Insights From Microsoft for Building High-Value, Scalable AI Projects

Many organizations are eager to adopt artificial intelligence (AI) but uncertain about where to begin, which projects to choose, or how to avoid costly missteps. Recently, our CEO, TVN Reddy, explained why moving forward with AI requires courage and a willingness to learn. If you missed that piece, you can read it here: Turning AI Anxiety Into Advantage.

To help companies overcome this fear of AI and develop programs with a clear purpose and a realistic path to production, we sat down with Eduardo Kassner, CTO of Cloud & AI for the High Tech Sector at Microsoft.

Eduardo Kassner has guided more than 900 AI projects across global technology organizations, giving him an expert perspective on what separates successful AI initiatives from expensive experiments.

In our conversation, Eduardo explained how companies can identify genuine AI opportunities from a broad list of ideas. He also outlined Microsoft’s three-step model that can help you more easily turn early concepts into solutions that deliver value in your daily operations. As Eduardo often reminds his team, “catching the AI tiger” is only the beginning—you also need a strategy for how to nurture it within your care.

How Your Organization Can Identify Viable AI Use Cases

Q: Most companies approach AI with expansive wish lists and ambitious timelines. How can they cut through the noise to identify genuinely viable use cases?

Eduardo Kassner (EK): The worst thing you can do with AI is invest huge amounts of money without a solid business case. Otherwise, you’ll end up asking "what did we get out of this?" That’s a shame, because AI is a super useful tool—so long as you understand what it's for.

No AI project is equal to another, so you need to find something important enough that pushes people to do the research and solve a problem.

Q: What framework does Microsoft use to separate high-impact AI opportunities from expensive experiments?

EK: We apply three filters to every project.

First: Is there a clear purpose backed by real need? Not something that's interesting, but something a key stakeholder can't live without—and that your CEO will agree is worth investing in. It needs to matter all the way up to board level.

Second: Can you measure it? Not with a general benefit statement that can’t be verified, but with measurable results which nobody can dispute.

Be careful about the metrics you choose, because you can only squeeze so much efficiency from your business. I get excited about projects that drive growth and make people more successful; projects that use AI to raise all boats.

Third: Is it technically feasible and viable right now? You should be able to build it using existing tools and technologies without having to invent anything new.

Most companies come to my team with lists of 60-100 use cases. These three filters usually reduce that list to three to five projects, which is exactly what you want. This process helps you avoid “AI washing.” You don’t want to build something just because everyone else is doing AI.

From Concept to Production: Microsoft’s Three-Project, Three-Month Model

Q: Many companies struggle with the transition from AI concepts to scaled production systems. What's your methodology for de-risking this process?

EK: We have a very specific progression: three projects in three months, following a clearly defined process.

First, we encourage companies to do “hackathons,” which get people excited and generate ideas. Then we apply our three filters to narrow those ideas down to a focused project list.

Next comes proof of concept (POC) to demonstrate that these specific ideas could work with that company’s data and systems. Maybe you're not sure this part can connect with that piece of data, or you need to verify that one component will work with your existing systems. Now’s the time to work it out.

Then you build a minimum viable product (MVP). This is where organizations figure out how it's going to work in production, what it's going to cost and who's going to manage it. Once you catch the tiger, who’s going to feed it and clean the cage?

When you've followed this process successfully with your first few projects, you can scale it to dozens of AI initiatives. Your organization understands how to adopt AI, knows the process and has built the internal capability to execute plans systematically.

Q: A phased approach prioritizes organizational learning as much as technical delivery. How essential is education and mindset to AI success?

EK: AI adoption is as much psychology as it is technology. You're dealing with something very new that makes people uncomfortable, and you need to prove the process works.

But you also need to understand that you're going to fail often with AI, and senior stakeholders can find it hard to learn that risk and failure are part of the process. Traditional business thinking says, "be right a lot," but AI involves uncharted waters.

The key is learning how to label "failures" as experimentation. This is more important in AI than any other development I've seen in my career. Every setback should teach you something that you can apply to your future AI projects.

Technical and Cost Considerations for Your AI Adoption Framework

Q: What should business leaders understand about the technical challenges of deploying AI projects?

EK: The biggest misconception is that AI involves throwing data at a large language model (LLM) and getting magic back. But no AI project uses only LLMs or generative AI. Every single project uses a lot of data, querying, standard databases, multiple data sources and old-school code.

Every project that goes into production uses, on average, 6-14 different models. You might use translation services, computer vision, document recognition, speech-to-text, data vectorization and/or various data science models.

What's more, most enterprise AI systems work as multi-agent architectures; different AI components handling specialized tasks that need to work together seamlessly.

Let me give you an example. Say you want an AI tool that helps calculate taxes. You need one agent that validates whether the user's question is appropriate. You need another that checks if the user provided all necessary documents. You need sub-agents that handle different aspects: one for property taxes, another for different state requirements and so on.

Then you need a coordination layer that manages all these specialized agents, and finally a quality assurance component that evaluates whether the answer meets your standards based on the last 10,000 responses you've provided.

This is why we do POCs and MVPs: to understand all these moving parts before committing to full production.

Q: Cost management is a critical but often underestimated aspect of AI strategy. How can companies balance technical ambition with financial sustainability?

EK: AI can get very expensive: not because the units are costly, but because consumption can explode quickly. You need to make the solution viable for the long term. Is the juice worth the squeeze?

Sometimes what seemed exciting during the POC phase becomes a budget nightmare six months later.

You'll probably end up using multiple types of databases and balancing smaller and larger models to keep costs manageable and make the solution viable for the long term. The key is building something you can not only run and operate but also finance sustainably.

And remember that these systems will evolve over time—you can swap models as newer, better ones become available.

Clarity of Purpose Is a Deciding Factor in AI’s Success

Q: What's the one key thing that determines whether an AI project succeeds or fails?

EK: Clarity of purpose. If you're building something because it's really cool or because everyone else is doing AI, it’s going to fail.

I can't tell you how many projects I've worked on where I ask, "what are you doing this for?" and the company responds, "because it's critical for us." That's not an answer. It's not specific, meaningful or useful for project success; the company hasn't done the due diligence to understand the real purpose.

Don't get caught up in the technology or the hype. Focus on finding those two or three projects that will really move the needle. And approach it systematically, being prepared to experiment and learn from failures.

The companies succeeding with AI are the ones that treat it as a strategic capability to build over time, not a magic solution to implement once and forget about.

The Importance of Partnership in Successful AI Programs

The insights Eduardo shared underline the importance of structure when shaping an AI framework for business. Purpose, feasibility and long-term responsibility form the foundation of any strong AI adoption framework.

Here at Aptean, we’ve embodied this framework by experimenting, failing fast, but learning faster to build AppCentral, our innovative AI-powered connected workspace. AppCentral gives you a single environment where your core applications, data and AI capabilities work side by side, automating business processes and streamlining day-to-day decision-making.

These insights were originally featured in The Algorithm, Aptean’s digital magazine for becoming an AI-first company. Download the latest issue of The Algorithm now.