What I Learned from Building AI Agents
Enterprise employees waste 21 full days per year on repetitive tasks, creating $50 billion in annual bottleneck costs. In this post, I break down Microsoft’s artificial intelligence (AI) strategy to address this issue and reveal what I learned from building AI agents for my own business. My real-world implementation faced unexpected challenges. The results taught me that AI’s greatest value isn’t just efficiency, it’s expanding how we approach problems entirely.
“What I Learned from Building AI Agents” Podcast
Picture this: Your sales team waits weeks for legal approval of a standard contract while a hot prospect grows cold. Meanwhile, your marketing team manually tracks campaign metrics across five different platforms, and your finance department gets bogged down building spreadsheets to model cash flow and revenue projections.
Removing Workplace Bottlenecks Creation
These everyday bottlenecks cost enterprises $50 billion annually, while the average enterprise employee wastes 21 full days per year on repetitive, menial tasks that could easily be automated. McKinsey found that “about half the activities people are paid to do globally could theoretically be automated using currently demonstrated technologies. In about 60 percent of occupations, at least one-third of the constituent activities could be automated, implying substantial workplace transformations for all workers.”
But here’s the question: Which AI platform will lead this workplace transformation and remove these bottlenecks?
Microsoft’s Strategic Evolution: From Browser AI to Enterprise Ecosystem
Back in early 2023, we asked “Is Microsoft Poised to Become Our AI Copilot?” as the company partnered with OpenAI to build AI-powered search within Bing and Edge. At the time, Microsoft’s long-term vision was to incorporate this AI technology across its entire product suite.That vision has now become reality: Microsoft has seamlessly woven Copilot into the tools 400 million+ paid seats and 3.7 million companies use daily—i.e., Word, Excel, PowerPoint and Outlook—creating an ecosystem where AI isn’t just a separate tool you switch to, but rather an integrated partner in your everyday workflow.This integrated approach gives Microsoft a significant competitive advantage. While other AI platforms require you to learn new tools and build custom integrations, Copilot transforms the applications we already use daily. Why would an enterprise invest in standalone AI platforms when Copilot seamlessly enhances Word, Excel, PowerPoint, and Outlook while also connecting to thousands of third-party applications through established Microsoft integrations?
This creates what some might call a “walled garden,” but for enterprises drowning in inefficient processes, it’s far more practical. The switching costs and learning curves of alternative platforms create barriers to adoption, along with compliance concerns, while Microsoft’s approach removes friction entirely.
Copilot Studio: Where Human Experience Converges with AI Agents
Last week at Microsoft Build 2025, Microsoft announced new multi-agent capabilities in Copilot Studio, Microsoft 365 and its Azure products, all of which seamlessly work together to deploy agents that, as Microsoft’s Chief Product Officer Aparna Chennapragada explains, will be able to “exchange data, collaborate on tasks, and divide their work based on each agent’s expertise.” This allows enterprise businesses to hyper-scale their workforce and productivity. As Microsoft’s Aparna Chennapragada explained, AI agents become even more valuable when they can “speak like you do, think like you do, and work like you do in terms of your own business”. Essentially becoming extensions of your organization’s unique expertise and culture.
Microsoft’s Chief HR AI Officer and Corporate Vice President, Human Resources, Christopher J. Fernandez shared the practical applications as a way to facilitate three types of agents:
Retrieval: These agents can retrieve information from data they are grounded in, reason, summarize and answer user questions.
Task: These agents take actions when asked, automate workflows and replace repetitive tasks for users.
Autonomous: These agents operate independently, dynamically plan, orchestrate other agents, learn and escalate.
What Mr. Fernandez stresses is the importance of “human-to-AI integration, not separation.” In all cases, agents function as an extension of human knowledge, human insight and your vision for scaling what you, as the domain expert, view as important.
Here’s how Microsoft Copilot agents could work collaboratively: “The system enables scenarios such as a Copilot Studio agent pulling sales data from a CRM, handing it to a Microsoft 365 agent to draft a proposal in Word, and then triggering another agent to schedule follow-ups in Outlook.”
Enterprise Results that Validate the Vision
Microsoft’s vision isn’t just theoretical: Early enterprise adopters are seeing remarkable results that validate the integrated approach. For example, Wells Fargo deployed an agent supporting 35,000 bankers, providing instant access to 1,700 procedures and slashing information search time from 10 minutes to just 30 seconds. T-Mobile created an agent that connects to over 20 device manufacturers’ websites, instantly assembling product information that previously required manual research across multiple sites. HCLTech’s agents resolved support cases 40% faster while allowing it to redeploy 30% of its 500-person support staff to higher-value work.
These results demonstrate exactly why Microsoft’s product ecosystem creates such a competitive moat—and why I wanted to see whether similar outcomes were achievable for a small business like mine.
My Real-World Test: Building Multi-Agent Workflows for ScheinerInc.
The Problem I Wanted to Solve: As the owner of ScheinerInc., a brand and digital marketing consultancy, I spend countless hours balancing my client responsibilities with prospect tracking, engagement analysis for posts like these and follow-up coordination—exactly the kind of repetitive work that should be automated. So, I decided to build my own multi-agent system to see if I could address my AI agent needs, and let’s just say that it didn’t go as smoothly as the demos suggest.
The first task was mapping out what I needed. I identified three core agents:
Agent 1: Engagement Analyst: This Agent would monitor and analyze my email reports to:
- Identify highly engaged readers
- Track which blog topics generate the most interest
- Identify engagement patterns (time of day, device type, content categories)
- Develop lead-scoring methodologies
Agent 2: Outreach Coordinator: The Outreach Coordinator would help me maintain relationships with prospects and network with contacts by:
- Creating categorized prospect lists based on engagement data
- Drafting personalized follow-up messages for my review
- Maintaining a calendar of follow-up touchpoints
Agent 3: Content Matchmaker: Following my content library spreadsheet and matching rules criteria, the Content Matchmaker would:
- Maintain a database of my downloadable materials and blog posts
- Match content to prospect interests based on their engagement history
To support my team of AI agents, I planned to::
- Review and approve outreach messages
- Provide feedback to refine agent recommendations
- Update the content library as new materials and blog posts are created
The Implementation Reality Check
Working with Claude AI, I plotted out a realistic implementation path that didn’t require programming. Claude provided step-by-step instructions using Google Drive and Zapier, a platform that helps you develop and automate repetitive workflows between apps with no coding required.
The reality? It was incredibly frustrating due to integration issues between Zapier and Google Sheets that proved impossible to resolve. The main problem was that Zapier’s CSV parsing couldn’t handle my Saleshandy report format, and manual workarounds created more bottlenecks than they solved. After two days and about 10 hours of trying, I had to settle for a much simpler version than what I had set out to build.
What I Actually Built
While not the seamless automation I envisioned, the system I ended up building now provides:
- Automated detection when I upload email reporting spreadsheets
- Structured tracking for high-engagement readers
- A content-matching framework based on reader interests
- Consolidated workflow that saves me three to four hours weekly
Although it’s a significant improvement over manual execution, it’s nowhere near the enterprise-level automation that Microsoft demos.
The Unexpected Value: AI as Thinking Partner
Despite the technical challenges, I learned three valuable lessons from building even this simplified system:
• Understanding AI Possibilities: It allowed me to understand the capabilities of AI and build workflows I would have never been able to create manually.
• Reimagining Work Processes: If we’re going to evolve, we constantly need to explore and take risks.
• Expanding Creative Possibilities: My friend Neil Perkin elaborates on his post about AI and inflection points, and how experimentation drives transformation, noting that “New technologies (like AI) are not just about efficiency, but about enriching creative possibilities.” He’s absolutely right. I discovered that AI’s greatest value isn’t just streamlining bottlenecks; rather, it’s expanding our way of thinking.
Think about it: you can present countless scenarios to AI and ask it to react or provide various outcomes, helping humans make more informed decisions based on questions we would never have otherwise considered. While it does create efficiencies, it does something even more important: it expands our cognitive reach. AI became my thinking partner, helping me consider scenarios and solutions I would never have explored or thought of alone.
The Future of Human-AI Collaboration
I believe that the more we view AI as a collaborative partner, the better problem-solvers we become. As I detailed in my earlier post, “The path to effective human-agent teams isn’t just about deploying technology; rather, it’s about transforming organizational mindset.”
Microsoft’s integrated ecosystem may have the competitive advantage in the enterprise market, but the real opportunity lies in how we fundamentally rethink our relationship with intelligent systems. Whether you’re using Copilot Studio or building workflows with other tools, the key is approaching AI as a thinking partner, not just an efficiency engine.
So, while AI does create efficiencies and streamline bottlenecks, it does something even more important: it expands our way of thinking.
As always, thank you for reading or listening. Questions? Feel free to email me here.
