Read My Blogs

Briefing on Generative AI Strategy, Architecture, and Program Leadership

Executive Summary

Agentic AI – Using function calling, RAG, and the Model Context Protocol, agents solve complex, unpredictable problems dynamically.

This document synthesizes a strategic framework for integrating Generative AI (Gen-AI) within a large enterprise, focusing on a disciplined architectural approach, high-impact use cases, and the principles of effective program leadership. The core philosophy is to leverage Gen-AI as a powerful assist layer for insight and acceleration while maintaining control, stability, and governance within existing automation and workflow platforms. This is encapsulated in the guiding principle: AI recommends, Platforms decide, Automation executes.

The proposed target architecture unifies disparate tools (Pega, UiPath, APIs) under a single, headless orchestration layer. This “API-first” model establishes a standard entry point for all requests, intelligently routing tasks to the most efficient executor—whether a modern API or a Robotic Process Automation (RPA) bot acting as an adapter for legacy systems. This approach, summarized by the flow Copilot -> Orchestration -> APIs -> Events, ensures modernization can occur gradually without disrupting service consumers.

In this future state, Gen-AI is poised for massive, bank-wide adoption across six key areas: Document Intelligence, Automation Development (Autopilot), Workflow Creation (Pega Blueprint), Developer and QE Productivity, Agentic AI for “Autonomous Operations,” and Enterprise Knowledge Management via copilots. These applications represent a shift from isolated task automation to intelligent, end-to-end journey automation. Successfully delivering this transformation requires a program leadership model founded on cross-functional alignment, structured end-to-end delivery, strategic communication, and the use of AI tools to enhance program visibility and efficiency.

1. Defining Generative AI and Large Language Models (LLMs)

A clear understanding of Gen-AI and its core components, Large Language Models (LLMs), is foundational to the strategy.

The Relationship Between Gen-AI and LLMs

Generative AI is a broad field of artificial intelligence focused on creating new content. LLMs are the critical deep-learning models that power modern Gen-AI by understanding and generating human-like language. The relationship can be understood through an analogy:

As one source document states, “LLM is one of the most well-known types of Gen AI.”

How LLMs Function

LLMs are not knowledge databases that memorize facts; they are statistical models that learn linguistic patterns. LLMs as “models that understand and generate human-like text using deep learning and massive datasets.”

Their process involves:

  1. Training: Ingesting massive datasets of text from books, articles, websites, and documents.
  2. Vectorization: Converting text into multi-dimensional numeric vectors.
  3. Pattern Recognition: Learning the statistical relationships and patterns within the language.
  4. Prediction: Predicting the next most likely word or sequence of words to generate coherent, context-aware responses.

Notable examples of LLMs referenced include OpenAI’s GPT-4o, Cohere Command, Google Gemini, and Meta Llama.

Core Enterprise Benefits of LLMs

The integration of LLMs into an enterprise enablement platform provides several key benefits:

2. A Framework for Responsible AI Integration

To balance innovation with the stability required in capital markets and other high-risk environments, a structured framework for integrating Gen-AI is essential. This involves clear architectural boundaries and a multi-tiered platform approach.

The Guiding Principle: “AI Recommends, Platforms Decide, Automation Executes”

This simple yet powerful framework is used to explain the architecture to executive leadership and mitigate fears of “AI running the bank.” It clearly delineates roles:

This model preserves governance and ensures that the fast-changing, probabilistic nature of Gen-AI does not introduce risk into stable, deterministic production systems.

Architectural Model for Onboarding Gen-AI

To enforce this principle, Gen-AI is onboarded as an “assist layer,” not a replacement for the core automation estate. The prescribed architectural flow is:

$Gen-AI → API Gateway → Workflow (UiPath / Pega) → Bot or human task.$

A key lesson learned from an early failure was to never embed Gen-AI logic directly inside an RPA bot. Bots must remain “boring” and deterministic, consuming only validated outputs from Gen-AI via controlled APIs. This boundary became a platform rule, increasing automation coverage and reducing bot failures while satisfying risk and audit requirements.

The Three-Runway Platform

To resolve conflict between data science (seeking experimentation), engineering (seeking stability), and business (seeking speed), the platform was split into three distinct “runways”:

  1. Research: A space for Gen-AI exploration and experimentation, isolated from production systems.
  2. Enablement: A layer providing shared, governed services like headless orchestration APIs, standard logging, evaluation metrics, and guardrails. This is where Gen-AI onboarding is centralized.
  3. Production: The stable environment where RPA and workflow automation execute deterministic tasks.

This structure allows teams to innovate without blocking each other or introducing risk to production operations.

3. Target State Architecture for Enterprise Automation

The future-state vision moves from a collection of siloed tools to a unified, intelligent platform where capabilities are exposed as services.

Current State vs. Future State

Current StateFuture State (Target Architecture)
Many tools act separately (Pega, UiPath, FileNet).All tools exist but are unified under a single entry point.
Teams manually choose which tool to use.Headless Orchestration routes requests to the best tool.
Bots often fill gaps where APIs don’t exist.APIs-first becomes the default; bots act as legacy adapters.
No single “brain” coordinating work.Event streaming provides real-time updates across the bank.
AI is used in small, isolated pockets.AI is layered across the platform for extraction, guidance, and assistance.

The Core Architectural Flow: Copilot → Orchestration → APIs → Events

This flow describes how work is initiated and processed in the target state:

  1. Copilot (Ask): A user makes a request in plain English through a copilot interface.
  2. Orchestration (Decide & Route): The platform’s central “brain” analyzes the request and determines the best execution path (e.g., call a document API, initiate a workflow, dispatch a bot). The user does not need to know the underlying tool.
  3. APIs (Do the Work): The orchestrator calls the appropriate API. If a modern API doesn’t exist for a legacy system, an RPA bot is triggered behind the scenes to perform the UI-based task, acting as a stand-in for a future API. The caller is unaware of this distinction.
  4. Events (Tell Everyone): Once the action is complete, an event (e.g., document_fetched) is published. This allows dashboards to update, AI models to learn usage patterns, and other systems to react in real-time.

A one-line summary of this vision is: “We unify all tools under one brain (orchestration), one language (APIs), and one heartbeat (events), with Copilot as the front door.”

4. Strategic Enterprise-Scale Applications of Generative AI (“The Big 6”)

The future-state architecture enables Gen-AI to become a major, bank-wide capability. The most significant usage will be concentrated in six transformational areas.

AreaKey Technologies & ConceptsWhy It’s Massive
1. Document IntelligenceIBM WatsonX, RAG agents for document interaction (summarize, explain, compare).Affects millions of documents in FileNet, ERT, etc. Becomes the front door for all document knowledge, enabling end-to-end straight-through processing.
2. Automation DevelopmentUiPath Autopilot generating RPA workflows from natural language.Touches every automation team (A&W, RPA COE). Provides a 40-60% faster bot build time, creating a mass productivity uplift.
3. Workflow CreationPega Blueprint using Gen-AI to turn ideas into functional workflows.Reduces build time from months to days. Impacts nearly all enterprise workflows across Credit, Fraud, Banking Ops, Wealth, and Claims.
4. Dev & QE ProductivityRovoAI (Jira/Confluence), BrowserStack AI (test generation), FIGMA AI (UX), Deque AI (accessibility).Pervasive impact across all SWE and QE teams. AI assists in creating test scripts, UX artifacts, design documents, and stories.
5. Agentic AIAI agents that can plan, reason, and act across multiple tools, orchestrated by Maestro or Pega.Moves the bank from task automation to journey automation. Agents can autonomously execute multi-step processes like regulatory reviews.
6. Enterprise KnowledgeAI-powered copilots with RAG, exposed via a headless orchestration API.Creates a unified AI knowledge layer for all teams (Delivery, Ops, Business). Becomes the single entry point for institutional knowledge.

5. Principles of Effective AI Program Leadership and Execution

Technology and architecture alone are insufficient. Delivering on this vision requires disciplined program leadership focused on alignment, execution, and communication.

Core Competencies

Case Study: Enterprise Automation Platform Migration

A case study demonstrates these principles in action:

Description
SituationAn organization needed to migrate 100+ business-critical automation processes, upgrade core platform components, and reduce a high volume of incidents while maintaining uptime.
TaskLead a multi-team program across engineering, Ops, infrastructure, risk, and business stakeholders, involving 12 pods and over 150 contributors.
ActionA multi-phase plan (readiness, migration, validation, stabilization) was created. A unified execution framework was implemented to coordinate all pods. AI-powered reporting tools were built to summarize progress and blockers. Strict cut-over, rollback, and monitoring plans were established, with weekly structured updates provided to leadership.
ResultThe migration was completed ahead of schedule with zero major outages. Incident volume was reduced by over 70% in subsequent quarters, and platform reliability, governance, and visibility were significantly improved.

Leveraging AI for Program Management Productivity

Program management itself can be enhanced with Gen-AI and low-code tools: