日本語
한국어
Tiếng Việt
简体中文
By Zachary Scott, Managing Director at Publicis Sapient
Banks are moving beyond artificial intelligence that simply recommends actions. The next stage involves agentic AI systems that can complete tasks, make decisions within defined limits and coordinate work across multiple banking functions.
However, an AI agent can only perform as effectively as the technology environment surrounding it. If an account-opening agent needs identity information from one platform, screening results from another, policy guidance from a separate repository and approval from a manual queue, the bank must first connect those resources and define how the agent can use them.
Modernization is the foundation for scaling agentic AI in banking. That does not mean every institution must begin with a lengthy, enterprise-wide transformation. Banks can make practical progress by focusing on the dependencies that prevent agents from completing real work: legacy systems, disconnected enterprise information and undocumented decision-making.
Why Legacy Technology Can Limit Agentic AI
Legacy technology often becomes the first major obstacle when an AI agent moves from a pilot project into a live banking workflow.
Consider an agent investigating a disputed card transaction. It may need to retrieve transaction records, review previous disputes, apply the appropriate rules, initiate an action and document every step. If those functions depend on outdated systems, limited application programming interfaces, batch processing or manual handoffs, the agent may be unable to complete the process consistently.
Rather than replacing every legacy platform immediately, banks can identify the specific capabilities that repeatedly block automation. Practical modernization efforts may include:
- Exposing important legacy functions through secure, governed APIs.
- Separating business rules from outdated application code.
- Replacing manual handoffs with reusable digital services.
- Improving access to data needed by high-value banking workflows.
AI can also support the modernization process. It can analyze older code, identify system dependencies, extract embedded business rules for review and generate test cases. This allows banks to prioritize modernization based on the capabilities that agents actually require in production.
Give AI Agents a Reliable View of the Enterprise
Connecting systems is only one part of the challenge. Once an agent can access information across the organization, it must understand how that information relates to customers, products, policies, permissions and business objectives.
An enterprise context graph can provide that structure. Banks have already invested in common data models, master records and authoritative sources. Agentic AI requires those resources to be connected to the policies and relationships that determine how information should be interpreted and used.
A governed context graph can show an agent:
- Which data source is authoritative.
- How information about a customer, product or account is connected.
- Which policies apply to a particular situation.
- What permissions govern access and action.
- Which business objectives should influence the next step.
For example, an agent assisting a homeowner who has missed multiple mortgage payments may need to review payment history, previous communications and hardship policies. The available options could also depend on the loan type, investor requirements, geographic location and the stage of delinquency.
A context graph helps connect those facts to the relevant policies and decision rules. That gives the agent a clearer understanding of which information matters, which options are permitted and when the situation should be escalated.
Convert Institutional Judgment Into Guardrails
Banking operations frequently depend on judgment that experienced employees apply without documenting every individual step. This creates a challenge for agentic AI. To scale safely, banks must make more of that judgment explicit.
Agents need clearly defined objectives, measurable outcomes, decision rights, escalation thresholds and controls that cannot be bypassed. These elements turn informal institutional knowledge into practical guardrails.
In customer onboarding, for instance, a bank may want to reduce account-opening time while continuing to identify fraud, sanctions risks and other threats. An agent could gather documents, reconcile information and approve clearly defined low-risk applications. A sanctions match, ownership discrepancy or elevated risk score could trigger deterministic checks or a human review.
Deterministic systems rely on fixed, pre-programmed logic. They are particularly useful when a decision must follow a consistent rule and the bank needs a clear audit trail. Agents, by contrast, are better suited to situations that require flexibility, interpretation or coordination across multiple sources.
Bank leaders should not expect agents to determine the bank’s priorities independently. They should establish a focused set of objectives and define how the system must handle competing goals. Clear guardrails make decisions more consistent and give risk, compliance and audit teams concrete standards for testing and monitoring the workflow.
The First Agentic AI Deployment Should Not Be the Final Design
Modernizing technology, organizing enterprise context and codifying decision rules can help an agent reach production. The next question is whether the bank learns from that deployment and improves the workflow over time.
The first version of an agentic process should be viewed as an interim design rather than a permanent solution. Production activity can reveal which decisions truly benefit from an agent and which processes are predictable enough to be handled through simpler automation.
An onboarding agent may initially manage a wide range of applications. After reviewing performance, the bank may discover that many cases follow the same sequence of checks. Those cases can be transferred to deterministic automation, allowing the agent to focus on ambiguity and employees to handle exceptions requiring professional judgment.
As a result, a mature agentic workflow may use less AI than its original version. The objective is not to maximize the amount of AI in a process. The objective is to use agents where their flexibility creates measurable value, then simplify the workflow as the bank learns which activities can be standardized.
Build Reusable Foundations for Future AI Use Cases
This approach makes cost, control and performance part of the original design rather than issues addressed after deployment. It also offers banks a practical way to measure whether they are genuinely scaling agentic AI.
Each implementation should leave behind reusable APIs, organized context, documented decision rules and performance data. These assets should make the next AI deployment faster and more reliable. At the same time, earlier workflows should become more efficient, less expensive and easier to govern.
Successful scaling is therefore not simply a matter of launching more AI agents. It requires building an operating environment in which agents can access the right systems, understand the relevant context, follow explicit rules and improve through real-world learning.
Banks that take this foundation-first approach can move from isolated AI experiments toward a sustainable model for agentic automation across customer service, onboarding, fraud operations, lending and other critical functions.
Source: TheFinancialBrand.com