Creditworthiness assessment and credit scoring of natural persons, and risk assessment and pricing in life and health insurance, are high-risk under Annex III, with fraud detection excluded. After the 2026 amendment, those duties apply from December 2, 2027.
Can AI agents make lending, insurance or investment decisions without human approval?
In most firms they do not, and in many cases they should not. Rules such as Regulation B, suitability duties and insurance conduct rules require explainable, accountable decisions. Most institutions let agents prepare the case and recommend, then have a named person approve anything that changes a customer’s position.
How does DORA affect AI agents in financial services?
DORA has applied since January 2025 to banks, insurers, investment firms and other EU financial entities. Agents depend on outside models, clouds and APIs, so DORA’s rules on ICT risk, incident reporting, resilience testing and third-party oversight apply to the services behind each agent.
What is the safest starting point for agentic AI in a financial institution?
Start with internal, read-only work such as policy search, document classification, research preparation and case summaries. Staff check every output and the agent cannot change financial records. That phase builds the logging, data access and evaluation habits needed before agents take bounded actions.
How should financial institutions measure the ROI of agentic AI?
Measure cost per completed task after counting review time, exceptions and incidents, not just time saved. Add quality measures such as reversed decisions and errors, and oversight measures such as override rates and audit log completeness. Then tie each measure to a subsector outcome like claims turnaround.
