After a Month with WorkBuddy: Why It Shouldn't Compete Head-On with Research Report Agents

Recently, I came across an article about AI automatically generating financial research reports. The architecture was quite complete. Dify handles the frontend, FastAPI the backend. When a user inputs "Please write a CATL (Contemporary Amperex Technology) research report," the system breaks it down into subtasks, calls data interfaces like Akshare and East Money, retrieves from vector databases, then uses a polishing Agent to generate charts, finally aggregating into HTML, Markdown, or PDF. It looks like a beautiful investment research agent pipeline.

But after using WorkBuddy for about a month and recently connecting sources like Huibo Investment Research Info, knowledge bases, PDFs, and Excel, my judgment has become more conservative. The real difficulty in financial research reports lies in whether the sources, dates, publication status, and data standards can be verified after generation. I am now more inclined to treat WorkBuddy as an auditable structured workbench rather than a machine that automatically writes research reports.

The architecture in that article essentially solves how to break a research report task into multiple agents. Requirement parsing, data querying, code execution, knowledge base retrieval, content polishing, chart generation, and report aggregation each have corresponding modules. This approach certainly has promise. Renmin University's Gaoling School FinSight award-winning system is similar, abstracting data, tools, and agents into programmable variables, using code to drive variable memory with high autonomy.

Doing financial analysis, especially after six years in foreign firms, I'm increasingly worried about one scenario: the system outputs a research report with beautiful formatting, smooth logic, and complete charts, but it mixes in old news from two years ago, or confuses publication dates, data cutoff dates, and institutional viewpoint sources.

I wrote a post about WorkBuddy's verification pipeline at the end of August. My core view then was to lock down sources, dates, and status; no matter how fancy the prompt, it can't fix source and date errors. Looking at that agent-based research report architecture now, I still hold this judgment.

WorkBuddy is actually smoother in this regard. It feels more like an office tool; it doesn't let me freely write Python to call financial data APIs, but it's very familiar with tasks involving office files, spreadsheets, PDFs, emails, and PPTs. My usage is to have it do field extraction and verification tables first, leaving the actual report writing to humans.

I give it a fixed template with fields including report name, issuing institution, publication date, data cutoff date, source file, core judgment, key figures, data standards, items pending verification, and whether it can enter the morning call. This template looks clumsy but is crucial. It forces WorkBuddy not to cleverly summarize, but to first break materials into traceable structured fields.

I started using Huibo Investment Research Info recently, less than a week ago. I've used the knowledge base for two weeks, PDFs for about four weeks, and Excel and WorkBuddy for about a month. These times aren't long, but linking these tools together over the past week has indeed stabilized my morning meeting preparation process.

Last Wednesday morning, I needed to prepare for an 8:30 AM morning call covering strategy monthly reports, US Treasury volatility, Beijing Stock Exchange interim reports, and AI computing power. Previously, my method was to open Huibo, download several PDFs, use Excel to cross-reference client holdings and industry views, and finally hand-write minutes. It took about 90 minutes.

This time, I tried using WorkBuddy. Step one, I created a "September Morning Meeting Verification" project folder in WorkBuddy, putting in several PDFs exported from Huibo, a client holdings Excel file, and two notes from the knowledge base. Result: it crashed the first time.

WorkBuddy could indeed organize the content, but with mixed data sources, it easily confused fields. For example, one report cited recent data in the body but had current-period strategy in the title; it almost wrote the data cutoff date as the publication date. Another scanned PDF table had intermittent field extraction, and the final exported Excel date column formats were inconsistent. I wrote about this pitfall before: when WorkBuddy handles mixed data sources, it's best to break down steps first, don't ask it to generate the morning meeting report directly in one go.

Later, I changed to a three-step process. Step one: only make metadata cards, outputting file name, source, institution, publication date, data cutoff date, whether it's old news, and whether it's citable for each file. This step ignores opinions, looking only at factual sources.

Step two: extract viewpoints and numbers. Generate 5 to 8 fields per report, including core judgment, key data, supporting logic, risk points, and changes from the previous period. I quickly checked this; the output was more stable than directly writing summaries because the task boundary was clearer.

Step three: assembly. I fed the structured results from step two back into WorkBuddy, asking it to output industry, judgment, data, risks, and items to verify according to my morning meeting template. Finally, it exported a Word version, and I manually edited it for about ten minutes.

Running this process, my tests showed original reading time compressed from about 3 hours to a 50-minute draft, plus 15 minutes of manual verification. Morning call prep time dropped from 90 minutes to about 56 minutes, saving roughly 37.5%. This number isn't rigorous—it's my own estimate—but the process is definitely more stable. More importantly, I no longer rely on AI's memory, but on the checkable fields it generates.

This is why I think WorkBuddy suits the daily routine of financial analysts but isn't suitable for directly replacing research report generation agents. Its value is turning fragmented materials into a traceable, verifiable, reusable office workflow. Don't expect it to generate a CATL research report in one breath.

To be blunt, WorkBuddy's shortcomings are also clear.

Unlike the Dify + FastAPI architecture in that article, which can split financial data queries, Python code execution, vector database retrieval, chart generation, and report aggregation into multiple agents connected via interfaces, WorkBuddy can't do that. Nor is it like FinSight, abstracting data, tools, and agents into programmable variables. WorkBuddy leans towards office scenarios, strong in documents, spreadsheets, PPTs, emails, and file organization. It can't handle infrastructure like real-time quotes, financial data interfaces, quantitative backtesting, or regulatory disclosures.

I tried having it process multiple PDF tables from Huibo reports. If the tables are complexly nested or scan quality is average, field extraction drifts. It can output charts, but they look more like office presentation graphics than investment-grade data charts. Asking it to automatically scrape Akshare, calculate ROIC, or generate DCF sensitivity analysis is not in its wheelhouse.

So my judgment is clear: WorkBuddy suits frontline analysts, sales, finance, and product staff who need to organize information and produce materials, but it's not suitable for replacing brokerage research systems. It's more like an assistant in the application phase of office scenarios, reducing data processing workload, but complex cross-stage decisions, real-time data links, and compliance audits still require professional infrastructure and manual verification.

What should be amplified about WorkBuddy is its structured verification capability. If teams treat it as a one-click research report machine, they'll likely be disappointed. If treated as a tool for material decomposition, field normalization, and process logging, the ROI is actually quite good.

My current habit is that any content entering a morning call or client briefing must first pass through WorkBuddy's structured template. Old news, date mismatches, and unclear standards are easily overlooked during manual reading, but once turned into fields, errors are exposed. Previously, I thought mobile AI memory functions were important, but later I valued task execution rate more. Now, with WorkBuddy, I don't care if it remembers my preferred format; I only care if it can turn every step's output into checkable fields.