Podcast Audio Transcript
Mimi: Hello listeners, this is Mimi; thank you for tuning in to yet another exciting podcast from Infosys BPM. Today, we’re tackling a question many banks are wrestling with: if your operations are still largely manual, do you start with small automation—or do you leap straight into Agentic AI and autonomous processing?
Joining me today is Megha Kochhar, Industry Principal, leading the digital transformation for Financial Services at Infosys BPM. She will help us understand how and where banks should start this journey. Welcome to the show, Megha! How are you today?
Megha: I am doing good Mimi. Thanks for having me on the podcast.
Mimi: Of Course! Megha, can we start by understanding what low digital maturity looks like today?
Megha: Great question Mimi.
Most banks aren’t sitting at zero. Low maturity usually looks like automation in pockets—and a lot of human glue holding it together. You’ll see RPA everywhere, but not necessarily end-to-end impact.
Here are some real patterns:
- In KYC onboarding, bots might download documents or create cases, but analysts still manually reconcile inconsistencies—names, addresses, document mismatch—across systems.
- In mortgage ops, there’s a workflow and some automation, but the process is exception-heavy: missing documents, policy nuances, and manual follow-ups drive cycle time.
- In AML alert handling, alerts are generated digitally, but investigations are still copy-paste research across tools.
So the issue isn’t “no tech.” It’s that tech isn’t connected into a coherent operating model.
Mimi: That’s very interesting. Many people argue you need clean data and standardized processes before you even think about agents. How can Agentic-first work in that reality?
Megha: “Agentic-first” doesn’t mean “Agentic decides everything.” It means Agentic orchestrates first.
Even in messy environments, agents can deliver value by:
- gathering context across systems
- checking policy/SOPs with citations
- identifying missing information
- routing cases and escalating based on confidence
- producing an audit trail of actions and rationale
As an example:
A bank running an older core plus a separate CRM had frequent onboarding delays because documents were scattered: branch email, portal uploads, WhatsApp images forwarded by relationship teams. An Agentic onboarding coordinator didn’t “approve” onboarding — it assembled the file, flagged missing items, and routed exceptions to the right queue with a reason. Cycle time improved without touching core decision rights.
So yes, data quality matters — but coordination agents can create value without waiting for perfection.
Mimi: Interesting, so where exactly should banks start if they want to go Agentic-first?
Megha: Start where humans already behave like agents — triage, coordination, and exception management — not where you need a perfect decision engine.
Three credible entry points, and I’ll tie each to what I’ve seen in mid-size/regional banks:
1. KYC Coordination Agent
- assembles customer profile from multiple systems
- checks completeness and mismatch patterns
- routes cases based on confidence and risk tier
- escalates exceptions with rationale
Policy interpretation often varies by branch. The agent doesn’t “override policy” — it enforces a standard checklist and flags when branch practice diverges, creating an auditable exception trail.
2. AML Case Assembly Agent
- pulls entity info, transaction history, prior alerts
- drafts investigation narratives + checklists
- recommends next-best actions
Investigators spend disproportionate time compiling evidence rather than judging risk. This agent removes the “manual research tax” while keeping investigators accountable.
3. Disputes Evidence & Routing Agent
- collects evidence artifacts
- drafts comms and case notes
- routes to the right resolution path
Disputes often get stuck between channels (branch vs call center). An agent that standardizes evidence collection and routing is a fast win.
This is Agentic-first that’s realistic: agents run choreography, humans run accountability.
Mimi: Thanks for clarifying Megha. So, If you’re Agentic-first, where do RPA, ML, and GenAI fit? Are they still needed?
Megha: Definitely, but think of them as capabilities the agent calls, not steps you must complete first.
RPA = the agent’s hands for deterministic execution: updating systems, creating cases, moving data when rules are stable.
Example: where APIs are limited, bots still do the “last-mile” updates to legacy cores or branch platforms.
Traditional AI/ML = the agent’s predictor: queue prioritization, anomaly detection, fallout prediction.
Example: an ML model flags onboarding cases likely to fail due to mismatch patterns seen historically in specific regions/products.
GenAI = the agent’s synthesizer: drafting narratives, summarizing cases, interpreting policy with citations, generating standardized documentation.
Example: GenAI standardizes investigation write-ups, reducing variability across branches/teams.
Conclusion : Agent orchestrates + calls the right tool for the right sub-task under guardrails.
Mimi: Okay. This clarifies lot of doubts. Can you give us an end-to-end example where the agent orchestrates and uses these tools.
Megha: Let’s take KYC onboarding in a mid-size bank with a legacy core.
The agent pulls customer details from CRM, core, and onboarding portal; requests missing docs proactively.
It uses GenAI to extract fields and draft a standardized onboarding summary with citations to source documents.
It calls ML to predict which cases are likely to become exceptions — based on mismatch patterns, region/product history, and prior rework loops.
It uses RPA to populate downstream systems when rules are deterministic — because APIs aren’t always available.
When confidence is low or risk is high, it escalates to a human with a clear rationale:
“Address mismatch between ID and account records; similar patterns caused rejection in last 90 days; requesting review.”
Another example: AML investigations in a mid-size bank:
- agent assembles history + evidence
- ML prioritizes alert queues
- GenAI drafts the narrative
- RPA updates the case system
- investigator confirms disposition
Net: faster closure, fewer manual touchpoints, improved audit readiness.
Mimi: It’s fascinating really. Are there any alternate views? Where do you draw the line? Where should banks avoid Agentic-first?
Megha: Yes. Avoid starting with agents as the final decision-maker in high-regret outcomes:
- autonomous credit approve/decline
- autonomous SAR filing decisions
- autonomous high-value fraud dispute closure, or
- High-Volume, Low-Variance Data Entry & Posting
Examples:
- Posting GL entries
- Journal uploads
- Standard transaction settlements
- Straight-through posting from operational systems
RPA or straight-through processing would be sufficient in this case for Deterministic mapping, Validation rules, Reconciliation before posting
Mimi: I see. From credibility perspective, how do you keep this safe and regulator-friendly?
Megha: You raise a very critical point, Mimi.
This is where Responsible AI is not optional—it’s foundational. Agentic AI in banking has to be designed as a controlled operating model, not an experimental capability.
There are five Responsible‑AI principles that matter most when you go Agentic‑first:
First: Clear role definition - The agent must have an explicitly defined role—coordination, preparation, routing, recommendation. If an agent crosses into final risk decisioning, that boundary has to be intentional, approved, and governed.
Second: Human accountability remains explicit. Even in Agentic‑first models, humans remain accountable for:
- regulatory outcomes
- customer impact
- high‑regret decisions
Agents support judgment—they don’t replace responsibility.
Third: Explainability by design, not afterthought. Every agent action should log:
- what data it used
- what rule or policy it referenced
- why it chose a particular action or escalation
This is critical for audit, disputes, and regulator review.
Fourth: Confidence‑based autonomy.
Responsible Agentic AI operates on confidence thresholds:
- high confidence → autonomous execution
- low confidence or policy conflict → human escalation
That’s how you avoid silent risk.
Fifth: Continuous monitoring and bias control.
Banks must track: false escalations, missed risks, drift in outcomes by segment, region, or channel
Responsible AI means not just building controls—but operating them continuously.
In short, Responsible AI allows banks to start Agentic early—without losing trust. Governance is not what slows Agentic AI down - Poorly designed autonomy is.
Mimi: That’s incredible. Megha, this has been a masterclass on Agentic AI. In practical terms, what would you advise banks to do in the next 90 days?
Megha: Thank you—
There are four design conditions I’d suggest banks use to choose their first area:
First: Pick work that is coordination‑heavy, not decision‑heavy.
Look for processes where humans are:
- chasing information
- assembling context
- routing tasks
- managing exceptions
If most of the effort is judgment-free coordination, that’s prime for Agentic‑first.
Second: Avoid areas with irreversible or high‑regret outcomes.
Early Agentic use should not be making:
- credit approve/decline decisions
- compliance filings
- final fraud or SAR determinations
The goal is confidence-building, not risk escalation.
Third: Choose something measurable end‑to‑end.
You want cycle time, rework, touchpoints per case, SLA breaches — not metrics like “AI usage”.
Fourth: Make sure your teams already agree the process is broken, so that the buy-in is quick
If operations teams feel pain every day, adoption happens naturally. If they don’t, even good AI struggles.
Once those conditions are met, then choose the use case.
I would start with disputes and chargeback operations—not because it’s glamorous, but because it’s operationally perfect for Agentic‑first.
Disputes are typically high volume, time-bound, document-heavy, and full of manual coordination between channels, teams, and systems
Here’s what an Agentic coordinator can do:
- interpret the dispute reason and required evidence
- pull transaction details, prior interactions, and policies
- assemble the evidence pack automatically
- draft customer communications and internal case notes
- route the case to the correct resolution path
- escalate only when confidence is low or policy conflicts arise
RPA handles system updates, GenAI drafts narratives and communications, Traditional AI helps prioritize time‑sensitive or high-risk disputes, and the agent orchestrates the entire flow.
Humans still approve final outcomes—but instead of chasing documents and stitching context together, they focus on judgment.
In many banks, this alone:
- reduces cycle time significantly
- cuts rework and customer follow‑ups
- improves SLA adherence
- and strengthens audit readiness
That’s a credible Agentic‑first win—fast, measurable, and low risk.
Mimi: Insightful, my summary of this podcast is that Banks can start Agentic first, even with legacy infrastructure—if they place agents in the right layer of work. Don’t start by making agents the final decision-maker, instead start by making agents the coordinator and conductor of work that today runs through inboxes, spreadsheets, and people.
Megha: Exactly, couldn’t summarize it any better Mimi.
Mimi: Thanks for listening, everyone! If you found this episode valuable, please subscribe and share it within your network.
Also, if you have any queries, do reach out to us through the email address on the podcast description. To contact us, visit us at InfosysBPM.com. Watch this space for more exciting podcasts coming up. Once again, thank you for tuning in, stay safe and sharp. Have a great day!
