Successfully Scaling Agentic AI Starts with the Bank Underneath It
By Zachary Scott, managing director at Publicis Sapient
Simple Subscribe
Subscribe Now!
Banks are moving from AI that recommends work to agents that execute it. That shift exposes a practical constraint: An agent can only operate as well as its environment.
If a customer-onboarding agent needs identity data from one system, screening results from another, policy rules from a repository, and approvals from a manual queue, the bank still has to make those systems accessible, establish which information matters, and define what the agent can decide.
Key insight: Modernization is therefore a prerequisite for scaling agentic AI in banking. But banks do not need another multiyear transformation program before they can make progress. They can modernize the dependencies agents need to do real work, starting with three areas: legacy technology, enterprise context and codified decisioning.
Once they have those foundations in place, banks can use what they learn in production to simplify and improve the workflow over time.
Need to Know:
- Legacy systems can become a stumbling block when a bank attempts to move agentic AI into production, but a consistent approach can focus attention on key blockages. AI itself can facilitate improvement.
- Using “enterprise context graphs” can help AI agents connect data to the appropriate policies and decision rules governing a given process.
- Many banking operations rely on judgment that humans apply implicitly. Scaling AI requires that banks make this explicit to agents.
- Sometimes the best approach is a blend of agentic AI with other approaches that can be implemented more simply.
- Ideally, each implementation of agentic AI should be a steppingstone that later implementations can benefit from.
Begin Where Legacy Blocks Execution
Legacy technology becomes an immediate constraint when AI enters a production workflow.
Example: Consider an agent investigating a disputed card transaction. Completing the work could require retrieving transaction history, checking prior disputes, applying dispute rules, initiating an action and then recording each step. When those capabilities depend on aging systems with limited application programming interfaces, batch processing or even manual handoffs, the agent cannot reliably execute the workflow at scale.
Banks can focus modernization on the components that repeatedly block these workflows. Steps to take include;
- Expose a legacy capability through a governed API.
- Separate business rules from old application code.
- Replace manual handoffs with reusable services.
AI can accelerate that work by analyzing legacy code, tracing dependencies, extracting embedded rules for human review, and generating test cases. Modernization can then follow the capabilities AI actually needs in production.
Read more: The Emerging Agentic Playbook for Community Banks and Credit Unions
Build an Enterprise Context Graph
Modernizing legacy systems gives agents access, but they still need to understand how information fits together across those systems. Once an agent can reach across the bank, it needs a reliable way to interpret the relationships among the data it retrieves and the rules governing its use.
A key tool: This is where an enterprise context graph becomes important.
Banks have spent years building common data models, master records and authoritative sources. Agentic AI adds the need to connect that data to the policies, customers, products, permissions and business objectives that give it meaning. The result is a governed map of what information matters, where it comes from, how it relates and which rules apply.
Example: Consider an agent supporting a homeowner who has missed several mortgage payments. Responding could require payment history, prior communications and hardship policies. Available options may depend on the loan type, investor requirements, geography and stage of delinquency.
An enterprise context graph connects those facts to the policies and decision rules governing the situation, allowing the agent to identify authoritative information and determine which options can be presented.
Read more: Banks Learn the Art of ‘Tokenomics’ to Make More Efficient Use of AI Tools
Turn Judgment Into Guardrails
Giving an agent that context creates another requirement: The bank must be explicit about what the agent is trying to achieve and the boundaries within which it can act.
Many banking processes rely on judgment that experienced employees apply implicitly, without documenting every choice. Scaling agents requires more of that logic to become explicit: workflow objectives, the metrics that matter, decisions an agent may make, escalation thresholds and controls that cannot be overridden.
Example: In onboarding, a bank may want to reduce account-opening time while maintaining risk detection.
An agent could collect documents, reconcile information, and clear predefined low-risk cases, while a sanctions match, ownership discrepancy, or elevated risk score triggers deterministic checks or human review. (Deterministic architecture performs tasks based on rigid, pre-programmed logic defined in advance.) These rules create a decisioning framework that constrains the agent and gives risk teams something concrete to test, monitor and audit.
Agents should not have to improvise the bank’s priorities. Leaders need a small set of clear objectives and explicit rules for handling tradeoffs. That turns institutional judgment into guardrails that can be applied consistently across the workflow.
Read more: AI Is Only as Good as the Data Behind It. Five Questions to Tell If Yours Is Ready
A Practical Test for Scale
Modernizing the technology, building enterprise context, and codifying decision rules gets an agent into production. What happens next may be the better test of whether the bank is actually learning how to use agentic AI.
Key insight: The first version of an agentic workflow should be treated as an interim design.
As it runs, the bank can see which decisions genuinely benefit from an agent and which become predictable enough to handle more simply.
An onboarding agent, for example, might initially work across a broad range of applications. Over time, the bank may find that many cases follow the same sequence of checks. Those can move to deterministic automation, reserving the agent for ambiguity and employees for the exceptions that require judgment.
In that sense, the mature agentic workflow may use less AI than the first version. The goal is to apply agents where their flexibility creates value, then simplify the process as the bank learns which paths can be standardized.
Key insight: That makes cost, control and performance part of the design rather than questions addressed after deployment.
It also creates a practical measure of scale. Each deployment should leave behind reusable interfaces, context, decision rules and evidence about how the workflow performs. When those capabilities make the next deployment easier, while allowing existing workflows to become simpler and cheaper, the bank is building an operating environment that can support agentic AI beyond individual use cases.
Read next: Guardrails Can Help Banks Reap AI’s Promise While Maintaining Trust
