The Rise of the 'AI Intern': How We Built a Multi-Agent Strategy Engine Without a Team of Software Engineers
Moving from data gathering to strategic thinking in Pharma Portfolio Management using Dataiku, Gemini 3.0, and Agentic workflows.
In the high-stakes world of Pharmaceutical Portfolio Management, our currency is intelligence. Every quarter, we produce a Pharma Market Overview—a critical document summarizing financial performance, M&A activity, and R&D trends. Senior management relies on this to make billion-dollar decisions.
But there was a problem.
Historically, producing this report was a manual bottleneck. Highly skilled Portfolio Managers were spending countless hours acting as “human scrapers”—combing through web search results, reading financial statements, and copy-pasting data. We were spending 80% of our time gathering information and only 20% analyzing it.
We didn’t need more people. We needed better agents.
Here is how we redesigned this workflow using Applied Agentic AI, built entirely by non-engineers, and why the “build vs. buy” calculation in enterprise AI is shifting rapidly.
The “Why Now?” Moment
The most important part of this project wasn’t the code—it was the fact that we didn’t need a team of senior software engineers to write it. The technology has finally matured to the point where domain experts (us) can build complex tools using low-code platforms.
We utilized Dataiku to orchestrate the workflow, proving that you don’t need to be a “geek” to deploy enterprise-grade AI.
The Architecture: Scout and Analyst
We moved from a linear, human-driven process to a multi-agent system consisting of two specialized roles:
The Scout Agent: This agent continuously monitors the horizon. Instead of illegally scraping full articles (a major compliance risk), we architected a system that ingests RSS feeds. It grabs snippets and HTML source links, allowing us to identify relevant signals—M&A rumors, trial results, earnings calls—and only pay for the full articles we actually need. This solved the copyright issue while keeping the pipeline full.
The Analyst Agent: Once the data is curated, the Analyst takes over. It synthesizes the disparate snippets into a cohesive narrative, drafting the sections of the report that used to take days to write.
Model Selection: The Cost vs. Capability Trade-off
One of the most interesting findings during our testing was the divergence in model economics. We benchmarked three cutting-edge models: Gemini 3.0, Opus 4.5, and GPT-5.2.
Using Portkey to measure exact token usage and costs, the results were illuminating:
- GPT-5.2 was powerful but priced at a premium—nearly 9x the cost of the alternatives.
- Gemini 3.0 and Opus 4.5 landed in the “Goldilocks zone”—a tight race between high capability and reasonable cost.
While the total cost per run is only in the single-digit USD range (making even the expensive models affordable for a quarterly report), the efficiency of Gemini and Opus proved that you don’t always need the most expensive hammer to drive a nail.
The Human-in-the-Loop: Trust but Verify
In Pharma, accuracy isn’t a “nice to have”—it’s a requirement. We cannot tolerate hallucinations.
We designed the system so the AI is the drafter, not the publisher. The workflow generates a draft, complete with source citations. A human Portfolio Manager then acts as the Editor-in-Chief. They review the narrative, cross-check the sources (which are always linked), and validate the strategic conclusions.
If a number looks “off,” the manager clicks the source link to verify. If the narrative lacks nuance, the manager adds it. Only when the human is satisfied do they click the button that packages the report into an HTML email and PDF for senior management.
The ROI: From 80% Grunt Work to 100% Strategy
The impact was immediate. We reduced the drafting time by approximately 80%.
But the real return on investment isn’t just time saved; it’s the shift in mindset. We are no longer data gatherers. We are architects of intelligence, supported by a team of AI agents that work 24/7.
This project proved that with the right low-code tools and a clear understanding of the workflow, “Agentic AI” isn’t just a buzzword for tech giants. It’s a practical reality for any department willing to reimagine how they work.