Swag Discussion Series | Week 4

I initially thought the challenge would be building good automations. What actually required much more intention was creating a repeatable system for identifying, preparing, designing, governing, deploying, and sustaining automations at scale.

And that's exactly why my recent thinking around agentic AI + the Case Backlog, the AA AD&M proposal, the flagship use case, and my insistence on defining the Problem Statement first are all connected. (Always defining the Problem Statement is my disciplined Lean Six Sigma learning background)

I find myself strategically moving from "How do we automate this?" toward "How do we systematically determine what should be automated, how it should be automated, and how we ensure the automation produces the intended business outcome?"

That's the scaling inflection point.

And honestly, at first, I thought it was he PDD, but now I’d put that ahead of the PDD as the biggest "I thought this would just work" lesson.

I thought agent orchestration would just work once the individual agents were performing well, but scaling taught me that clear handoffs, context management, and exception handling need proactive planning. I learned that aligning stakeholders on responsibilities, guardrails, and success criteria is just as important as getting the technology right.

I used to think that orchestration happened as an isolated functionality embedded within the agent itself or within the skill itself. Now that I have a better understanding of the Automation Anywhere landscape, I realized that orchestration takes place through the centralized platform and Mozart, because it allows you to create specific conditions that enable you to track the usage of IDP, RPA, and AI, and interconnect the workflow without losing important characteristics such as security, auditability, and scalability.

1: AI Prompting vs. Deterministic Code (Best for standard RPA background)

“Before the course, I anticipated that integrating AI into Automation Anywhere would mostly involve writing strict, complex logic to handle edge cases. What made me rethink this was learning how to leverage Generative AI packages and prompt engineering. I realized that instead of hardcoding every possible route, the real skill is crafting tight guardrails, structured prompt templates, and confidence thresholds so the AI can handle unpredictable data autonomously."

2: Document Extraction & Edge Cases (Best if you focused on IQ Bot / Document Automation)

“I expected that training AI models for Document Automation would be the most difficult technical hurdle. However, the training made me rethink where the actual effort goes—it isn't just about training the model, but designing seamless Human-in-the-Loop (HITL) workflows. I learned that 100% accuracy from AI isn't realistic, so orchestrating how human validation handles low-confidence exceptions is what actually determines a project's success."

3: Process Discovery vs. Automation Construction (Best for process-focused modules)

“I originally assumed the AI Engineer role would be purely about bot construction and API integrations. The training made me rethink this by showing that automating a bad process with AI just creates faster errors. Using tools like Process Discovery shifted my focus from 'How do I build this bot?' to 'Is this process structured enough for AI to handle reliably without breaking upstream systems?'"