From experiment to enterprise: Real lessons in building an AI agent

Over the past few months, Pega employees have been collaborating with a talented new co-worker: Intern Iris. Iris works remotely, but is available at all hours. She communicates solely through email. Her skills started out limited, but she is constantly learning how to handle new tasks. Iris is an AI agent.

When we first embarked on the journey to create Intern Iris, we set out to build more than just another chatbot. We wanted to develop an intelligent agent that could understand, execute, and optimize complex workflows while maintaining the highest standards of enterprise security and compliance. Along the way, we encountered real challenges that taught us valuable lessons about deploying AI agents in an enterprise environment.

The opportunity: Make every team member your best

Initially, we started our agentic journey to help arm client account teams with knowledge about the market and how they can better serve their clients. But we quickly realized that the opportunity for upleveling team members across functions was massive. Iris continues to evolve with new skills and knowledge across sources and our team finds new ways to leverage it every single day. Here are some of the early results we've seen:

• Simplified onboarding: Rather than taking weeks to learn how to navigate complex systems and processes, new team members now engage in natural conversations to quickly grasp workflows and organizational structures. Result: onboarding time dramatically reduced.
• Faster development: Rather than spending hours manually searching through scattered documentation, developers can now rapidly access implementation best practices and receive contextual guidance that ensures compliance with development standards. Result: hours saved per implementation.
• Supported account teams: Rather than spending days manually assembling and validating customer materials, account teams can now rapidly generate comprehensive documentation with integrated case studies and industry-specific requirements. Result: massive reduction in preparation time while maintaining quality.
• Strategic insights at your fingertips: Rather than starting from scratch with each new initiative, strategists can now accelerate ideation and content development through instant access to market insights and creative suggestions, all aligned with strategic messaging. Result: significantly enhanced strategic output with faster time-to-market.

And the roll out has sparked a new dimension of creativity within the organization. Team members across the board are now shifting from sharing documents to sharing prompts with one another, encouraging new ways to leverage Iris to increase productivity, effectiveness, and efficiency.

Learning from early missteps

One of our most significant early lessons came from an incident involving email automation. During initial deployment, Iris had the capability to send emails but lacked proper governance controls. This led to a situation where she directly emailed executives at a customer organization without first obtaining approval from the sales team - a clear violation of our business protocols.

This incident taught us a crucial lesson: AI agents need more than just technical capabilities; they need robust governance frameworks and business context to orchestrate workflows across stages, steps, and people. In response, we implemented several critical controls:

• Mandatory approval workflows for external communications
• Role-based access controls for different types of actions
• Clear delineation between internal and external communication protocols
• Automated checks against established business rules before any external action

This experience perfectly illustrates why what Gartner now calls BOAT (business orchestration and automation technologies) is essential for enterprise AI success. Strong orchestration isn't just about efficiency - it's about best practices, safety, and compliance.

Agents and workflows

Through developing Iris, we identified and refined three critical capabilities that any enterprise AI agent needs:

1. Workflow design: Just like how Pega GenAI™ Blueprint transforms ideas into application designs, we taught Iris to understand business requirements and suggest optimal workflow structures. However, early experiences showed us that unrestricted workflow creation could lead to compliance risks. For instance, Iris once designed a document processing workflow that bypassed required legal reviews. We solved this by implementing a hybrid approach - enabling AI-driven workflow design while maintaining human oversight and governance checkpoints.
2. Workflow execution: Real-world execution proved more complex than anticipated. We discovered that Iris needed not just to follow instructions but to understand business context. For example, when handling customer inquiries, she initially treated all requests with the same priority level. We enhanced her capabilities to recognize urgency levels, compliance requirements, and appropriate escalation paths. This is where strong orchestration technology proved invaluable, ensuring every action aligned with established business rules.
3. Workflow optimization: We implemented mechanisms for Iris to analyze patterns and suggest improvements, just like we do with Pega Process AI capabilities. A practical example: Iris identified that certain approval workflows were creating bottlenecks during specific time zones. This led to implementing automated time-zone aware routing and parallel approval paths, significantly improving efficiency.

Enterprise governance by design

Our early experiences taught us to build safety controls from the ground up. We now follow several key principles, each born from specific incidents:

• Implement strong authentication and access controls (after the unauthorized email incident)
• Maintain detailed audit trails of all AI actions (introduced after discovering gaps in tracking decision patterns) · Establish clear boundaries for autonomous decision-making (developed after several instances of overambitious automation attempts)
• Create fallback mechanisms for uncertain situations (implemented after observing hesitation in edge cases)
• Regular validation of outputs against business rules (instituted after finding inconsistencies in generated content

The role of orchestration

"The success of Iris ultimately came down to strong orchestration technology - what we know Gartner now recognizes as BOAT."

This framework ensures that AI agents don't just execute tasks in isolation but operate as part of a cohesive enterprise automation strategy. Our email incident perfectly demonstrated why orchestration isn't just about efficiency - it's about ensuring AI operates within appropriate business boundaries while maintaining security and compliance

Looking ahead

As we continue to evolve Iris and shape the broader landscape of enterprise AI, it's clear that the future lies in this orchestrated approach to intelligent automation. The key is not just in having powerful AI capabilities, but in how we orchestrate these capabilities to deliver real business value while managing risk.

The future of work isn't about AI agents operating independently, but about how well we can orchestrate their capabilities within our existing business processes and systems. That's where the real transformation happens, and that's where platforms like Pega - with their strong orchestration capabilities - will play a crucial role.

Want to start your own journey toward orchestrated intelligent automation? Let an AI agent design your next workflow! Try it now with Pega GenAI Blueprint at pega.com/blueprint.

Public link

Please use the above public link if you want to share this noodl on another website.