How I saved two days of drudgery with ChatGPT

I’m a writer who uses AI to help. I’d never let AI write for me (I’m too proud, it writes boring prose, and it makes mistakes). But there are still times when a tool like ChatGPT can do things that would be much harder any other way — especially on rote, tedious research tasks.

Turning a brainstorm into a database

For a book project I’m working on, my client and I will be reaching out to and interviewing dozens of startup companies. As with most case study research projects, those companies must have certain qualities:

• They were started relatively recently.
• We know the principals (or know someone who can introduce us).
• They are good examples of what we’re trying to demonstrate in one of the chapters we have planned.
• They exhibit some diversity — not just in who we interview, but in what kind of companies they are, what category they’re in, where they’re located, and so on.

To get started on this, we conducted a brainstorm meeting this week in person with my client and two of his colleagues. They threw out about 60 names and my client scrawled their names on a whiteboard, including notes about what they were doing. I also took detailed notes.

The result was a mess. We had a photo of the whiteboard and my notes, but what I needed was a sortable spreadsheet of these companies and what they did. The next obvious step was to look them all up on online and make a note of their short descriptions, categories, and markets. That was work, and not likely to be very enlightening.

So I decided to see if ChatGPT could help. Here’s the start of my dialogue with it:

I’m going to want your help with information and classification of a bunch of startup companies. I’d like the output to be a table (spreadsheet) with one row for each company and one column for the classifications I want to create. Can you do that?

ChatGPT said:

Yes, absolutely — I can help you with that.
We can create a table where each company is a row, and each classification you specify becomes a column.
Once the table is ready, I can show it to you and export it as a spreadsheet if you like.

What classifications (columns) would you like to include?

I’d like the following columns: (1) name, (2) date founded, (3) one sentence description (4) category (e.g. fintech, consumer hardware), (5) B2B or B2C, (6) state of financing (e.g. seed, series A, series B) (6) exit, if any (e.g. acquired, went public)

ChatGPT said:

Perfect — here’s the column setup based on what you specified:

| Name | Date Founded | One-Sentence Description | Category | B2B or B2C | State of Financing | Exit (if any) |

A couple quick clarifications to make sure it’s exactly what you want:

• If a company has had an exit, do you also want the acquirer or IPO date listed (if available)?
• If a company has no exit yet, should I just write “none” in that column?

Once you confirm, we can start building it!
Which companies would you like to start with? (You can give me a list, or I can suggest some.)

As a test, just after this, I gave it the names of about a dozen companies from my notes. ChatGPT searched the web, found companies that were a plausible match (same name, recently founded startup), and populated the table. I found its short descriptions of the companies helpful and fairly appropriate. It pointed out some problems with my list, and in those cases I realized that between my notes and the scrawl on the whiteboard, I’d gotten some company names wrong (it turns out there’s not startup called DataBog, it should be DataDog). I also realized it would be useful to include links to the company websites, so as I entered more companies, I told ChatGPT to add that detail to the table.

There were placeholders for much of the data in the table it showed. I asked about these, and ChatGPT told me it was still filling in that information.

But even when the full list was populated, we weren’t done. I uploaded my Microsoft Word file of notes. Then I told it to add a column for who the referrer was (that is, which of my client’s colleagues had suggested the company) and a column for my notes from the meeting. ChatGPT took on the tedious task of matching the bullets in my notes to the companies in the list, parsing out the initials of the referrer where I’d included the person’s initials, and then grabbing the rest of my notes on the company and populating the notes column with that.

Here’s how the result look when I imported it into the tool we’re using, Coda (I’ve blurred the actual content because that it not public information).

Here’s a closeup so you can see the first few columns better:

If you read my book, you know that a case study target list is essential early in any nonfiction book project. Now, after a one-hour meeting and a couple of hours with ChatGPT working mostly in the background, I had such a list, neatly sorted with information on each company. I could continue to populate it with which chapter the case studies would be slotted into, contact information for targets, and state of interview inquiries. I could sort it by industry or market or founding date. This was a huge step forward. If I’d done the same work to create the little database manually, at the rate of about 15 minutes of compilation and web research time per company for 60 companies, it would have taken two days. And that’s not two days of productive, creative time, either — I had avoided two days of drudgery.

Why this ChatGPT was perfect for this task

Right now, people are using AI to do everything. But it’s not good for everything. Here’s why this task was well-suited to AI-based automation:

• Much of the necessary information was online, and the rest was in my notes. There was no need to search inaccessible or copyrighted content.
• The nature of the task — assembling information from widely spread online sources and organizing it into a neat table — is exactly what AI-based automation is most helpful at. It did what I could have done myself, but much more quickly.
• I could modify the task as I went along without causing a problem. Things like adding the web site links didn’t faze ChatGPT. It’s a lot easier to do something like this with incremental improvements than to plan exactly what you need up front and possibly get it wrong. (If that sounds Agile to you, you’re getting the idea.)
• Some imperfect or inaccurate information would not cause a problem. As you read my account, you may have asked yourself, “What if it made a mistake? What if it looked up the wrong company with a similar name, inserted an inaccurate description, or screwed up in its judgment of whether company was B2B or B2C?” These types of errors are indeed possible (and anybody I hired to do the same work could make the same errors). But when we identify a company as a promising candidate and start to reach out to target it, I’ll research it in more detail. The descriptions in the table are just internal guidelines for the book team; they won’t be published anywhere, and they’re easy to correct.

This task has turned me into a true believer. LLMs are still not going to write nonfiction prose for me, because I still think I write better and more interesting text than AI can. But I stand ready to turn AI loose on tedious tasks like this. Doing so will allow me to spend more time on creative work. That’s what I’m best at and enjoy the most.