Alicia M Morgan

By

How to Build an AI Agent in Microsoft Foundry Now

Filed Under: Uncategorized Tagged With: , , , , , , ,

How to Build an AI Agent in Microsoft Foundry Is More Than a Challenge.

For most project managers, the conversation about AI feels uncomfortable right now.

We’re used to proving ROI through quick wins and PMI’s business value equation: Business Value > Effort + Expenses.

When AI productivity tools mostly handle meeting notes and file organization, leadership asks, “Where’s the real return?”

Many of us feel stuck.
In my case, I’ve lived that tension.

So instead of just studying AI fluency, I built something different: an agent shaped by real cross-industry PM experience.

This article shares the journey from codifying tacit knowledge to building and validating an AI agent in Microsoft Foundry.

The Executive Challenge

Meanwhile, organizations face similar patterns:

  • AI experimentation without consistent ROI evidence
  • Disconnected pilots across teams
  • Inconsistent governance practices
  • Rising risk exposure without clear ownership
  • PMs unsure how to lead change
  • Growing fatigue from hype without results

As a result, executives ask: “How do we move forward with AI responsibly without slowing innovation or damaging trust?”

What validation revealed: PMs are not struggling with AI tools. They are facing the challenge of positioning AI within organizational reality.

What I Built

In practice, I created an AI agent called Cross-Industry Insights in Microsoft Foundry after completing the Foundry FastTrack: From Idea to Prototype Challenge.

About this prototype: This agent was developed in Microsoft Foundry’s playground environment for controlled validation. It is not publicly deployed. This article documents the methodology and lessons learned.

Microsoft Foundry Cross Industry Insights Agent Playground Example.

To ground this work in enterprise reality, at AgentCon Dallas 2025, Microsoft leaders outlined the distinctions between copilots and full-agent platforms.

Enterprise agent development needs the right environment.
That includes evaluation, safety testing, and scalable architecture.

Azure AI Foundry is now Microsoft Foundry. Photo from AgentCon Dallas 2025.

The agent is designed to:

  • Interpret real PM questions about AI adoption
  • Identify situational context based on the operating environment
  • Provide structured guidance aligned to a staged AI-fluency journey
  • Emphasize governance, risk awareness, and business value progression

The framework:
9 Steps from Traditional PM to AI-Fluent Leader.
Not theory. Observed transitions from real work.

My role: Agent architect, and behavior designer.

I made deliberate design choices:

  • Authored detailed agent instructions defining purpose, tone, framework logic, and response structure
  • Structured knowledge files for optimal retrieval and stage mapping
  • Designed scenario-based acceptance tests to validate behavior

The work was translating PM expertise into agent behavior.

Why Pre-Work Mattered

However, the biggest lesson was clear: the work happens before you open the tool.

Before touching Foundry, I spent months codifying what I’d learned.
I built a GitHub repository called the
Cross-Industry PM Playbook for AI Transformation.

It captured:

  • Patterns from regulated aerospace environments where precision matters
  • Lessons from nonprofit work where trust matters more than authority.
  • Insights from corporate transformation where governance enables innovation
  • Frameworks from education settings where adaptation is constant

A pattern emerged.
PMs moving into AI don’t start from zero. They move through recognizable stages depending on the operating environment.

Before building the agent, I already had:

  • A structured 9-step journey with clear stage definitions
  • Three scenario-based test questions covering ROI, governance, and influence
  • Language patterns from PM conversations
  • A defined tone and intent for responses

Key takeaway: Detailed instructions beat clever prompts.

You’re not just telling the model what to answer. You’re telling it who it is.

With that foundation in place, here’s what the agent proved.

What the Agent Proved

As a result, I validated the agent with three acceptance tests — real PM questions about ROI, governance, and influence.

Test Question 1: ROI and Business Value
The system correctly identified the PM’s stage. It distinguished tactical AI usage from strategic business value.
It delivered a four-step approach linking AI improvements to leadership metrics.

Test Question 2: Governance and Risk
The model mapped this to high-stakes environments. It reframed GRC from overhead to speed with seatbelts.
It provided a Day One governance checklist.

Test Question 3: Influence and Identity
The framework recognized environments where influence flows through trust, not authority.
It offered stakeholder mapping and social capital-building tactics.

Four repeatable patterns emerged:

First, stage confusion. Teams often advanced AI efforts without identifying whether they were experimenting or piloting.
The agent surfaced the real adoption stage before laying out the next steps.

Second, business value translation gaps. AI activity was easier to describe than AI value.
The agent reframed work into executive-relevant outcomes.

Third, governance timing misalignment. Governance was introduced too late or too early.
The agent identified optimal timing without stalling momentum.

Fourth, influence without authority. PMs lacked structured approaches to build executive trust.
The agent emphasized sequencing and stakeholder mapping.

The agent diagnoses context first, then tailors guidance.
Pattern recognition makes this possible.

The Pattern Behind the Agent

Previously, long before I ever touched AI tools, early in my career at Raytheon, I led a $2.5M factory rearrangement that saved $3.2M and avoided 20,000 labor hours.

What made it work was structured empathy.
I listened to facilitators and operators.
At the same time, I mapped invisible constraints.
Before any changes, I built trust.

Years later, as I led nonprofit programs serving 45,000 students across 100+ institutions, I recognized the same pattern.
Different stakeholders. Same need for structure that enables movement, not blocks it.

This work earned recognition as a 2019 Dallas Business Journal Women in Technology Awards Advocate Honoree.
Recognition from TEDx and PMI global webinars followed.

Organizations fail at AI adoption for a simple reason.
They struggle to translate between operating contexts, risk tolerance, and governance cultures.

That realization led to another lesson.

What I Learned

Nonetheless, not everything worked immediately.

Early instructions were too vague. The agent drifted into generic advice.
I learned to be explicit about worldview, not just outputs.

Teams frequently conflate experimentation and piloting. Scenario testing exposed gaps.
The agent was less effective on technical questions but strong in organizational navigation.

Current constraints:
The agent performs best with structured PM questions on adoption, governance, and influence.
It is less effective in highly novel situations.

The prototype remains in a controlled playground environment.
Production deployment would require organizational sponsorship, security review, and governance approval.

So the natural question became:

Can This Scale?

Therefore, my background gives me confidence that it can.

I’ve spent years scaling programs across regulated industries: aerospace, nonprofit, corporate enterprise, and education.
I’ve guided initiatives where frameworks flexed across environments.
Along the way, I set governance that supported progress rather than slowing it.
In parallel, I earned trust without formal authority.
As a result, I showed value in ambiguous situations.

The prototype was built in an enterprise-aligned environment.
It supports expanding knowledge artifacts.
It also supports integration into internal PM playbooks and shared AI adoption language.

This positions the work to move from pilot to scale.

More importantly, it changes how organizations talk about AI.

What This Changes for Organizations

Consequently, when a PM uses this agent-guided approach, conversations change.

Before: Disconnected AI pilots with unclear ownership, unclear value, and delayed governance.
After: AI initiatives are positioned within a staged adoption journey with defined next actions, evidence expectations, and risk checkpoints.

Before: “We’re testing some AI tools.”
Executive hears: “Unclear cost. Unclear risk. Unclear payoff.”

After: “We’re in Stage 2. The pilot is defined at the outset. Evidence emerges to support decisions.
Connections to cycle-time reduction become visible. At that stage, governance enters the process.
Within sixty days, outcomes become clear.”
Executive hears: “This is controlled. Measurable. Safe to invest in.”

Business Value

In addition, here’s what that value looks like.

For PMs:

  • Identifies true AI adoption stage
  • Translates AI activity into business value
  • Sequences next actions realistically
  • Provides a confidence scaffold for executive conversations

For executives:

  • Replaces vague experimentation with staged adoption visibility
  • Surface governance and risk need to be addressed early
  • Connects initiatives to measurable outcomes
  • Establishes repeatable operating language

Version Progression

Version 1 (Current Prototype)

  • Core framework encoded
  • Agent instructions authored
  • Knowledge artifacts structured
  • Scenario-based tests executed
  • Behavioral feasibility validated

Version 2 (Planned Expansion)
Introduces The Evidence Ladder — a framework proving AI value, building trust, and scaling responsibly.

It helps PMs link AI actions to operational metrics, map evidence to executive thresholds, and show incremental value without premature ROI claims.

Version 3 (Planned Operational Model)

    • Human-in-the-loop design
    • Governance alignment for organizational deployment
    • Pilot-to-scale readiness
    • Integration with internal PM playbooks
    • Production environment deployment considerations

—

Agent Development Series

This article introduces my first agent prototype. For comprehensive technical documentation, methodology deep dives, and updates on additional agents under development throughout 2025, visit my Innovation in Action repository

This is where the Cross-Industry PM Playbook evolves into working AI systems.

—

Key Takeaways

Finally, here are the core takeaways.

Start with pre-work. Codify learned expertise before touching development tools.

Choose enterprise-aligned platforms. Microsoft Foundry supports evaluation, safety testing, and governance controls.

Design for behavior, not features. Define who the agent is, not just what it does.

Validate with real scenarios. Build acceptance tests mirroring actual user challenges.

Expect iteration. Early versions drift. Scenario testing reveals gaps.

Pattern recognition scales. Cross-industry experience is your advantage.

Resources

Closing Thought

AI fluency is not about learning tools.
It’s about translating lived experience into new environments.

Project managers already know how to navigate ambiguity, align stakeholders, deliver outcomes, and earn trust.

This agent gives that knowledge a new shape.

Executives don’t need PMs who only understand tools.
They need PMs who move AI forward with evidence, structure, and credibility.

That’s the problem this work solves.