Build an AI sales assistant with Google Sheets and Slack

Chris: It was more neat needs must than anything else. The agency struggled a fair amount in the last couple of years. And so I spent time learning how to code focusing on Python and that allowed me then to start building AI products. So I spend a bit of time on the side of the agency. Building and consulting with companies on AI products and processes.

And with that, I have a ton of, spent a ton of time playing with AI, should I say, maybe too much time. I get distracted really easily, but it's a lot of fun. I use mostly replit AI flawed and GPT forward to code. So while I did learn can do it manually now, I ironically rely on the AI models to do a lot of it.

And they do an incredible job. So spinning stuff up takes me next to no time, I want to say, because I have The experience now behind it of coding and I leverage those tools really heavily. I'm actually going to share my screen now. All right to do that? Yep, we should be good to go.

Stop.

Can everybody see my slideshow screen? Sweet. Yeah, so that was obviously a pretty brief intro. I'm here to talk about building an AI sales assistant with Google Sheets and Slack. It's something we've leveraged pretty heavily in the agency that we have. I've already gone through this. You can see on the right hand side, I tried an early attempt of AI photo generation, built some custom models that makes you look like a hamster, but hopefully doesn't look like me right now.

I've tinkered with so much stuff on the AI side, A lot of it initially was how to help Flixer because as I mentioned, we did struggle for a while there as a company. So I was trying to figure out ways, like how could we use AI to cut costs, identify new revenue opportunities make us more efficient internally.

Through that learning, I've now started taking on consulting on the side to help bridge the gap for the revenue for the agency. But going through that process, I did a lot of random stuff. I attempted to use AI to produce videos and we do a lot of UGC like social media, influencer based videos on TikTok and Meta and Instagram, YouTube, and my initial.

I guess concern six, nine months ago was AI's just gonna eat this business up because of all the AI influencer progress. If you've seen tools like Hagen or Synthesia, DID they're getting really good and it was a pretty big concern of mine. So I started tinkering with those tools and trying to figure out is that something we could do.

And we built something pretty cool there, actually been able to generate those videos fairly fast. We'd start by creating an influencer in one of the image models, and then we'd feed it into HeyGen with a script and a voice from Eleven Labs, and then produce a video. Tried selling that to our clients, and actually, they didn't want it, surprisingly.

I don't know if that'll change in future, but for now, there's a, an anti AI creative amongst our clients, at least. And now, granted, there are bigger brands like Zynga and King. com and folks like that. But they have an anti AI policy actually internally on their creative side of things.

So I don't know if that will change when authenticity gets better and it looks more real.

Ben: But is that Sorry, is that so that, that feels like the main challenge at the moment is people just pushing on a blanket. We don't want to use AI generated avatars in videos right now. You

Chris: They have policies internally, yeah that stop them from using that.

Even if the policies were not in place, I'm not sure it would take off yet. I just don't think the quality's good enough. And when you're working with brands like we are, they're pretty pedantic about The the feedback rounds and we're going through the video creation, right? They're paying 1, 000 for a video and then they're spending way more than that on the media behind it.

They have strict policies in place about how they look and There's a lot of feedback rounds oftentimes. We're doing three four or five feedback rounds I'm, not sure if AI will solve that problem. It definitely doesn't now I can't I don't know Give the answer of whether it will in a year or five years time, but obviously it's getting better.

Like I've seen some incredible quality AI influencer stuff in the last week, even, but you can still tell the lips are fake and it still actually looks to read the two perfect. If that makes sense, like the backgrounds are too perfect. It's not like a real human doing it. And that's the trend on.

The creative side of things on social media recently has been someone doing it on a selfie camera on their phone. So it looks very, it looks genuine, it is genuine, it's authentic. AI ironically has to get better on the face, but then worse in the surroundings to make it look more real.

We've also tried it for concept and script writing. Seemed like a really interesting opportunity. Where would it be, 38?

Sorry? Is that? Yes, course. Yeah. We attempted to use it to write our concepts and scripts. We do all that concept in the script writing with a human team from scratch for every single client. And we thought, all right, let's throw this into chat GPT or Claude. We'll give it some context of all the stuff we've done in the past and generate some concept ideas for us that we can push out to our clients.

What we found is it, if you send the, what the output of the chat GPT or Claude straight to the client, they could tell it was AI pretty much every single time. And so we have come to a. point now where we use it to generate some ideas and then the human is taking it to the next step and actually polishing it off and finishing it and make it sound more human.

We've had challenges on the scripts as well, sounding too fake. It's too enthusiastic at times, which we just have a human then finish that off. We also tried it. We thought it was a really interesting use case using vision models to analyze the performance of creative. So we take a set of videos and then we get the performance data from our clients.

And we fed all of that data into a vision and asked it to summarize like what drove performance, what didn't drive performance, give us some output from that. And it felt like initially it did a really good job of it, but then when we tried the same videos multiple times, there wasn't really enough consistency in it.

So while it might get there in future, when we attempted that, it didn't work enough for our use case. And then the final thing we've tried is on the sales side and has been by far the most successful use case for us internally for the agency. We were spending ages on lead research. We have an SDR over in the Philippines.

We used to have more than one actually, and it just took a long time to find the right companies to sell to, find the right people to sell to, then research those people, draft the emails, follow up, et cetera, et cetera. So we thought what are things we could build to make this process easier and get us a lot more efficient on that side of things?

And something I've, I feel like I've experienced that AI is when it's equipped with the right tools is really good at research and really good at summarizing information from lots of places very quickly, right? And so that's the angle we took and that's where we've had a lot more success than we have with the other tools.

You might ask why build ourselves? There are tons of AI styles tools out there. I feel like LinkedIn has just served me so much of this stuff now because I've liked some people's. Posts and I'm getting nonstop AI sales content from all these folks, big companies that build these tools. I took a look at a lot of them and I just thought they were too complex and or enterprise level.

High price, Flixer is a small company. We have a limited budget. I didn't want to go spending thousands a month on a tool that had all these bells and whistles that quite frankly we didn't need. So we thought let's try to build it ourselves. Something that has a really simple use case that a 12 year old could use.

I'm not saying their SDRs are 12, they're not, obviously but we wanted something super simple that you could just give to somebody and say, three steps, get it done. You're not going to be spending ages trying to learn this system and embedded where the team spends time. We spend most of that time in Slack and Google Sheets.

The hypothesis also is that a Google sheet or spreadsheet is one of the lowest common denominator tools that almost everybody knows how to use. I don't think I've met many people that haven't ever used a spreadsheet before, and that's the case for our SDRs too, right? They're all well versed in spreadsheets.

If we have to try to teach them another tool, there's a lift there, and it hasn't always gone as well as we'd like. So that's where we. We started and we've got three tools so far. I'll demo each of these shortly. We have a Google sheet, AI prospecting tool that automates finding people. It researches the company and each person that we have on the list, and then it drafts emails.

We don't allow it to send yet because I don't think we can trust AI to actually get it nailed every time on the emails they send out. We have it connecting to CRM and Apollo that sends automatic followups out. If they don't reply to the first email, which oftentimes people don't, we then automate the followups and ultimately RSDR just needs something simple and out of the box that's ready to go.

So the instructions for them is very straightforward. We built a Slack signals bot that researches the sales targets and existing accounts automatically from that sheet, and then every single week it sends an alert to us in Slack with product launches, new hires. Personnel changes things like buyer interest.

So we're taking each of the profiles we're tracking on from LinkedIn. We look at what they're liking and what they're commenting and what they're posting on LinkedIn. And then we summarize that and put this into Slack. And since we launched this product only three weeks ago, we've already got two upsells in with information we wouldn't have known.

Otherwise. Because we don't get that information out of our clients, right? We, it's impossible for us to track and understand every time Zynga launched a new game, they have a million of them. And that stuff is often in the press or talked about on social. And then the final one we built, which is the newest one is the Google Sheet LinkedIn profile analysis.

It really started by a deep dive. We've taken more of a comedic term with it for now, which I'll show on here. As a personal LinkedIn analysis and we actually roasts, roasted my LinkedIn profile and I'll roast Ben's on this as well. It's pretty funny. Let's see what comes back. Maybe I'll go in and actually do the demos now, and then I'll talk through about how. We experimented and started that before I do, Ben, any questions?

Ben: No, yeah, I guess the question is you said it connects to HubSpot and Apollo. Is that a requirement? Can there be other tools? Does it need to do that?

Chris: No, it doesn't need to. It's just how we built this custom for us. I have a friend of mine who I built one for now as well, and we're hooking it to Salesforce for them, for him. Doesn't matter which CLRM it is, if you have a CLRM or not. You don't have to, once you hook into it we, we are because it's where our reporting comes out of in HubSpot and just due diligence around our sales process.

And the only reason we did Apollo is because the follow, the automated follow ups, right? We thought, could we do it out of the sheets itself? And it, it was too challenging. In a spreadsheet. So we just build that into Apollo. And what happens every time you send an email from here? It automatically adds it as a contact.

The person has a contact in Apollo, which triggers a sequence then as well. And that's all custom built through here. But, when we've done that now, it doesn't take very long for us to do that moving forward. Yeah, this is the sheet. Pretty straightforward. We have a bunch of input.

So for Flixer, we're selling to anybody with this job title and we just went through and identified, these are the target job title keywords, more so the job titles themselves. So we track people who have that in their job title. We also track any jobs that the target companies are hiring for with those same job titles, as that is a signal for us, right?

If someone's targeting a creative production manager or trying to hire a creative production manager, we can go in there and attempt to sell to them a kind of interim services while they find those people we had just a company blurb about me, the sender. And an email display name, and we use that then to quote unquote train the AI.

It's not really training I'm just giving it context to improve the output. And then we have the signals that we're tracking in here into Slack. Company LinkedIn posts, press mentions, relevant new hires, relevant job posts. And then posts and engagement from the target buyers. And then we are working on these at the moment, where we find game launches in the press, and in new apps, or apps that are moving up the charts, in the app stores, which is important for us to track.

And then any industry influencer posts. There's several influencers out there that are well respected in our industry. In our space, mobile gaming, and they often get a lot of engagement on their posts. So we'll start tracking those.

Ben: So that says they're not yet live. Is that not yet live in this tool or is that, yeah, how's that?

Chris: We are working on these right now. Yes. Then

Ben: you could set your own signals that you would be looking for.

Chris: You could customize the signals pretty easily. Yeah. So this is for mobile gaming specific, but something we've thought of is, could you track like public company earnings reports? Yeah, we could do that as public information.

So we could scrape that. And if you could scrape off the web Is a data source we could do a buddy of mine was texting me over the weekend. And if we could, he has a self storage facility, completely different business. And he was wondering if we could track prices from competitors and we have the URL.

So yeah, we could pull that down as a single too. And if they change price we could pull that into Slack easily.

You see some of the pre fill stuff in here. So I'll

I'll do a gaming company play Tika. You'll see the stuff above for anything that's pulled in for previous ones, just pon you tested. And it takes a few moments to start pulling in, but you'll see gradually the information start pulling out. We just from an input target company or all, that's all we need to tell the sdr r to do is an enter the URLs and then it starts pulling in company name, the LinkedIn number of employees industry.

We have the last LinkedIn post, the date, time of that. A list of press mentions when that came in the three month employee change we actually did find a way to get even more information on what company insights in particular departments at companies we can pull a recent executive hires.

We can pull isn't in the sheet at the moment, but we've just found ways to do that. Are they hiring for anybody in our target? Target job posts are marketing or created for us. Yes, they are. We're pulling the LinkedIn about, and then we ask GPT force to a quick summary. How does the company made money?

And what does the company do? And the goal of that is to help us when we actually draft the emails.

Ben: And is the data from LinkedIn. So yeah, and just asked in the comments, is it called. from the LinkedIn API. Is that how you're doing it? Like an agent?

Chris: I will actually show that shortly and go through the process that we use.

There's not unofficial, we haven't got access to an official LinkedIn API so it's scraping via a third party API which I can show some of those sources for those. There are quite a few of them out there now. Cool. So that's the company level stuff. So there's more information we can pull in where we started and very straightforward instruction for the SDR is just paste the URLs in here and it will start pulling down.

And then after that, automatically we get our target contacts and this, instead of preloading everything and then going from there we created a way for the SDR to select the ones that are relevant. As you can see in here, play are the one I just ran. There are, I dunno how many, a ton of results.

25. A lot of them are the same job title. And so we wanted to have the SDR actually go through and critically look at some of them and say, this the right person. Not I don't think I've seen this many of the same job title elsewhere. If you look at Zinga in here, they've got some with creative. The matches is probably the most relevant for us to Pinterest.

They have a creative strategist. They're the ones we prioritize and all else. We can pull in, CMO and we'll take a look at one of the one once you click enrich on the SDR clicks enrich, it actually kicks off a process then. To to analyze those individual people and pull in their contact information as well.

So you'll see here, we just did Gil looks at the last LinkedIn post for that person. That was four days ago. Any sign of that person mentioned in the press, to be honest, I've not seen any individual we've pulled. And this is obviously a demo sheet, but even across the other one where we're pulling hundreds and hundreds, I haven't seen any individual mentioned in the press yet.

So it's probably. Not needed. We have AI, summarizer, buyer motivation, why would this person buy from Flixr based on the services that we have, and then we have it draft an email. My experience with email so far, it's a good job of drafting. I personally, I always change or adjust the message a little bit.

It feels a little too forced sometimes. But given it's in a spreadsheet, we can do that, right? You can just change it whenever you want. And you can actually then send the email from here automatically. We also built in a way to change. The body automatically. So you'll notice in here this will change shortly.

It drafts a new version and we could change the subject as well. We'll draft a new version. Once we click that

pretty, pretty straightforward. Pretty cool. We can pull this information in and use it in our outbound. It's something that otherwise the SDR would have spent ages doing, right? They click on the profile and they start. Diving in and trying to understand more about it. Whereas we try, the goal here is to bring this all in one view on the spreadsheet.

So it's easy for the team to use.

Any questions on, this is the last part of this sheet. I

Ben: don't think so. I assume the CC BC, C, you just, you can add your own email.

Chris: Yeah. Yeah. Cool. Yeah, that's that's pretty much it. And then it'll just send that with that on the CCC or BC I don't really know why we had that is the, I think SDR wanted to BC, CA particular HubSpot email.

Added both those in. Works, works well. Of course that's the sheets. What we then added on top of here is a, the Signals product that we use. This is both for our new business like Cold Outbound from here, but also we have a separate sheet with our list of accounts and our contacts at those accounts.

I'm going to open up Slack and show you the output of that. Every single week we run A list. We've actually run this process and these are demo ones that the team put together for this morning, but we pull in a summary of activity from the Company and from the the list of people every single week.

And you see what Paul here is like any job postings that are relevant for us any engagements, any insights on any other posts they're making. Things like that might be relevant for us. Something that we've had success here two Zynga launched a new game and talked about it. On LinkedIn, which we had no idea about, and that has led to an opposite opportunity for us.

Working on a new creative campaign, which is awesome. So immediately that was payback. And then another company we're tracking is a new business target, site play announced they were looking for a connected TV creative strategist, which we had no idea about either. And we sell. Part of our product is a connected TV creative angle, and we have a call set up with them to actually try selling them a new creative campaign on the TV side of things.

But both signals that we wouldn't have known about otherwise, and takes, a second to run every single week. The difficulty we've had with the signal so far is continuous improvement. There are lots of signals out there. Sometimes some companies don't have very much. Some companies are very active and trying to get those signals distilled down is the next challenge.

We've done a pretty good job on these. And if I run into some of these, you'll notice. These aren't too bad either, there can be a lot, and some of them are irrelevant. So we're working through how to make sure they're relevant, and asking GPT 4 to give it a relevancy score before posting it, and then just giving us the top five is the current approach.

With that, it'll involve some more tinkering to make sure we do bring things that are relevant. We also don't miss stuff that is that I think that's given the challenge involves some testing. We're working on that right now, but it all keys off this spreadsheet. So any company that's added on here or any spreadsheet in the Flixer spreadsheet that we have, it pulls it from that company list and it pulls it from the contact list in here and just gives us a lot of intelligence on those.

That allows us already to upsell from the accounts and they're creating opportunities and interesting signals to reach out about, right? Something I think I struggle with and I've seen in market is there's a lot of noise out there right now when you try and cold outbound and actually having some relevant messaging around signals that you've tracked.

It gives you a higher chance of response from these potential customers.

Ben: So is it like, obviously you this is the first part is this prospecting sheet, and then you actually add the companies that you're looking for, but then you have those signals coming through the Slack. How do they communicate with the system to create that outreach email, or is there no actual connection there?

Because obviously you're saying this would automatically populate an email and then you might have a signal next week. It would be the perfect time to go and draft an email. So how does that all work?

Chris: Yeah, the current, the only drafting this email is based on the signals from this sheet at the moment.

It doesn't come in from the continuous signals later on. That's an interesting idea, right? To think about how we do that at the moment, we're relying on the account management team and, or the SDR to use those signals and then draft the email themselves. When it's signal based. This one is just based on this sheet, which is the sheet is a one time snapshot of that, of what we've detected at that moment in time.

It could be an interesting next step of the signals products to take those and actually drafted the email or the SDR for the account manager to use. But we haven't got to that point yet.

Cool. I'm just checking on the time. I think we have enough time to go into the next one. So the LinkedIn profile analysis tool. This is our newest one. Then I'm going to find your URL. Bear with me a second. Oh dear.

Ben: And how Remco's just asking in the, in that, in the other prospecting sheet, how does the trigger of entering a URL work to start off the filling of the data?

Chris: It's just a Google Apps script trigger. So it detects when there's a change in any cell in that column and it kicks off the process at that point.

Yeah it's pretty straightforward and rudimentary to be honest. And then we hook it up to a library that we built. And it really is a, I'll go through the process shortly after this tool, but it pulls information from APIs and then it structures that, sends it in multiple different calls to GPT for to do all the kind of summary behind that.

I'll walk through more of that process and how we got there shortly.

Cool. All right, let's analyze Ben's LinkedIn, didn't

take too long. We have a kind of a comedic element to this, the roast, and then. We actually go through some more serious stuff that. Professionals could use underneath, and if you guys can read that, I'll be a bit small,

but to do nothing is to look busy. That's a life

Ben: motto of mine. So that's fine.

Chris: So we do a roast and here that's just for shits and giggles. Summary of the profile, how to optimize it keyword analysis, strengths, weaknesses, stuff is still loaded and then increase engagement It started initially with looking at the posts and the likes and the comments from everybody's posts to try and understand for us internally, LinkedIn and I was trying to figure out how can I get more engaged with my posts and learn like what works, what doesn't, right?

I post pretty much every two or three days. Certain things do better than others. I don't have any kind of scientific analysis of that right now. I just post. I don't have a strategy. And my goal here was build something that will allow me to figure out what is the right strategy we post on LinkedIn.

Is there something that I can do better topic wise media content wise? And I think this is not perfect, right? But it is an analysis that we can start to use to increase that engagement. There's a fun little tool that we put together to help that and send this out to you, Ben, if you want to take a look at.

Ben: Yeah, that'd be great. It's

Chris: helpful.

Cool. And that's this actually took this tool pretty straightforward, right? So same kind of flow as the other stuff, maybe a bit simpler overall. We're just taking the data that we pull from the LinkedIn API and the three different endpoints we use pulling the profile data, pulling the engagement data, and then pulling the posts.

And then it really is GPT 4 doing the rest of the work here. We've had to tinker with the prompts, but GPT 4 is doing the lion's share of it. So it's fairly straightforward in the backend. It's like an input to GPT 4 and get the output back. And that's generally how we've done most.

Of the tools. They're all pretty straightforward. We identify the data sources and I'll show one of them in here. A rapid API is a, where we get most of our end points from. I don't know if anyone's used it before. There's a ton of stuff on here. You do need to be technical to, to get access to it, obviously.

But once you figure that out it's, there's a treasure trove of information. This one end point in particular. You can get profile data just from the username, URL, you can search for people on LinkedIn, you can get profiles, posts, comments, but there's so much information here and that's just on the pro on the individual profile.

You can get information on the company insights, that's stuff that we just had just been released on here. We can pull out like employee movements, recent executive hires, things like that. So what we initially did, we just. In the playground here, we got the data, what was returned, I copied it and then literally just pasted it into Jack GBT, just to see what would happen.

And so roast my LinkedIn profile was a good example, really easy. I just pasted the payload, hit go, and just to see what it comes back with. It's an easy way of experimenting with the prompting for us. We do a fair bit through the API too, but. Rather than having to write code immediately, then deploy and test.

We just thought let's throw it in here. And you can see here, it comes back with a pretty good option. If we make it shorter, roast this LinkedIn profile, but keep to one paragraph, that's too long.

It's an easy way to go back and forth and figure out what's the right way for us to get the output we want. So we did that extensively using all these data, different data providers who have the LinkedIn APIs that give us all LinkedIn for Rapid API, for anything search based, which is press things, I think in future, if we want to pull company results, like stock market results, things like that, we'd probably look in here.

And then crawling, we use something called Firecrawl to crawl multiple nested layers of pages as well. So If you want to pull like a company blog as an example or job posts, you can start with the main page like blog, and then it will actually go pull all the sub pages to so we just find a call for that.

And then we pump the raw data into Chattopadhyay and Claude. I use Claude a fair amount to see what we get. Inca with those, and then we put it into code at that point. And that's allows us to scale things, right? Like back and forth one by one in chat. GPT is no slow. We can do on mass in the spreadsheet, Google app script with code.

And once we got that nailed, then we started adding integrations on. So pushing into Slack HubSpot, Apollo as well. And all in all, they didn't take us that long to do this because of that process, it was really iterative. Results wise this is the cell stuff has worked really well for us as a business Flixer really did struggle last year at the start of this year. It's probably not all down to this, but the fact we've saved. The SDR times so much it's allowed us to push up our outbounds, which in turn, our qualified leaders increased and it's up three X from April to July which is awesome.

And for the amount of time we put into it, isn't it really strong return investment and our existing customer revenues up we've identified the opportunities. From those clients again, not all down to the signals that are coming in, but it really has helped the team like take a holistic view and understand what's going on across that list of clients.

We don't have a massive amount of clients in business. We have 20 that, our accounts we manage. And even managing 20 and understanding what's going on at all these big companies is pretty challenging for the team. So it's been a massive help for us. And hopefully that carries on as we go through the rest of the year.

As we add more accounts into that, that core account list use this tool to grow them. What's next? We're working on more simple tools. It's simple as the key word here can we do things in a spreadsheet in Slack? Something we still spend too much time doing is identifying target companies.

So if you remember back to that tool, we have to add in the company URLs which necessitates you having that list of time companies already. So what we have to do here is generate that list of target accounts for the SDR. Manually, really and give that list over to we want to come with a tool that practically identifies target companies for us.

There'll be different facets to that identifying press mentions more broadly in the industry. We're going to track key influences on social media and engagement in their posts. Competitor social media, so who's engaging with the competitor posts. And then probably the biggest one for us is app store tracking.

We sell to mobile games predominantly, but also have worked with mobile app developers in the past. It doesn't necessarily need to be games. And tracking movements in the app store is a one key indicator for us and something we can do automatically. That's next on our list. We also need to refine those signals in Slack.

There is too much noise. How do we get the summaries right? Make sure it's delivering the right information. It's really like trigger, tinkering with the prompts more than anything else to make sure that it's coming in properly. I'm sure there are a bunch of other ideas that other people might have, or we might get in time that we can add in here too.

make things a lot more efficient for us. And happy to hear more answer questions if you guys have any ideas too.

Ben: Yeah. One thing I wanted to, I think someone asked in the chat earlier was, you said the sequences, like, how does that trigger a sequence? You said it adds a contact to Apollo.

And then I guess you, you've got to write the sequence yourself in Apollo. Is that how that works

Chris: right now? Yeah. We have a pre baked sequence in Apollo. That's just a followup sequence. Basically. I'm sure that could be improved too, it's generic. Right now, the easiest way for us is to a generic follow up sequence that we hit the AP the Apollo API.

It's funny. You have to create a contact first and then you have to add a contact that sequence. Remember, right? So actually it's two API calls. Not a big deal. And it's fairly straightforward, but It would be nicer if it was just one. Yeah that's how that's working at the moment.

I would like to have it from the sheet because then we can maybe come up with some more relevant messaging because at the moment it's just like generic, right? Oh, have you seen this email kind of thing? I get those emails all the time and I think they're garbage, but it's just the easiest way for us to get going on that.

Next step, is there a way to customize those? I'd love to find that out too.

Ben: Yeah, that makes sense. If other people have comments and stuff, if you can raise your hand on Zoom, I can try and get to a few of you. And yeah, so how does someone go and use this, use a system like this, use this tool or these tools?

How how do you do that?

Chris: That's a good question. There's a fair amount of custom code involved in there, right? So if you want to follow me after, we can look at creating one for you, maybe. I've created one for another friend of mine at the moment and he's using it extensively as well.

I guess it depends on how much customization in there. If you have engineering capabilities, then it's something you could go and build. That's the one hurdle, I guess it's not necessarily off the shelf because it is custom tool that we've built and doesn't necessitate some engineering resource, but it is fairly quick.

So if you have that ability to write in Google Apps Script or Python yourselves then building these tools really isn't that challenging. Yeah, it probably took us a day. Overall, I did that LinkedIn profile on last week and it took probably two or three hours. As I mentioned at the start, Claude or Repl.

ai or GPT 4 are doing pretty much all my coding and the way my floater works. I give it a scope. I write scope out first. I send it into one of those LLMs. It gives me a draft code base. I plug into. Replit or Google Apps Script in this case, and then run it and test. Invariably it doesn't work first time.

And so I have to go through multiple steps of bug fixing. The bug fixes don't take very long and then you can run it. It works. The process, I think, Ben, I'm sure you've done some of this, right? Just. Testing and learning going through the LLM chat providers.

Ben: Yeah, exactly. Yeah, so yeah, what I'd love to see is just make a copy of this sheet.

And then there's an instructions page that is if you want to tweak it, do this. Want to do that. And I wonder if you and I sync up afterwards to figure out how we can either. Do something like that. Or if we do a tutorial showing basically what you've done. So that, yeah, we can show others how to do it as well.

Cause I definitely want to do that prospecting sheet. We need to figure that out ourselves. So yeah, we'd love to do that. I think others would appreciate a tutorial on that as well. So yeah, we'll try and figure that out. I just

Chris: have to figure out a way to, you have to put your API keys in there.

How do we get the API keys in there safely. I can figure it out on my side.

Ben: Like a blank template with into API key here. Otherwise, yeah, cool. Yeah. If you folks want to see a tutorial on that, so we'll try and figure that out. Yeah. Rem Remco has asked a couple of questions here.

I don't know if you want to come on live Ram code to ask, if not, I can ask on your behalf,

I'll do it. So what are the SDR stats? Are the open rates replies and meetings box. I'm using this system.

Chris: I don't have them to hand right now. I'm afraid. I know our leads have gone up. I can get those and follow up on those.

Ben: And then, yeah, info on the ballpark costs of running this system.

Chris: Yes, the prospecting sheet is cheaper than the signals because the signals are continuous, right?

So the prospecting sheet is a one off. We get people from the Apollo API and we pay a subscription for the Apollo, for Apollo at the moment which I think is nine to nine bucks a month. I don't think they have any limits on the amount of people you can pull, but there are rate limits on how often you do it.

So that's the only limitation there. Other than that, the LinkedIn data API from RapidAPI, that I think is like half a cent per run, so it's not particularly expensive and GPT 4 we use mini and mini is extremely cheap. They're negligible. For that the biggest cost is a LinkedIn data API.

And I think that you think about pushing this out to more people, the rate limits are really just where the challenge lies. So how many can you do in one go? So the friend that we built this for is using our account. And if we're using it alongside that account and we start to hit rate limits, I hit rate limits yesterday actually already.

So I just need to increase my plan size on those or figure out solutions to I think that's a much bigger lift. The scrape ourselves from LinkedIn. There is reliable third parties for now. Cool. Overall pretty cheap, right? Like it costs us a huge amount of money to run and you just need subscriptions, different things.

Yeah.

Ben: Yeah. So it seems like obviously Apollo. Which is why I've just cancelled is one of the ones that I need to re renew.

Chris: There are ways around it. There are other search things. We just chose Apollo because we have a subscription there. You can pull results from Google directly like backdoor in Google custom search.

There are people that are pretty expensive. There are search. APIs on RapidAPI 2, which we haven't looked at, but we can look at those as well and look at alternatives to try and make it as like a simple, seamless as possible without having to get a bunch of different subscriptions to other tools.

Ben: Yeah, makes sense. Any other questions from anyone in the crowd? Yeah, if you raise your hand on the zoom button I can then get you on to ask if there's anything else

pop it in the chat. If you don't want to speak live, Cam, you have a question.

Guest: Chris, thank you so much. That was awesome. Love the using the Google sheets to figure all this out. It's great. Instead of having to learn a new tool. That was really cool. I think all of us in the audience, we want to build something like this.

What are the things that you had wish you had known before you did it? So we can not step in the hole or on the rake and hit us in the face. Like it did to you when you had to build this thing.

Chris: I'm honest, not a lot hit us in the face. It was really experimenting with the data API. So I think finding the right data is key.

As soon as you've got the right data, GPT 4 does a really good job of Summarizing, I think it's one of AI's best use cases summarizing. We did initially send the emails without checking them. I wouldn't do that. I, the, I've seen plenty of Ai, SDR things out there now, and I feel like I'm getting sold them all the time.

I just don't think AI is in a place where it can replicate your own voice in a realistic tone. It would send out emails that weren't terrible. It just, I can tell that AI generated, it's like way too over enthusiastic way too salesy. And that's not the tone that our team employs. We don't tend to try selling a first email anyway.

We tend to just ask a question and it has got better by feeding it a bunch of examples, but it's still, I wouldn't trust it to send out straight away. So I think that kind of QA. Step from a human is really important. I know that our team changes emails before they get sent out. Yes. They use the text as inspiration, but they do change some of the wording.

Ben: Yeah. I think in our,

Chris: other than that, I'm here.

Ben: I was going to say in our previous workshop with Barbara, she talked about how the tone and just kept on saying you gotta say, make it better. And yeah, Kashav on our team says, there's one thing that he says in the prompt that has helped to any of the writing, which is make it chill vibe, but still business focus.

Cause otherwise it was too to extreme on the chill vibe. Cause yeah, in the example email that was generated, then it's still you can just tell that even first sentence, you're just like, Oh, it's a bit, Oh, you've been sharing great content on this thing. You're like, yeah. Okay. So yeah, I think that there's like a load of prompting and like you said, giving examples and everything else to try and keep making that bit better.

I think Claude always seems like the best. writing style I've seen, or most natural at least. But it's almost as soon as you say, write an email, it's like when you say you write a tweet, but don't put hashtags, it'll always do a really out there tweet with hashtags. So I wonder if there's something about the context of You're writing an email to a prospect that just automatically prompts the LLM to sound like that.

Yeah, very interesting. We've got a question from Jacob. I will click the button for you.

Guest: Hey Chris, thanks for doing this. Did you ever evaluate any of the generative AI sales tools? I know Reggie AI is a big one and we're seeing some of our clients purchase it. So just curious if you've heard of it or how you compared it to your own process that you built out.

Chris: I haven't heard of regi.

ai. The big one we have added is Clay. I think Clay is probably, I'm going to say the darling of it. It's just, they seem to have a lot of momentum right now. And I could use Clay, but I'm technical, right? So I think when it came to our SDR team, we wanted something that they were just Working in that's a spreadsheet and clay is a technically a spreadsheet within a in an interface, but you have to set up all these other flows and I Zapier.

I love Zapier, I use Zapier a lot in our business, but I think you need to be somewhat technical or product brain to be able to use Zapier. There are some things in clay where you have to pass variables in from different steps to different steps and a lot of. Folks aren't technical enough to get there, I don't think which is fine.

RSDR and sales team are not I want to something that they're used to dealing with and spreadsheet is pretty much always common denominator there. And all we have to tell her was enter URLs in here and it won't do the job. Without any sort of other. Complications there. At the signal side, I've looked at a few common room, probably to be rather based in Seattle as well.

It looks like a really cool system, but just, it's expensive and over complicated for what we need. I was like, let's just get into Slack. It's super simple. Our team are in there eight hours a day in a way. I don't want them going somewhere else. Let's make it as easy as possible. Get it into Slack.

They can use that information as they want. And maybe in future it goes somewhere else. I don't know let's see. But for now, simple was best for us.

Ben: And Alex, I've.

Guest: Yeah. First of all, thank you. The this was really impressive just fantastic workflow. I have a short question iot, several tools you have mentioned in your slides, such as, let's say EX a.ai or fire crawler. Could you please share maybe a couple of words about them? So just like to make us familiar with them better.

Thank you.

Chris: Exa. ai, their whole spiel, I think it is accurate based on how I've used it they're like AI, it's a different way of search, right? So if you use Google search Google search API it doesn't do a great job of saying like companies similar to X, whereas Exa. ai, you can go in there and say, give me, Competitors to a URL, right?

Competitors to Flixer Studios, and they'll give me 50 and they'll tell me like a snippet of information or a highlight from that page. And Google search just doesn't do a great job of that. So I think XR is built for, I don't know the right way to put it as AI search, whereas a Google is traditional search, if that makes sense.

It's worth experimenting with it. If you get a bunch of free credits and there's a playground in there, so you don't have to be technical to do it. You can just go in and play around with the searches and see what comes back. Fire crawl, I like a lot because of the nested crawling. Typically if you use a crawler you will go to a page and it'll crawl the contents from a single page.

The use case for us is can we pull from Zynga's blog, because they announce things on their blog, right? They announce game launches on there too. If you just use traditional crawl, oftentimes what happens is you We'll get the blog page and the titles and everything else on there, but you won't go deeper.

Firecrawl is native. If you want it to go into all the sub pages on that blog page, it will go and do it and it will pull it all in for you pretty instantaneously. So I've been really impressed with Firecrawl so far. Obviously it does come across some roadblocks, right? We did try Firecrawl for crawling some job pages where people use third party it's like a, where you put a job out there and it's not necessarily on your own website, it's on a third party website. That sometimes hits an error if they block it, but typically Firecore would seem to work really well for our use case. Greenhouse, that's it yeah, there's one like a greenhouse, I'm trying to find the word for it.

Yeah,

Ben: And then there's another one you said, was it SERPA? What's that one?

Chris: Serpa, yeah that's a Google search tool. You also can try that out. You get 3, 000 credits, I think. And that's great for news. There's like news APIs out there, which are really expensive. News. com or something. And it's I'm not paying that for for pulling news on companies.

And Serpa does that for you. If you just go in and create your query, it pulls the news, like Google News, basically. And it does a really good job. And It's not free after a thousand credits, but it's cheap and it's our solution for pulling the news press mentions for any company that we're tracking.

Ben: Awesome. Chris, yeah, thanks so much for your time and showing us this. We all want to use it now, so we've got to figure that out. But yeah let's figure out a way to either you make it super simple or we do a tutorial or both. Thanks everyone for joining and asking questions and stuff.

And yeah, we'll see you on the next one, but yeah. Cheers, Chris. Appreciate it.

Chris: Thanks everybody.

‍