AI in Practice : Why Building It Yourself Is Harder Than It Looks

Abhay Jajoo, founder and CEO of Customer Insights.ai, joins Telmo Silva for a wide-ranging conversation about data complexity in life sciences, the architecture behind intelligent agentic systems, and the real cost of the build-versus-buy debate in the age of AI. With 40 years in pharma analytics, Abhay brings a grounded, no-nonsense perspective on what it actually takes to turn messy industry data into decisions that move the needle, and why getting the architecture right is the only bet that survives every technology wave.

Questions We Asked

• Why is pharma data so much more complex than data in any other industry, and has regulation really changed anything over the years?
• What does it mean to build an AI system that is truly LLM agnostic, and why does it matter when token costs can vary 100x between providers?
• Can companies really build their own agentic analytics systems, or is the “build it ourselves” wave about to hit a wall?

What We Learned

• Complexity in pharma data is not about volume. It is about variety: different sources, different structures, different rules for every drug, every market, every disease area.
• Architecture reigns supreme. Being data agnostic, infrastructure agnostic and LLM agnostic is not a constraint, it is the only sustainable design choice in a market where providers change faster than contracts.
• The “build ourselves” wave is real but fragile. Boards are pouring AI budgets into IT teams that are great at maintaining systems but not at solving business problems. Most projects run out of money before they have anything to show.
• Hypervertical AI is the next frontier. The third wave of AI innovation will not come from Nvidia or OpenAI. It will come from narrow, industry-specific companies that encode decades of domain knowledge into deployable, maintainable products.

Forty Years in Pharma, One Core Problem

Abhay Jajoo spent the first part of his career at IMS Health, one of the world’s largest pharmaceutical data companies, where he and Telmo Silva both worked. That shared background is the foundation of the episode. The conversation opens with a detailed breakdown of how the pharma commercial ecosystem actually works: providers, payers and life sciences companies, three distinct segments with different data, different incentives and different analytical needs. Telmo draws the contrast sharply: selling a can of Coke gives you a complete, traceable data trail. Selling a drug gives you abstraction at every step, a doctor who can be overridden by a pharmacist who can be overridden by an insurer.

To me, the complexity comes not from volume. It comes from variety.
Abhay Jajoo, Founder and CEO, Customer Insights.ai

That insight shapes everything Abhay has built. His first product, Parthenon, was designed around a simple observation: prescription data for a cholesterol drug and prescription data for a diabetes drug look structurally identical. Standardize the ingestion, standardize the transformation, build the visualizations once, and you can serve any pharma company in minutes instead of months.

From Parthenon to Athena: When Agents Replace Workflows

The shift from Parthenon to Athena is the architectural pivot at the heart of the episode. Where Parthenon hardcoded workflows and visualizations, Athena replaces them with intelligent agents, each one understanding its environment, knowing what data it needs, and operating inside an orchestration layer that ties them together. The user interface collapses to a conversation. Training costs drop. Adoption accelerates. And the 40 applications built on Parthenon are now being re-released as use cases inside Athena, one by one.

The key architectural decision is agnosticism: agnostic to the database, to the cloud provider, to the LLM. Abhay explains why with a single data point. Anthropic charges $56 per million tokens. Open source models charge 56 cents. A business that bets its cost structure on a single LLM provider is one pricing decision away from a broken model. The IP at Customer Insights.ai is not the tooling. It is the pharma metadata, the analytical workflows and the standard operating procedures that tell the agents what to do when a question arrives.

The Build Trap

Both Telmo and Abhay are watching the same wave hit their respective markets. Boards are allocating AI budgets. CIOs are receiving the money. Six-month requirement-gathering projects begin. Token budgets evaporate on experimentation. And five months in, there is nothing in production.

Abhay is direct about where this leads: building is easy, maintaining is not. A bug in a prompt-engineered system is not a two-line fix. The companies that have already gone down this road in other industries are starting to come back. The happy medium, in his view, is a clear boundary between what a company should own internally and what an industry-specific provider should deliver pre-built, pre-validated and ready to deploy.

Three Waves, One Bet

The episode closes on what Abhay sees as the map of AI innovation. The first wave is infrastructure, the Nvidias and data centers fighting a CapEx battle that will shake out within a year. The second wave is horizontal platforms, the Snowflakes and CRMs gaining conversational interfaces. The third wave, still about two years out, is hypervertical: narrow companies solving specific industry problems better than any general-purpose tool ever could. That is where both Telmo and Abhay are placing their bets. And in both cases, the foundation is the same: get the architecture right before anything else.