“AI research is getting faster” startled me recently
Last Wednesday night, I was revising a client workflow in a shared office in Hangzhou. The testing kept failing. After staring at the screen for two hours, I suddenly realized: if AI can accelerate research, the time we spend every day on repetitive searches, reading materials, and drafting could get repriced too. My first reaction wasn’t excitement, but unease: could the execution skills I’ve built become less valuable?
It doesn’t do the work for us—it adds another layer to the research process
The core idea is simple: AI doesn’t only help people use AI; it also participates in researching AI itself. Researchers can have models help gather sources, write code, and compare approaches, then have a person check the results. My colleague Xiaolin, who worked with a knowledge-base startup in Shenzhen in April 2025, used AI to screen interview transcripts before manually checking the original quotes. A three-day sorting job took half a day. His biggest takeaway wasn’t “fully automated,” but moving his time from moving information to judging it.
To replicate this today, we don’t need a research system
The lowest cost is roughly RMB 0–100 per month using an existing chat tool. I’d budget two hours to learn how to write effective prompts and verify sources. The technical barrier is knowing how to copy, paste, and check materials; programming knowledge isn’t required. First step: open the AI chat tool you already use, start a blank conversation, enter “Compare two options for me. List the evidence and the uncertainties,” then test it with the most tedious task on your desk.
Based on where my work is now, this is how I’d arrange it
When I’m just getting started, I’d use it to organize requirements and build a competitor list without rushing into a complicated workflow. The experiment is cheap, and mistakes are easy to reverse. With one or two clients, I’d hand repetitive research and document sorting to it while keeping client communication and final recommendations for myself. While scaling, I’d record every point where I manually revise the output, gradually turning those patterns into a reusable team checklist. Only then would a more stable team solution be worth considering.
Not everyone needs this tool, and there’s no urgency to try it now. I might start with one task that isn’t urgent and can be undone if it goes wrong, spending 30 minutes on it. The point isn’t to let AI make the decision for us. It’s to see which stretch of repetitive labor it can take off our hands.