Below is a transcript for our conversation, The Check with Joseki Tech | Episode 1: HubBoost #1.
Adam: Hi, my name is Adam and I'm here from Joseki Tech. Joining me today is Jasmin, our director of marketing, and Dan, our VP of engineering. We're here to show you what we've been working on lately, which is a new product feature we're calling Hub Boost, where we're trying to do an integration between HubSpot and OpenAI to start using generative AI in our sales process. With that, Dan, do you want to show us how it works?
Dan: Sure. This is HubSpot here, and we're going to proceed into contacts. What we're trying to do here is use OpenAI to help us in generating a more personalized sales email. So we proceed to a contact life begins with a contact life apparently begins slowly. Yes, it does. Okay, so here we have a couple of contacts. I'm just going to select one.
Adam: So we've downloaded some data from LinkedIn about this. Obviously, Jasmin's not taking email for Bill Gates, but we wanted to have a little bit more interesting records here to play with in HubSpot.
Dan: So we'll go to send them an email. And this is the first point or place where we start to see the Hub Boost integration. If you look carefully, you'll notice that there's an additional menu link here, hub boost. So again, we're going to use AI to help us try and generate a sales email.
Adam: So Dan, what we're looking at here, this is a. Is it a Chrome plugin that we developed? I mean, how do we do this basic integration here with HubSpot?
Dan: Yes, that's correct. HubSpot does have some extension mechanisms, but we rolled with a custom Google Chrome plugin. And once the plugin, you know, it injects the menu link here and then we have a handler on the menu link. So when you press that, it will use the information within HubSpot in order to try and generate a personalized email. In this particular case, we're doing two things. We're generating both the subject and the message. Jasmin, did you have any feedback about this message that you'd like to alter?
Jasmin: So thank you for kind of pivoting to me, Dan. I was introduced to HubSpot just a few months ago. Our goals in this was more so to reach, like Dan said, that higher level of personalization and also reduce our lead up time to send out these messages. Personalization can take a very long time, especially if it's not automated. So adding in our own automation onto HubSpot has been kind of a small lifesaver for us so far. Hub Boost has kind of been integrated into our personal systems. This is not something that is for sale necessarily. So I have used it a small amount and now I just kind of want to take a look at what we are getting fed back from the AI just based off the profile that we are actually looking at.
So once again, the profile that we're looking at is Bill. And so let's take a look at the email. Dear Bill Gates, just off the top of my head, it is a little bit long, a little bit wordy. I would understand that we are trying to target a kind of big fish in this situation, something a little wordier would make sense. However, I think in this case I would give some feedback.
Dan: Okay, so we have the ability here to provide feedback to the mechanism. So much like when you're using ChatGPT, you have the ability to kind of tweak the results. So that's exactly what the feedback does here. If you need something tweaked, then type in the instructions and press regenerate. It'll take that feedback and then communicate with the AI engine again to come up with a better sample.
Adam: And what are the sort of the combined inputs that the AI is using to create the email?
Dan: Right now we're using the name, title, contacts, title and the contacts about
information.
Adam: Okay. And that we're pulling into the context of this sort of conversation, if you will, that OpenAI has, that doesn't, that information doesn't sort of enter the cloud magically. So that we're teaching the terminator machines anything about our customers. Does it stay private to us there?
Dan: Correct? Yeah, it stays private to us. We're actually using a special feature of OpenAI called knowledge bases. So we're doing essentially knowledge-based retrieval for that information, but it's kept local.
Adam: Okay, so the. Sorry, just to dig in a little bit deeper. So we're using information we have from HubSpot, and so far we're just using those fields. And I think as we sort of work and find what's working for us, we could add other fields we think might be more interesting. Maybe their past job history, maybe their skills profile, maybe other people they're following to help enrich the conversation. But I think.
Dan: Exactly correct. We're in line with version two now that incorporates both skills and positions.
Adam: Oh, cool. So the other side of this to me for potentially for other customers is it doesn't have to be the LinkedIn info. It could be what their sales history was or what their contact history was, or what insurance products they currently use with your company so that we could inform the conversation better in a bespoke manner. There. Now, the other side of it is the baseline to build the message. We've also fed it into the knowledge base, a bunch of information about who Joseki Tech is. So, I mean, Jasmin, didn't you supply Dan with a bunch of documents? What was the baseline there?
Jasmin: So I did. I actually was hoping to pivot to a little bit more about the creation of Hub Boost in general. So this is a great time. Initially, we would set this up with the OpenAI assistant feature. So this assistant that we kind of created was using a bunch of proprietary information from Jaseki already. So any standards that we have for just general email writing, images that we might include, and kind of just the knowledge base of the kind of work that Jaseki tech does, as well as the kind of work that we want to do for our clients.
Adam: Yeah, and if I recall, we started with essentially taking a copy of our website, putting that information in there, and then a series of. We had some notes sitting around of projects we've worked on in the past. I think we fed that in and then we fed in some documents that you use with our clients about sort of best practices for marketing emails.
Jasmin: Exactly. So we started off with, like you said, a list of our old projects that we had listed previously on our website. So it is taking in all the information from our website currently, as well as any of our previous sites that we made, as well as any information that I have written that we share with our clients about email marketing and other general marketing tips that we would like our OpenAI assistant to abide by. And so by having all of that information kind of fit in there already, it tailors it even more to just how we want to message this person and the voice that we have and combining the voice that they are presenting out to the public. So, like I mentioned, LinkedIn has what the person chooses to put out into the world.
So that is how they want to be represented. So that is how we want to essentially have them represented in our contacts and in our personalization, by combining kind of our life story and our life history with their life story and their life history or career history. Better said, we can actually get that super tailored message using OpenAI with information that has already been fed so that we don't need to do it every single time. So part of the problem that I have had using AI in the past was that it would take a super long time to just type in every single bit of information that I had over and over again. So that's where the AI assistant came in and we could backload all information.
But now with the combination of Hub Boost, we can target without me necessarily needing to know exactly everything about this person. But then we go in and we tweak it, and as I learn more about our potential client, we are joining into the best message for them.
Adam: So in terms of actually getting this shipped out, what's the next? How do I actually click send on this?
Dan: Sure. So once you are happy and you've polished up your message and perhaps subject, it's just a simple matter of pressing copy here and transferring your subject over, and then transferring the message over and press and send. Now, obviously we probably don't want to send this to Mister Bill right now.
Adam: Excellent.
Jasmin: One other note for right now. So I know that this is something that we are working on right now, but I wanted to see where were on it. I had a conversation with you, Dan, previously about how we may tailor this message for the person based off their profile, but that may not be the message that I want to send to them. So now we have the ability to hopefully type in a little bit of a sample message of what we would like to send them and then generate a new message for them based off that base knowledge that we give them to the ChatGPT function that you have where you can go back and forth and say this is the basis. That is what we're kind of hoping to as one of our next steps.
Dan: That's correct. We'll either be starting with a combination of either a template or free form text. So if you select a template here and you have a prefab email that you'd like to start with as a baseline, or you can type, you type here, and when you press hub boost, that information will flow forward to the next screen and it'll be provided through to AI. Use that as a base template in generating its response.
Adam: Yeah, I think that's going to be super helpful because right now it's just everything starts from the same point. And we do know we may have a specific message. Knowing the customer a little bit, we want to hone in on, and they're probably going to fall into a couple of different basic patterns. So now Jasmin can just go in to create a new email template in HubSpot. She doesn't need any help from the IT side and we can just get it rolling. So I think that's going to be super useful for speeding this around. But I do think what we're looking at here is a really strong baseline. We always talk about you got to get your plumbing set up first. So there's a lot behind the scenes here that I think is worth mentioning.
Adam: We talked about we set up the service so we stood up that infrastructure and it OpenAI things. So it's talking to ChatGPT-4, it's got our knowledge base there. It's got the baseline offer. That service also understands the notion of a conversation. So when you do that regenerate, it's keeping track of what's going on. So you can add on the next just like you were doing a chat session there, but you're doing that remotely. And likewise, Jasmin can be doing one chat session and I can be doing another chat session, and we don't bump into each other to kind of cross wires. So we had to do a lot of plumbing to get that set up. So I'm feeling like we got a good baseline there.
Adam: And we've also got a good baseline over here in HubSpot where you don't have to go out to a different agent. You can stay right in the tool. It's going to pass that customer data over to the service call. It has the means for you to do that conversation and we can do the cutting and pasting. There were a few challenges there, and perhaps we'll talk about that on another call where we dive into the, you know, the joys of building a HubSpot chrome extension. But there were a couple challenges we had to overcome to get that working. But I think this sets us up for a really strong starting point in our next conversation. We're going to get into the test in tune for all of this. We built an email that's in there and we've shown that we can tweak it.
Adam: But how do we fundamentally up the quality of our starting point to get a better baseline? That's what we're going to talk about next. Thanks everybody, for the time and attention today, and we will hopefully see you all soon.
Dan: Thank you.
Comments