15 real AI agent use cases in research, each with its source and a prompt to try it yourself.
About 15k pages of discovery stood between a solo practice and a summary judgment motion. Claude ran OCR, built embeddings and search scripts, then found evidence for each assertion, with traces to check.
Result Trusted because it left traces to follow
Tracking which newsletters break into Substack's top 200 meant clicking and copy-pasting. Claude in Chrome now scrapes the charts every Tuesday, and Cowork turns the data into a spreadsheet report.
Result Each scrape now takes a couple of minutes
When a new OpenClaw beta added Telegram message streaming, they asked their agent what else had changed. It read 15+ merged pull requests, compared them to the current version, and gave an upgrade recommendation.
Result Costs about $10-15 a day, with spikes to $30+
Before a conference with 40 people they had never met, LinkedIn research normally takes an evening. Claude Code researched the attendees in parallel via LinkedIn and Chrome MCP, producing a full briefing in 5 minutes.
Result Also books flights and hotels in the company travel system
A master's thesis in a fast-moving field and too many new papers. A weekly Codex automation searches arXiv, triages only new papers and files promising ones into Zotero, with fewer hallucinated citations.
Result Sorts papers into keep, skim, drop or duplicate
Financial data for 100 schools, from tax filings into a Google Sheet. Operator clicked through it in the background, but rounded numbers, sometimes grabbed the wrong school, and was so slow it was stopped at 25.
Result Then o1 turned the sheet into an interactive chart
Which zip codes, with a hundred-mile radius each, cover most of the 50 states? DeepSeek R1 on Perplexity wrote the Python, fixed it from pasted tracebacks, and drew the coverage map.
Result Plus a script that maps the coverage
Treated as a child for a rare condition, they asked Gemini about it years later. It explained the test used back then was the wrong type. Two specialists confirmed it. Their advice: verify, don't trust blindly.
Result Two specialists confirmed the old mistake
Ten years ago a similar scraper took them a horrendous amount of time and wasn't half as good. With free Claude and ChatGPT, this one grabs images, reads them with OCR and crops the hits.
Result About 5-10 hours over a weekend
Audio circuits had always intimidated them. They took apart a stereo effects pedal with ChatGPT, quizzing it on every part, then designed their own version in KiCad. A college-level crash course in a week.
Result Grilled it on every node, at all hours
Describe a research question and OpenClaw searches Semantic Scholar and Google Scholar, builds a clean Zotero collection and tries 9 ways to fetch each PDF. It gets about 60-65% automatically.
Result Also runs their to-dos, briefings and meal plans
Surveys, supplier risk, emissions reports, modern slavery checks: the data gathering never ends. With Claude Code they built tools that now do about 70% of it. Plus a game of Frogger, for breaks.
Result Plus a game of Frogger, for breaks
A photo of the dead-start switch panel from a historic supercomputer. Claude Opus read the switches and disassembled them into commands in about two minutes. Not detailed, but impressive.
Result Not detailed, but impressive for a historic panel
Bessemer Venture Partners' investment and legal teams lived in manual research. Perplexity Enterprise now handles it and tracks news on 400+ portfolio companies, per Perplexity.
The data team wanted to automate complex analytics and lean less on outside agencies. Claude now runs those workflows, and also helps HR draft feedback and marketing produce content.