018 - Welcome Custom GPTs and ChatGPT Team

When you want a GPT that finally *gets* you... and your practice.

Hey Friends,

Let's chat. Like between you and me, not with the robot. Some exciting AI things happened this week.

But, literally, exciting AI things are happening every week, so it's difficult to find anything exciting anymore.

This week, though, it's actually extra exciting. Especially considering the implications for tax, accounting, and financial advisory firms.

Two things happened on the same day:

• OpenAI opened its 'GPT Store'

• OpenAI announced 'ChatGPT Team' Subscription

Let's talk about each and why we might care.

GPT Store

What's a Custom GPT? A custom GPT is a type of special-purpose chatbot created by someone using tools provided by OpenAI. OpenAI has even created and shared their own custom GPTs:

• 'Game Time' - A custom GPT explaining instructions for board and card games. Handy if a family member tries to introduce House Rules to your Monopoly game.

• 'Sous Chef' - A custom GPT that will craft recipes based on your likes and available ingredients. Interestingly, ChatGPT is a great recipe creator, so this GPT fits well within its skill sets.

• 'Math Mentor' - A custom GPT for parents struggling with new math, old math, or just math. (A reminder that you *can* take a picture of completed math homework and have ChatGPT Vision check over it for you).

How are custom GPTs made? If you want something other than an off-the-shelf custom GPT out of the store, you can create your own custom GPTs. ChatGPT will even helpfully fill in the information and come up with a name and a logo for you. I started this one as an example. (This is a custom GPT that you might consider building for your practice that answers tax facts. This will save you and your staff from having to pull out a Quickfinder to look up the tax brackets or who can contribute to an IRA.)

And here we are, a little further along (notice that the description will change along with the discussion) …

It's so easy to create a custom GPT; you can even accidentally create them if you didn't know that the "save" button means you're ready to publish 😬

Won't everyone just make the same GPT then?

Yes, yes, they will. The last I checked, there were 3 million GPTs in the GPT store. There were even seven 'TaxGPT's. (The best way that I check a TaxGPT's source information is to ask it trust and estate income tax questions - they usually struggle in that category.)

But this article isn't about you making a million dollars off of creating a custom GPT; it's about you supporting yourself and your firm by creating an internal custom GPT. And to do that, you'll make it relevant to your team under the 'Configure' tab.

You'll notice that all of the work that ChatGPT and I did to set up this custom GPT is reflected on this tab. This is the 'tinkering' tab - once you've got 80% of the way there, you add your 20%.

And you see that 'Upload files' button? The one with the red arrow pointing directly at it? That is where you add your own knowledge and secret sauce. Because if we know one thing about ChatGPT, it sometimes makes things up and doesn’t know everything.

But What About…

• Custom GPTs can be configured so that your data isn't used to train the model (meaning, your data isn't sent into the big robot hopper in the sky for future generations to learn tax facts from).

• Your data is still retained on OpenAI's servers under their privacy and data policy, though (please, no PII in the robot)

• The Builders of custom GPTs won't be able to see your questions, but if they have 'Actions' that move your data outside of OpenAI's servers, they *will* have your data.

  For example, here are some custom GPTs in the store:

To be a particularly useful custom GPT, you need to have not only GPT4's brains but also outside information. Otherwise, anyone with a good prompt will figure out how to replicate it (and you'll have to work around knowledge limitations). AND you wouldn’t upload your entire database of nature trails under the 'Knowledge' button. Can you imagine keeping that up to date? So, if you’re someone like AllTrails, you're going to create a little data connection under the 'Action' button. Which will send the data from this custom GPT to your servers to process.

Once again, be careful with your data, and be extra cautious with client data.

Why doesn't Microsoft have something like this?

They do! Just not the 'GPT Store' aspect. But you can create your own custom copilots and GPTs using your internal company data with Microsoft Copilot Studio. Right now, it's $200/month for 25,000 messages and going to be used by those larger companies looking for a secure, internal-data-based resource. For example, a national firm would get an efficiency boost by creating a custom Copilot populated with FAQs and resources created by their firm's subject matter experts.

If you don't want to pay $200/month, you can create a bot in Microsoft Teams using the Teams Toolkit. And before you say, "I couldn't possibly build one of these, I'm an accountant, not a programmer!" One of my CPA friends created one in an evening that answers PTO/Vacation questions for her staff. If we can parse tax code, we can figure out this stuff 😉

I'm sold! Who can make custom GPTs, and how much does it cost?

If you have a ChatGPT Plus (recommended anyway) or a ChatGPT Enterprise (or, spoiler alert, a ChatGPT Team) account, you can create your own custom GPT. There's no additional cost; you can keep it to yourself, share a link, or publish it to the store (just… please… no more TaxGPTs). You access the custom GPTs at the top left.

What's this about a ChatGPT Team Subscription?

Well, it's brilliant, really. Before now, you could have ChatGPT (free), ChatGPT Plus ($20/month), or ChatGPT Enterprise (untold riches). But to share your 'Really Cool Prompts' with your coworkers, you had to create a 'Prompt Library.' And the best 'Really Cool Prompts' are typically a more extended conversation than just a one-off prompt.

Here are the items included (with a heart next to the most exciting items):

GPT-4 with 32K context window

What's a '32K context window', and why should you care? Right now, the context window for ChatGPT Plus is still the meager 4,000 tokens from way back in the early days. That means that after about 3,700-ish characters, your ChatGPT conversation will start getting a little… weird. It will start forgetting things you had discussed before and eventually losing the thread. Hard.

There are ways around this, including building a prompt that requires ChatGPT to repeat the most important items throughout the discussion. Still, if you used ChatGPT as a 'Business Diary Bot' during early 2023 tax season, you may not have known that. 😢

32K context window means that instead of losing the thread after 3,700-ish characters, you have about 28,000-ish characters to spill your guts in. This isn't an earth-shattering number by any means - Anthropic's Claude 2 has a 200k context window - but for almost everything you'll probably be using ChatGPT for, this will be the perfect number of tokens (unless you're dropping in entire publications).

No training on your business data or conversations

We all know that anything we drop into our friend ChatGPT goes into the training hopper, right? Even if we pay for Plus? For example, if you drop something confidential into ChatGPT and haven't switched on the 'Incognito' option, it'll get chucked into the training data to fuel the model upgrades? We all know this, right? (I saw on Twitter/X recently that someone had fed a confidential/NDA contract into ChatGPT to understand the terms better - please don't do that).

With the ChatGPT Team subscription, your data and conversations won't go to train the model. That doesn't mean you should just let client data fly - your conversations will *still* sit on the OpenAI server, subject to their data and privacy policy.

Secure workspace for your team

Right now, even if you're paying for ChatGPT Plus for your whole office, you're still paying for little islands of ChatGPT outside of your control. This brings them together and under your oversight.

From what I can tell, you get fewer admin controls than you would if you had the 'Enterprise' subscription (costing untold riches). However, you still bring your whole team together.

Create and share custom GPTs with your workspace

The ChatGPT Team subscription allows you to create and manage custom GPTs within your secure workspace. Right now, even if two employees had ChatGPT Plus, one person would have to share a link to a custom GPT for the other person to use it. And links are notoriously difficult to control once they are shared outside your organization. "Hey, my job has this really cool AI chatbot that you should try out…"

Having the ChatGPT Team subscription allows you to control the sharing of your custom GPTs (presumably loaded up with really cool tools that you created for your staff). This allows everyone to use the 'Really Cool Prompts' that have been turned into a custom GPT without having to go to the Prompt Library (which they weren't doing anyway).

But What About…

• If you pay monthly, it costs $30/month, and you must have at least two people using it.

• The message cap is now 100 for every three hours. I usually only hit message caps when I'm building prompts. Still, I could see a new staff member needing the 100-message cap (when I was a new staff member, I was told that I had contacted our tax software company for help more than anyone else in the office).

• The rumor is there will be a ChatGPT Business subscription at some point. The Business subscription would have more control over how folks use the tool but less expensive than Enterprise. For now, though, Team is a nice step between Plus and Enterprise for a business.

Bonus: Best Practices for your custom GPTs

1. Think about all of the small things that get in the way of your staff doing their job now and create a custom GPT for it:

   1. Emails (fill 'er up with templates!)

   2. Tax Facts (hey, can you tell me what the annual exclusion is this year?)

   3. Accounting Facts (where does the debit go again?)

   4. Company policies (when are we closed, when are we open, how do I ask for a day off?)

   5. And more…

2. A custom GPT is only as good as your prompt and your data. Take some time learning prompt work so your team can lean on GenAI for the easy stuff and you for the hard stuff.

3. GenAI still needs an outside plugin to math well. It can math okayyyy, but only if you set it up correctly.

4. Create single-purpose custom GPTs first rather than trying to build a Swiss-Army-Knife GPT. GenAI does better when it knows its job and has tidy data. Introducing lots of random jobs and data will go about as well as just putting your Thanksgiving dinner in a blender because "It's all going to the same place anyway."

5. Add Knowledge, but even better is to learn how to create 'Actions' (basically, this is another word for 'Plugins'). This feels like the perfect summer intern project…

Thanks for hanging out with me! Let me know what dream custom GPT you would build in the comments.

Happy Chatting!