I’ve been busy lately with graduation from my MBA program and moving. Amid all these transitions, I’ve found it helpful to pause, ground myself, and reflect: What did I actually learn during my two years in business school?
As we all know, pursuing an MBA in the U.S. costs an arm and a leg. I might as well make sure the lessons stick by writing down the highlights. I’m organizing my MBA takeaways into these broad themes: AI & Technology, Strategy and Operations, Capital Allocation, Leadership & Management, and Core. I’m still figuring out exactly where this series will lead, and that is part of the fun. Let’s see how the journey takes shape.
I’ll start with AI fluency (AI & Technology) because I’m pursuing this space after my MBA. The field is still nascent, and the world is learning in real time, myself included.
I was fortunate to spend my MBA internship at an enterprise AI startup, where hands-on experience complemented what I learned in the classroom. For this section, I’ll draw from two favorite electives: Business Analytics II: Foundations of AI and Generative AI for Business (Side note: I wish I had taken BAII earlier so I could’ve taken Business Analytics III, but it’s all good).
Three key takeaways, AI fluency edition:
- AI is grounded on data and probability. LLMs learn patterns from vast amounts of data and use those patterns to predict what comes next.
- AI is a production system. Having the most sophisticated LLM alone does not guarantee success. The plumbing around just as important as the model itself.
- Competitive advantage is beyond the technology in itself. The moat is around the model, which often comes from distribution and execution.
1. AI is grounded on data and probability; key terms
At its core, AI learns patterns from data and uses those patterns to predict what is most likely to come next. Large language models (LLMs) apply this principle to language: they are primarily trained to predict the next token based on the text that comes before it.
One new thing about these models is that we can use it using simple English (natural language); a progress from relational databases and tables that can only be accessed through SQL (structured query language) and spreadsheet apps (Microsoft Excel, GSheets).
Think of a child shown many pictures of cats and dogs. Over time, the child notices that cats often have pointier ears and dogs often have longer noses. When shown a new picture, the child makes an educated guess: “This is probably a cat.”
Apart from the actual mechanics, I also learned a couple of key terms:
Supervised vs. unsupervised learning describes how AI learns from data.
- Supervised learning learns from labeled examples. For instance, a model studies customers labeled “churned” or “retained” to predict which current customers may leave.
- Unsupervised learning works with unlabeled data to uncover patterns or natural groupings, such as customer segments with similar behaviors.
Discriminative vs. generative AI describes what AI does with what it learns.
- Discriminative AI primarily classifies data or predicts outcomes. Examples include spam filtering, facial recognition, credit scoring, recommendation engines, and churn prediction.
- Generative AI learns the underlying patterns in its training data and uses them to generate new content, including text, images, audio, video, and code. Examples include writing, summarization, coding, and synthetic-data generation.
Overall, I probably should have paid more attention in my statistics and math classes. Kidding aside, AI ultimately boils down to logic, and uses that logic to create. For machines, creativity emerges from pattern recognition. For humans, creativity often seems to exist in tension with logic. Usually the creatives in our society are the most intuitive and emotional. Is machine-generated creativity the same as human creativity? I digress. My broader takeaway is this: data is the lifeblood of this new technological wave.
2. AI is a production system
AI is a production system, not merely an LLM, which makes systems-thinking essential for anyone pursuing an AI-forward career. Even the most powerful model cannot compensate for poor data, weak evaluation, or a broken workflow.
I learned in Business Analytics II that a model can perform well during training or cross-validation yet fail in production. Overfitting and biased data are the usual culprits. Generative AI for Business extended this thinking to deploying any technology across an organization. An LLM becomes useful only when it is given the right context, integrated into actual workflows and connected to relevant tools and information.
Many AI projects fail outside the model because of:
- Inputs: Fragmented, biased, or unrepresentative data
- Process design: Weak evaluation, poorly designed workflows, or inadequate system integration
- Post-launch: Unclear ownership, low user adoption, and missing feedback loops or ongoing monitoring
Applying AI in real-world scenarios requires human-built plumbing: well-written prompts, well-designed workflows and adequate guardrails. Their value depends not only on the model’s intelligence but also on how reliably the surrounding system turns its outputs into useful actions.
Human-in-the-loop remains essential in setting this whole production system. Humans set the direction, review AI-generated decisions, handle exceptions that the system cannot, and apply the creativity and judgment that technology alone cannot replicate.
3. Enterprise AI advantage often comes from distribution
The last lecture of Business Analytics II was remarkable for me. It challenged a common assumption that having the best technology, or even the best product, is enough for world domination.
Tools like Codex, Claude, and Cursor have made building AI products faster and easier. Anyone ambitious and resourceful enough can vibe-code their way to a working product. But when building becomes easier, copying does too. Technical features can be reproduced quickly, particularly in crowded enterprise AI markets.
Building a great product is one challenge; nailing the go-to-market motion is an entirely different animal. Making people use and continue using your product goes beyond technical capability. You’re dealing with behavioral psychology and competing with numerous products out there.
Distribution is competitive advantage. Products become even more defensible when they are deeply embedded in customer workflows, creating stickiness and high switching costs. The moat is generally not “AI” itself. I plan to discuss further what I’ve learned on competitive advantage in my next post.
To conclude, I see AI as today’s Internet in the early 90s. It’s new, it’s sparkly, it’s cool. AI is a paradigm shift, yet still relies on the fundamentals of capitalism. Value is created when one know what it is exactly, how to apply it in real life and someone is willing to pay for it.