AI Agent for Data Analysis

Every Friday I used to spend two hours building the same report. Traffic numbers from Google Analytics. Conversion rates from our dashboard. Competitor pricing from their websites. Revenue figures from Stripe. All of it manually copied into a Google Sheet, formatted, and turned into something my team could read at Monday's meeting.

Two hours. Every week. For a report that took the team ten minutes to skim.

The agent does it now. Every Friday at 3 PM, it collects the data, compiles the report, and drops it in Notion. I review it in five minutes, add any commentary, and it is ready for Monday. Two hours became five minutes.

Data collection is the boring part

The actual analysis - finding patterns, drawing conclusions - that is the interesting part. The part before it is pure tedium. Logging into five different dashboards. Exporting CSVs. Copy-pasting numbers into a spreadsheet. Making sure the date ranges match. Formatting cells.

This is exactly the kind of work AI agents are built for. Repetitive, multi-step, involves visiting multiple websites and collecting specific information. The agent uses browser control to visit each data source, reads the numbers, and compiles everything into a single document.

Competitive pricing analysis

One of my favorite recurring tasks. Every Monday the agent visits fifteen competitor pricing pages. It records every plan, every price point, every feature mentioned. Then it compares to last week.

Most weeks nothing changes. The report says "no pricing changes detected across all fifteen competitors." Fine. Peace of mind in thirty seconds.

But twice in the last four months a competitor changed their pricing. Both times I knew about it within 24 hours. One was a significant price drop that we needed to respond to. Without the monitoring, I might not have noticed for weeks.

The agent handles JavaScript-rendered pricing pages, cookie consent popups, and sites that require scrolling to see all plans. Real browser, not just HTML scraping.

Market data aggregation

Last month I needed salary data for software engineers across European cities. Normally this means visiting Glassdoor, Levels.fyi, LinkedIn Salary, national statistics sites, and half a dozen blog posts with survey results. Then reconciling all the different numbers.

I told the agent: "Collect software engineer salary data for Berlin, Amsterdam, London, Paris, and Zurich. Check at least 5 sources per city. Present median, range, and source for each."

Forty minutes later: a clean table in Notion with all the data, source links, and a note about where sources disagreed significantly. I would have spent half a day on this. The agent did it while I was on a call.

Pattern detection in collected data

Once the data is collected, the language model can actually analyze it. This surprised me - I expected the agent to be good at collecting and bad at interpreting. Turns out Claude is genuinely useful at spotting patterns.

When I feed it four months of competitor pricing data, it notices things like "Company X has raised prices twice in Q4 while Company Y dropped their starter tier by 20%. This suggests X is moving upmarket while Y is chasing volume." That kind of analysis is not revolutionary, but getting it automatically as part of a weekly report saves me the mental effort of doing it myself.

Limitations

• - No direct database access. The agent collects data from websites and APIs it can reach through the browser. It cannot query your PostgreSQL database or pull from internal dashboards behind VPN.
• - No chart generation. Reports are text and tables. For visual charts, export the agent's data to Google Sheets or a BI tool.
• - Statistical limits. The agent can calculate averages, find trends, and compare numbers. It cannot run regression analysis or build statistical models. For heavy quant work, use Python or R.
• - Login-required dashboards. The agent can access public websites and pages where you have provided credentials. But feeding it passwords to analytics platforms is a security decision you should think carefully about.

Setup

Skills: browser control (visits data sources) and web search (finds sources). Notion MCP for saving reports.

Key instruction pattern: "Every [frequency], collect [specific data] from [specific sources]. Compare to previous period. Note any significant changes. Save report to Notion and send summary to Telegram."

The more specific you are about what data and where, the better the reports. "Check our competitor pricing" is vague. "Visit pricing pages of [Company A], [Company B], [Company C] and record all plan names, prices, and listed features" produces reliable, consistent data every time.

Frequently asked questions