In this second part of my discussion about ChatGPT and your small business, I want to share some small business ChatGPT tips. Also some tricks. Thus, this post provides more mechanical advice pointers.
Note: In the first part of this discussion, Using ChatGPT in a Small Business, I talked about how we’ve been using ChatGPT in our small CPA firm. That post amounted to a laundry list of tasks and projects we’ve used ChatGPT for.
ChatGPT Tip #1: Start Today
My first tip. Start today. In fact, before you read another paragraph, open your browser, enter chatgpt.com into the address box, click the Signup button, and then follow the on-screen instructions. Sign up for the $20 a month option to start.
Once you’ve done this, start using ChatGPT in place of the Google or Bing search engine. And start asking it the questions you might have in past asked a friend, a co-worker, or some online forum.
ChatGPT Tip #2: Use Detailed Prompts
You know how to use a search engine like Google. And ChatGPT sort of works the same way. You can type in a simple question.
But what you want to do with ChatGPT? Provide much more detail. My typical initial prompt probably runs 100 to 200 words. I provide background information. I explain my current understanding. If appropriate, I may cite or quote sources like a blurb from an Internal Revenue Code statute or from a Treasury Regulation. I also provide instructions for ChatGPT. I might ask for it to provide links to its source, for example. Or for a table that summarizes its analysis or its answers.
Think about the instructions you might provide to an entry-level employee. Or someone very smart but also very new to their job or role. That’s the level of detail or instruction you and I want to use.
ChatGPT Tip #3: Ask Follow-up Questions
You’re going to want to ask follow-up questions. Lots of them.
For example, maybe you realize with your first prompt and ChatGPT’s response, that you didn’t phrase things correctly. Or that you weren’t precise enough in your instructions. Ask ChatGPT for more information. Or to redo its work. Or ask for clarification about some bit of its response.
I’ve found ChatGPT so “human-like” in its response, that I’m inclined to treat it as I would treat a human. But what counts as good manners when interacting with a co-worker or professional? Like not re-asking nearly the same question ten times in a row? That’s not the right approach with something like ChatGPT. Rather with ChatGPT, you or I probably do want to ask numerous follow-up questions all based on slightly different wordings of the facts.
ChatGPT Tip #4: Watch for Errors
ChatGPT makes errors. And it’s hard to spot them because it does such an articulate job when it writes its answers. Also, most of the time? It gives you or me a good answer. I guessed in the earlier blog post ChatGPT gives the right answer maybe 80 percent of the time. And maybe that’s even too low.
But my point is, it makes mistakes. (People call these mistakes hallucinations, by the way. Because ChatGPT presents its answers or responses with seemingly unshakeable confidence.) Thus, you want to be aware of and watch for this.
By the way, often when you spot an error, you can explain to ChatGPT why you believe it made an error. And it will quickly agree. But not always.
This example from the world of tax accounting. So most small businesses can use a Section 199A deduction to avoid paying federal income taxes on the last 20 percent of their business income. However, tax law adjusts the deduction for high income small business owners if they don’t pay enough in wages or if they work in a professional service. ChatGPT probably can’t work with the adjustment formulas. (I’ve tried to get it to do this a bunch. And unless I explicitly provide the detailed steps and calculations, it stumbles again and again.)
This characteristic of ChatGPT to make errors (and of other large language model, or LLM, AIs to make errors) suggests a best practice: For really consequential chats? You and I probably want to stick to topics in which we have pretty deep knowledge.
ChatGPT Tip #5: Curate ChatGPT’s Memory
As you or I work with ChatGPT, it collects memories of business or personal facts that may matter for future discussions. Personal information such as your age, education, ethnicity, cultural background, marital status and net worth. Details about the type of business you operate including its revenues and profitability.
These memories help ChatGPT do a better job at providing you with answers and reports. But you probably want to manage and edit the information ChatGPT collects. Some information may inadvertently be incorrect. Other information may become out of date. And then some other information may be too personal to let ChatGPT use that information to infuse its answers to your questions.
You can do this curation in possibly two ways. First, you can tell ChatGPT to forget details you don’t want it to remember. (“Please ignore the information I provided about expected revenues for next year. That information was wrong.”)
Second, in the past though not currently, ChatGPT has alerted you when its memory is full. I mention this here because the feature might become available in the future. If it does, you can click a button, review its list of memories, and then remove some memories. Watch for this feature if it appears. I think it’s a good way to clean up memories that aren’t correct, up-to-date or accurate.
ChatGPT Tip #6: Use Attachments
ChatGPT lets you attach documents to your prompts. You click the plus button. It displays a dialog box that lets you point to the file. You select the Microsoft Office document, PDF or other file that should be part of the prompt.
Attachments let you provide lots of information to ChatGPT, thereby saving you from typing in a bunch of text. Also, you can include much more data when you use attachments. (ChatGPT does not at the time I’m writing this provide any firm recommendation about how much text you can enter into the text box. But it does suggest you work and it more easily works with attached, formatted documents and spreadsheets than with stuff copied or entered into the prompt box.)
And a tip here: Those JavaScript calculators I’ve been adding to our blog with ChatGPT’s help? Attachments make those possible. What I do is build a small Excel spreadsheet. Verify the spreadsheet formulas correctly make some set of calculations. Then attach the spreadsheet file to a prompt that asks ChatGPT to write JavaScript that collects the needed inputs, makes the same calculations and displays the outputs. I then copy and paste the JavaScript into a blog post.
ChatGPT Tip #7: Proceduralize Use of ChatGPT
I asked ChatGPT whether it had tips about how to begin employing an AI like ChatGPT in a small business.
Its response was lengthy, And I won’t copy and paste it here. But to generalize, it suggested creating standard procedures for using ChatGPT (for example, two-pass reviews and verification of key facts), adding explicit disclaimers to the deliverables when appropriate, recycling results when possible, and standardizing best practices.
All good ideas.
ChatGPT Tip #8: Consider p(doom)
I end with a brief discussion of p(doom), the probability that artificial intelligence creates an existential threat to humanity. I’m going to slightly redefine the term here though and focus on the possibility or probability that AI leads you or me to make some sort of catastrophic business decision or leads us down some absurdly dangerous path in our entrepreneurship or management.
Before I do that, however, let me first disclose my default bias about this sort of stuff. I’m almost always a skeptic. Seriously. To give an example that’s probably old enough to not be polarizing, I did not think there was any chance the Y2K bug would cause the power grid to fail and society to crumble. (I heard that concern from numerous, smart, well-intended friends and colleagues as the 1990s drew to a close.)
Yet I think we need to consider p(doom). Given the hallucinations an AI like ChatGPT regularly produces, given the incredibly polished, hypnotically good answers it provides, and given the supreme confidence it continually displays, we need to be careful it doesn’t send us off a cliff.
A personal example of this: Recently as part of a chat about staffing in the face of talent shortages, private equity firms muscling into the profession and then adjusting billing rates for inflation, ChatGPT recommended a tactical approach for our firm that appeared be a brilliant gamechanger.
Following its plan, we would apparently be able to massively improve our profitability. Two cups of coffee later, I was on-board.
Another cup of coffee, and I stumbled on the tactic’s terrible flaws. Boiled down to its essence, ChatGPT was suggesting we bump our prices by 50 percent more than the prices our able, smart, and highly skilled competitors charge. That tactic then would not have worked, I am pretty sure. Furthermore, the tactic irreversibly deployed very possibly would have caused our CPA firm to implode. Which counts, I think, as a p(doom) threat.
The takeaway then: Do consider p(doom) risk.