Girish Mukim

Posted on Mar 1

BeSA Batch 09 Week2 - Building with Agentic AI

#ai #agents #solutionarchitects

Week 2 – Building with Agentic AI
Batch 09 – BeSA Cloud Academy

Disclaimer:
These notes were drafted using AI for clarity, structure, and readability. They are intended solely for learning purposes.

These are the structured notes from Week 2, focused only on the two role plays. Writing this mainly for my own revision and to continue this learning series.

Role Play 1 – Solution Architect and Customer (Insurance Claims Automation)

Context
This conversation was structured as a discovery session. The customer represents a SAS platform aiming to automate insurance claims processing using Agentic AI.

Business Problem

Current state:

Manual claims processing.
Takes days to weeks.
Error-prone due to:
- Reading bulky documents.
- Cross-referencing policy clauses.
- Performing deep reasoning to accept or reject claims.

Target state:

Faster processing.
Reduced human effort for low-risk claims.
Intelligent automation for complex scenarios.

Technical Requirements and Constraints

Latency

High-volume, low-risk claims should be processed in under 5 seconds.
Complex claims involving reasoning can tolerate higher latency.
Not all steps need the same performance profile.

Data Types

Unstructured documents.
Handwritten forms.
Images of vehicle damage.
The solution must combine text understanding and image analysis.

Integrations

S3 for document retrieval.
Policy databases for clause validation.
Third-party APIs for fraud detection.
Agent must orchestrate across multiple systems.

Team Constraints

Team has DevOps and backend engineers.
Upskilling in AI/ML.
Strong preference for managed services.
Avoid managing infrastructure.
Managed services like Amazon Bedrock are preferred.

Model Selection Framework

The Solution Architect proposed evaluating models across four dimensions:

Capability

Does the model support deep reasoning?
Can it handle multi-step logic?

Latency

Is it fast enough for high-volume workflows?

Cost

Is it sustainable at scale?

Features

Tool usage.
Structured output.
Multi-modal support.

Key insight:

Not every task requires a premium model.
Complex reasoning may need higher-tier models.
Simple extraction can use lighter, cheaper models.

Multi-Model Strategy

Instead of using a single expensive model for the entire workflow, a four-stage pipeline was proposed:

Document extraction
Image analysis
Policy reasoning
Decision generation

Each stage can use a different model optimized for that task.

Result:

Right model for the right task.
Cost reduction estimated at 60–70%.
Performance optimization without overpaying.

Resilience and Governance

Inference Profiles

Used to manage throttling.
Provide resilience across regions.

Application inference profiles

Useful for SAS providers.
Track cost per customer.
Ensure data residency compliance.

Data assurance

AWS does not use customer data to train or retrain models.
Important for compliance-sensitive industries like insurance.

Key architectural takeaway:
Agentic systems must be designed not only for intelligence but also for cost control, compliance, and operational resilience.

Role Play 2 – Enterprise Architect and Solutions Architect
The Evolving Role of the Architect in the Age of AI

This conversation explored how Generative AI impacts the role of Solutions Architects and whether AI reduces or transforms their value.

The Arithmetic vs Mathematics Analogy

Concern raised:

AI can generate complex architecture diagrams in seconds.
Does this make architects obsolete?

Response:

AI handles arithmetic (technical plausibility).
Architects handle mathematics (judgment and context).

Meaning:
AI can produce technically valid architectures.
But it cannot:

Understand politics.
Navigate budgets.
Handle human constraints.

The What vs The Why and How

AI provides:

The “what”.
A plausible list of services and patterns.

Architect provides:

The “why”.
The “how”.
Adjustments based on:
- Finance constraints.
- Organizational policies.
- Stakeholder expectations.

Example:
A technically optimal architecture may not work if:

The finance team requires fixed budgets.
The CTO prefers certain vendors.
Compliance constraints override design decisions.

Research vs Judgment Split

Tasks to hand off to AI:

Knowledge retrieval.
Service comparisons.
Summarizing long documentation (e.g., regulatory texts).
Pattern generation.
Code scaffolding.

Tasks to retain as an architect:

Managing stakeholder biases.
Handling organizational politics.
Making trade-offs (availability vs consistency).
Long-term strategic alignment.
Defending architectural decisions.

Three-Loop Workflow

Loop 1 – Discovery

Use AI as a research assistant.
Summarize meeting notes.
Identify gaps in requirements.

Loop 2 – Design

Use AI to generate architecture drafts.
Validate and refine designs.
Ensure feasibility and alignment with constraints.

Loop 3 – Delivery

Use AI to draft executive summaries.
Draft architecture decision records.
Refine documentation.

AI becomes:

A fast assistant.
Not the final decision maker.

Shift in Identity

Earlier:

Architect’s value was in memorized knowledge.
Service limits, certifications, patterns.

Now:

AI fills the knowledge moat.
Real moat is synthesis.

Synthesis means:

Combining technical capabilities.
Understanding human constraints.
Applying empathy.
Building coherent, real-world architectures.

Strategic Advice

Trust but verify.

AI can hallucinate.
AI can sound confident even when incorrect.

Treat AI like:

A fast summer intern.
Productive.
Needs direction.
Requires oversight before delivery.

Week 2 Consolidated Takeaways

From the first role play:

Agentic AI architecture must align with business latency, cost, and compliance constraints.
Multi-model pipelines are practical and cost-efficient.
Governance and inference profiles are critical in enterprise scenarios.

From the second role play: