How I Outsourced My Google Research to AI (And You Can Too)
GenAI 30 Project Challenge - 5
GenAI 30 Project Challenge - 5
Published: April 4, 2025 URL: https://buildtolaunch.ai/p/ai-research-automation-workflow Engagement: 42 likes, 16 comments, 2 restacks Word count: 1027
When you're buying a car, how many hours do you spend scrolling through reviews, comparing features, or deciding on a dealership?
Or when planning a family trip to a new city, how long do you spend googling cool places, kid-friendly activities, and food spots?
For me, the answer is always: hours. And I was totally used to that — until I ran into a problem I couldn't afford to spend hours on.
The Problem
I had a task: research a large list of preclinical drugs.
That means digging through Google, checking clinical trials, reading SEC filings, and pulling together anything that can help build a clear, professional picture for clients.
In short, I needed to answer questions like:
- What are the available technical details?
- Which companies are developing the drug?
- What clinical trials have been conducted?
- How many patients have been enrolled or treated?
- What's the market value — now and projected?
Just researching one drug could take hours. It's tedious, overwhelming, and very easy to get lost in a sea of information.
So, I decided to use this GenAI project as an excuse to streamline everything.
Did I know how at first? Nope. Did that stop me? Also nope. Cursor had been impressing me lately, so I thought: why not just explore it together?
First Try — Shockingly Good
I picked a drug from my list, threw it into Cursor with the five big questions above, and hit enter.
Here's what happened: Cursor just… did its thing.
It was fascinating to watch it pull info from across the internet. In just a few minutes, I had a nice chunk of raw data.
Of course, it looked a bit meh. Not client-ready.
So I asked Cursor: "Can you organize this into dedicated markdown files?"
Then, just for exploring different storage formats, I told it to put the info into Excel format.
Mind blown. Total time: under 10 minutes.
The Next Step: System Prompting
Sure, typing in five questions and waiting a few minutes is fast. But when you're dealing with dozens of drugs, even that becomes a chore.
So I worked with Cursor to create a repeatable system, divided into 3 phases:
- Search online
- Compile into markdown files
- Extract and format into Excel
I wrapped it all into a prompt. Now, I just plug in the drug name and let it run.
A few hiccups:
- I still had to feed in drug names manually
- Cursor stops every 25 steps and waits for me to type "continue"
- Sometimes the formatting was inconsistent
At that point, I realized: maybe it's time to bring in a proper AI agent.
Batch Processing? Let's Go!
Instead of manually poking Cursor, I rolled up my sleeves and built a crude AI agent to automate the whole thing.
Here's what I wanted:
- Automated data collection via search engine APIs
- Smart detection of missing info and supplementary searches
- Data synthesis + spreadsheet generation
I broke it into two agents:
Planner Agent:
- Designs the research plan
- Detects info gaps
- Verifies sources
- Coordinates tasks
Executor Agent:
- Fetches data (market value, clinical trials, mechanism, etc.)
- Extracts and attributes info
- Keeps data clean and organized
- Documents the process
The test run? Surprisingly okay! Data came in clean, Excel files generated, and structure followed. Until…
API rate limits. Apparently, I made way too many calls and got rate-limited.
Tried again the next day — still blocked.
So I had to explore every free-tier API I could find: DuckDuckGo, Google, Bing, Serpstack. They all worked… briefly. Then I hit the quota walls again.
The Hybrid Strategy
I looked at the data I had and realized something funny:
All my results were pointing to just two sites — clinicaltrials.gov and sec.gov.
So why use search engines at all?
I pivoted: instead of asking Google where to find data, I went directly to the source.
Along the way, I found a few more go-to sites like PubMed and NCBI.
Now my wild AI orchestra has become a clean, reliable system:
- Scrape and store data from high-quality sources (via API or scraping)
- Send the raw content + my key questions to an LLM
- Parse the results into an Excel sheet, with source references
And the results look reliably good!
Lessons Learned
I can't believe I went through all of this in just a few days. Pre-AI me would've spent weeks figuring this out manually.
- How to ask questions so AI gives better answers
- When to use the right tools
- How to document findings as reference points
- Why separating roles (planner vs executor) makes things scalable
- How to recover after hitting limits and pivot smartly
My coworkers were impressed by the very first version.
AI is not just a productivity booster — it's a new way of working, thinking, and building.
Final Thoughts
It's a little sad that my complex agent orchestration system didn't become the final solution.
But hey, who says I can't use that system to plan a trip? Or find a car dealership? With full references, reviews, and a beautifully formatted spreadsheet?
Have you built your own AI workflows like this? I'd love to hear how you're using agents in your own life.