Agentic AI in Financial Services and Insurance

Agentic AI in Financial Services and Insurance

A Q&A with Dmitry Tikhomirov, Global Head of Microsoft Business Group, EPAM Systems, Inc.

Why is agentic AI especially relevant to financial services and insurance now?

Financial services and insurance are full of rules, documents, exceptions and handoffs, and that makes them a good fit for agentic AI. Most banks and insurers have digitised the customer-facing parts of the business. But if you look inside the operation, a lot of important work still depends on people moving between emails, PDFs, legacy systems, spreadsheets and compliance checks. That is costly, slow and increasingly hard to defend.

Agentic AI is well suited to these industries. Unlike a chatbot, it can pursue a goal across several steps: gather information, interpret documents, call tools, hand work to specialist agents, flag anomalies and escalate when needed. In regulated industries, that is useful because it takes repetitive work away from people who should be spending more time applying judgement.

Which processes are best suited to agentic AI today?

The best place to start is with work that is document-heavy, repetitive and spread across multiple systems.

In banking, that includes onboarding, KYC (Know Your Customer), lending intake, trade-finance operations and financial-crime case preparation. In capital markets, post-trade exceptions, reconciliations and surveillance triage stand out. In insurance, underwriting intake, claims triage, policy servicing and fraud referral are obvious targets.

KYC is a good example. An agentic workflow can collect documents, extract and validate fields, reconcile entities, screen sanctions and adverse media, assess risk indicators and prepare a case for review. In underwriting, agents can ingest broker submissions, extract exposure data, and triage them to ensure they are within the firm’s risk tolerance before prioritising them for the underwriter. In claims, they can classify first notice of loss, retrieve policy details and route suspicious cases early.

A lot of the value comes from removing the manual coordination that sits between systems today. In other words, AI agents excel where humans are acting as costly middleware between systems.

What should remain firmly human-led?

Anything that creates a material customer, legal or financial outcome should remain accountable to a person.

Complex credit decisions, claims denials, suspicious activity judgements, product suitability, complaint handling and major underwriting calls should remain human-accountable. In these cases, context, proportionality and conduct matter as much as efficiency.

The right model is not machine-led decision-making. It is machine-prepared, human-accountable execution.

In commercial insurance, relationships are essential to large and complex risk placement, where trust is at the heart of doing business. Agents have a limited role to play in protecting these crucial human touchpoints. However, AI can still support these activities by providing research, preparation and insights.

How should leaders think about risk and control as systems become more autonomous?

By treating autonomy as something to calibrate rather than celebrate. Agents need specific mandates, clear permissions and visible boundaries. What can the agent read? What can it write? What can it trigger? When does it need to stop and ask for approval? In regulated industries, those are not technical details; they are the architecture.

This is where Microsoft’s stack is useful in practice. Microsoft Entra ID governs identity and access. Microsoft Agent 365 provides emerging control-plane capabilities for discovering, registering and governing agents across the enterprise. Microsoft Purview handles protection, retention and compliance. Microsoft Defender strengthens security monitoring. And telemetry across Azure services provides the audit trail regulators and internal control functions will eventually ask for. The aim is controlled delegation: faster processes with stronger traceability.

Where is the highest-value opportunity in banking and capital markets?

Three areas stand out. First, client lifecycle and financial crime: onboarding, KYC refresh, alert handling and investigations. These are labor-intensive tasks and full of duplication. Agents can gather evidence, enrich alerts and prepare cases for analysts.

Second, lending and trade operations: much of the delay lies not in decision-making but in assembling the file. Agents can extract, compare and organize information before a human applies judgement.

Third, post-trade operations: settlement breaks, reconciliation mismatches and operational exceptions are ideal agent territory - structured enough to automate, messy enough to need orchestration.

And in insurance?

Insurance may be the most natural fit for agentic AI, because so much of the work involves reading documents, making judgement calls and dealing with exceptions.

In commercial and specialty insurance, underwriting teams often receive large and varied submission packs. Agents can help by reading those packs, pulling out key exposure data, helping with triage and routing, summarising long documents, researching clients, assets and perils and comparing policy wording. They can also help underwriters see how a risk fits within the wider portfolio before the policy is bound.

Claims are another major opportunity. Agents can help from the first notification of loss to coverage verification, triage and routing to the right handlers, subrogation analysis, and fraud or anomaly detection, bringing material improvements to all these processes.

In insurance, value is not based solely on productivity or speed. The material value comes from giving experienced people more time to consider the new agentic-generated insights. That drives better decision-making, which brings the promise of improving financial performance.

What role does Microsoft play when building these systems?

Microsoft matters because it offers more than models. It offers an operating environment where agents can be built, deployed, managed and governed safely.

Microsoft’s appeal lies in its ability to assemble not just AI tools but a usable, secure enterprise stack.

Work IQ, Fabric IQ and Foundry IQ form the intelligence layer, helping agents understand how work is done, where trusted data resides and how enterprise knowledge should be accessed.

Foundry Agent Service is the runtime, where agents are deployed, coordinated and managed in production.

Beneath that sits the developer layer: Microsoft Agent Framework for more disciplined engineering, and Microsoft Copilot Studio for lower-code, business-facing agent design.

And finally, governance. Microsoft Agent 365, alongside Microsoft Entra ID, Microsoft Purview and Microsoft Defender, provides the controls for identity, policy, compliance and security that regulated firms will insist upon.

Read the full Q&A here. 

Discover how EPAM and Microsoft help global enterprises turn complex challenges into scalable, responsible AI solutions that accelerate transformation and unlock real business value: epam.com/about/who-we-are/partners/microsoft