How Leaders Build Their AI Agents

Last week, I did something completely different—I took a two-day break from AI to explore dinosaur skeletons.

Yes, quite the opposite of AI.

Stepping back gave me a valuable reminder: AI is a fantastic collaborator, but human creativity requires moments of disconnection and reflection.

So when I read Angela Yang’s “​Need to attend a meeting, order groceries or book a flight? There's an 'AI agent' for that​”, I reflected on a fundamental truth.

Too often, leaders rush into AI without taking a step back to see the bigger picture.

Because as we’re nearing the end of our Implementation Program, a four-week trajectory where ​Lead with AI​ graduates turn theory into action, some themes emerge.

Learning from those, here are five steps to elevate how you think about building your AI agents.

Five Steps to Successfully Build AI Agents

1. Don’t Start with AI—Start with the Problem

In a previous edition, "​Successful AI Implementations Start Specific​," I already shared about choosing the right AI tasks.

Most AI failures happen because people jump straight into execution without truly understanding the problem they’re solving.

The best AI applications focus on tasks that meet my ​G.E.D. + R.​ Framework:✅ General – Could a great generalist do the job?✅ Error-Friendly – Is there room for mistakes – since AI will make them?✅ Digital – Is the work online? AI can’t cook yet.+✅ Recurring – Apply AI where you spend the most time.

Added to that, consider Joy. A recent joiner, Jael Chng, was ​surprised​ to hear me mention this. But it makes sense: a task could be perfectly G.E.D.+R., but if you enjoy doing it - why apply AI to it?

If AI isn’t saving time, reducing costs, or improving efficiency, it’s not the right tool for the job.

2. Write Out Your AI Process Before Touching Any Tools

I see this mistake all the time: Someone gets excited about AI, starts setting up a new tool, and then gets stuck.

Instead, write down your task and the ideal outcome of your AI implementation.

Then, map out the entire AI workflow:

1. What triggers the AI? (e.g., an email, a document upload)
2. What input does it need? (e.g., a customer request)
3. What does AI process? (e.g., summarizing or classifying data)
4. What happens to the output? (e.g., saved to a CRM, emailed to a user)

We did this on a Figjam board and had a lot of fun.

Once you’ve written out the flow of your “Minimum Loveable Product” (MLP), ask ChatGPT to roast you.

Is what you’ve conceived actually the best way to reach your objectives?

3. Test AI Manually Before Automating Anything

Too often, people rush into Zapier, Power Automate, or API integrations before knowing if AI even generates useful responses.

To avoid this, test out your idea in a regular ChatGPT chat window.

• First, test AI outputs manually in ChatGPT, Claude, or Gemini.
• Tweak your prompts until the response is consistently useful.
• Only then, automate it into a workflow.

Let’s take an example: A finance team wanted to automate invoice categorization. Instead of jumping into automation, they tested AI in ChatGPT to see if it could accurately classify invoices. Only after verifying accuracy did they integrate it into an automated workflow.

4. Start Small, Get Quick Wins, Then Scale

AI projects fail when you try to solve everything at once.

The best implementations solve ONE problem well, then expand.

For example, instead of launching a full-scale AI chatbot, start with one feature, like:

• Automating appointment scheduling
• Generating meeting summaries
• Improving email drafts

Once that works, expand the AI’s capabilities step by step.

You’ll learn so much from your own experience or user feedback, which is a great way to improve your solution over time.

5. Choose the Right AI Tool for the Job

Not all AI tools are created equal. Some overpromise and underdeliver, while others pose security risks.

When choosing an AI tool, ask:

• Does it integrate with your existing systems?
• Does it have strong security & compliance?
• Is it cost-effective for long-term use?
• Will the provider still be around in 3 years?

These questions are why I do almost everything in the Microsoft and Google ecosystems rather than trusting some startup with my sensitive data.

The Bottom Line: AI is a Journey

In short, successful AI implementation isn't about using the latest tool—it’s about:

✅ Defining the right problem

✅ Mapping the workflow before touching tools

✅ Testing AI manually before automating

✅ Starting small and scaling strategically

✅ Choosing AI solutions that are secure and sustainable

Taking this into account, your AI solution may just be around as long as the dinos in the photo.

Now, I’d love to hear from you:

• What AI implementation are you currently working on?
• What’s your biggest challenge in making AI work for your business?

Let me know—I’m always curious to hear what’s working for you!