AI-Powered Customer Support
Custom GPTs, Assistants API, Vector databases, and how to make sense of it all.
Hey there!👋
I still can't wrap my head around the fact that it's Friday once again. And since it most certainly is, here's another issue of the Minty Analyst newsletter.
Getting Started
Lately, I've been playing with AI more and more. I believe it's safe to say we all have! I've always been a massive fan of the idea. I used OpenAI when they first granted access to their GPT 2 API, trying to use AI to do financial analysis as part of our FP&A platform Magnimetrics. It didn't work back then, but since ChatGPT turned the world on its head, it's become more feasible.
In the second half of this year, I started thinking of AI, especially LLMs, as more than a toy, as something that can be effectively applied within a business setting.
Most of those models cost too much to build and run within a small business, so initially, the benefits were reserved for large enterprises. However, OpenAI, all the open-source projects, and various other players effectively changed that.
Current State
You can now sign up for ChatGPT Plus for $20/month and chat briefly with ChatGPT (pun intended!). It will custom-configure your GPT implementation to access proprietary information you provide as files.
This is not really something new. Vector databases have been around for some time, and using them to expand LLMs' knowledge beyond their context window has become more accessible and cheaper than ever before.
However, OpenAI made it much more straightforward with their Custom GPTs for frontend users and their Assistants API for developers.
The possibilities are nearly endless (and things will get much more exciting in the future). Still, my interest over the past few weeks has been solely around using AI to automate customer support.
Don't get me wrong, I love engaging with Magnimetrics customers, but I can only reply to so many chat messages and emails. Additionally, I can only work for so many hours. Wouldn't it be great if AI could answer for me?
It would. And it is, right now!
Right now, the chat bubble on our website is not manned by anyone. It's an AI model that I trained on our website pages and a few FAQs. And this has become so simple!
I wanted to prove how simple it really was, so I sat down with my father and asked him to develop the knowledge base to serve his clients. He put together a large doc file and uploaded it to the service I'm using. After some help on my end in picking the branding colors and adding a logo (yes, that was the hard part), his chatbot was up and running.
If you want to try it for yourself, I suggest checking out AidBase! It’s awesome and they have a great free plan, which is practically unlimited if you bring your own OpenAI API key. It’s a pretty sweet deal!
Build or Buy
There is an abundance of services that add additional features. Still, when all is said and done, it's a very similar experience everywhere.
All those SaaS businesses would initially use their API key and charge customers via usage-based plans. But that quickly changed to more and more companies charging you a flat monthly fee while asking you to bring your own key (BYOK... excellent term, just came up with it). That way, you incur the usage fees at a lower price, and you only pay $20-$100/month for their chat widget and additional features.
Building your own implementation is more straightforward than ever if you want to cut down on those costs even more. You don't even need to code. Every no-code web builder out there has an OpenAI integration. You can set up an assistant and ask questions with a simple form.
Granted, it is a bit harder to implement if you want to leverage the entire conversation functionality where you effectively chat with the model. But you can still do it quite easily, even with no-code.
What else?
So, what else can we do with the Assistants API and the Custom GPTs? I don't really like chat apps and don't believe successful apps will be ones where you effectively chat with the AI.
I can really see the enormous value in what I described above - customer support chatbots. It's incredible how easily we can take care of more than 80% of the chat messages we receive on our websites by having an AI read our website and respond instead of us.
But I can also see another very interesting way to leverage such fine-tuned models as internal helpers.
I recently had this idea about a service that has access to all IFRS and US GAAP texts and can answer specific questions. I reached out to some accountants and friends in finance and received some lukewarm interest. The idea was so simple that it would take me a day to build the whole thing with no-code tools (including payment collection).
I'd use Softr.io with Airtable and Make for the MVP to take care of all the plans, payments, usage, etc. Of course, payments will happen through Stripe's payment links (the easiest no-code payment solution).
But before I did that, I needed to get all the published IFRS and US GAAP, not only the standards but also all the accompanying documents that discuss the application of the standards. I contacted the standard issuing bodies behind both IFRS and US GAAP. I was told the standards themselves are copyrighted material and I am not allowed to train AI models. So, the project stopped before it started.
However, I recently needed to consult IFRS, and it took me around 45 minutes to find it online. I got annoyed and built a custom GPT. I downloaded a 5,000-page IFRS guidebook from EY's website, split it into 10 files of 500 pages each, uploaded it to my custom GPT, and asked the same question.
The whole thing took me 25 minutes (10 of which was spent figuring out how to split the PDF...), and the answer was much better.
This is another area I see getting huge in the coming months, especially with the expected rollout of OpenAI's marketplace.
We live in a time when the advancements in AI, particularly in LLMs like ChatGPT, have opened the floodgates to incredible possibilities. The applications are endless, from building internal tools with proprietary knowledge to building customer support bots or whatever else awaits us shortly.
We have yet to see even faster innovation in this sector, and we probably can't even imagine how AI will revolutionize our work and play over the next year.
So, I say strap in, grab some popcorn and enjoy the ride. Be open, adopt new ideas, and stay on the frontlines of this movement so you are not left behind.
I guess that's it for today. Honestly, it ended up being a bit different than I expected, so if you made it this far, thank you for reading! As always, extra brownie points for those who share the post with a friend or two.
So, think about how custom GPTs and the Assistants API can help you daily, and go for it. Who knows, you may come up with a $100m startup idea.
If you want to ask or share anything, just hit reply and let me know!
Oh, I almost forgot: Merry Christmas!🎄
Till next time,
Dobri