Effectively Using AI as a Roofer

Pete: All right everybody. We are live back with another masterclass. Excited to see everybody jump back on here in the chat already going crazy. We are joined today by a special guest star.

One of our own Roofr employees here, Rachel Gould. So welcome Rachel.  

Rachel: Thank you.  

Pete: Rachel is our resident expert for this topic, so, we thought it was pertinent that you jump on here and, and help us, wade through this since Nic and I are not experts on this.  

I think it's cool.  

Nic: I just like smashing the keyboard and seeing what happens. Yeah, cool. Stuff comes out and, Rachel hates me for it, but it's okay.  

Pete: So, and Nic joining us all the way from Italy today.

So welcome Rachel. I'll give you a second here to introduce yourself and, a little bit of what you do here at Roofr.

Rachel: Hi everyone. Thanks for joining.

I'm Rachel. I am the engineering manager here at Roofr. I manage our growth team as well as our new AI platform team. So I've been working for the better part of the last two years on, everything to do with AI at Roofr. And I'm excited to intro the topic to everyone.  

Pete: I think we've got a poll here to start it up. How familiar are you with AI here?  

Nic: What is an LLM?  

Rachel: I guess I can get into that. A large language model or an LLM is basically a deep learning model.

So at the most basic explanation, if you're an expert in data science, you might. Post in the chat that this is not the technical explanation, but the way I like to understand it is that it's like a next word predictor. So basically it's a machine learning algorithm that's been trained on a ton of data, like the entire internet.

And, given a sequence of words, it will predict the next word in the sequence. So when we ask, have you used an LLM before? We mean chat GPT, we mean Claude, Grok, copilot. Those are all large language models also known as generative ai.  

Nic: So which LLM do you guys have most experience with?

Pete: While people are answering here, explain to them what generative means. Right. So I think there's a couple different versions of ai.

Rachel: Yeah, so generative AI basically means given a training data set, you're generating new information. So that can be text, it can be images. I think on the promo for this webinar, there was a picture of me.

With shoulders. There were no shoulders in that picture. It was cutting off at my neck and somebody from our marketing team used some sort of generative AI tool to give me shoulders. So that is an example. And LLMs are also generative ai, but they're focused more on text.

Although some of them are dabbling now in images and things as well.  

Nic: You're saying this one here.

Rachel: Yeah, there were no shoulders. Yes.  

Nic: That is wild.  

Rachel: So that's a scarf.  

Pete: You would never know.  

Rachel: No, that was a scarf, not like a sweater or whatever it turned into.

So this, this is a good segue, I guess, into sort of how generative AI works is like whatever model was used to create my shoulders here had been trained on. Photos of women with shoulders. So they took the general sort of average of what a shoulder looks like and they applied it to me.

Maybe I'm like, some swollen, muscular person. This doesn't look like my real shoulders. They didn't know that. It's just the average.  

Nic: I heard an interesting stat the other day that almost all women have shoulders. And that blew my mind.  

Pete: Almost all. I like that. Almost all.  

Nic: It was like 99% of women have shoulders, so it threw me off.

Pete: We did a survey, what, back in 2025. Where we surveyed a ton of Roofrs just to kind of figure out how they were using tech, how they were employing Roofr, that type of thing. Only 24% of roofers have adopted advanced tech in their business.

Still a lot of people potentially under utilizing tech throughout the business.

Nic: It's very interesting if you think that you're behind, you're not because your competition isn't using AI yet, but you wanna get ahead of the curve.

The stats show that high tech Roofrs they're closing jobs 16.4% faster than low tech ones. So using AI and using technology is gonna keep you ahead, and the landscape is changing so quickly. You don't have to just use it on the speed to lead, race to face side of things. You can use it a lot on the marketing side too.  

Pete: I've essentially replaced going to Google for stuff with chat GPT, right? Now it's more conversational. The average search in Google is like eight words, and the average chat GPT conversation around the same topic is like 352 or something like that.

The LLMs constantly feeding you follow up questions and narrowing the scope of what you're talking about.  

Nic: I, brought up a good question that I wanted to ask Rachel. When Pete was talking about searching and I was talking about searching, those are prompts that we're building out, but there's context too. So what's the difference between prompts and context?

Rachel: So prompts are actually a part of the context. A prompt is like, you know, if you go to chat GPT and you say like, what's the weather in Paris? Right? That's your prompt. And the context is all of the additional information that it has access to, that it's consulting in order to.

Answer your question. So the prompt is your question. The context is like what the AI knows, why it cares, who it is. And I think one thing that I find really fascinating about these models is, so I talked about it being like a next. Word predictor. If you're using Chat GPT or Claude or something to write an email to your customer, you maybe wanna like sell them on something.

And you tell them like, okay, act as you know, a customer service rep and answer this customer email. Versus you're a detail oriented roofing expert who helps sell to customers. Now answer this email. Those are all part of your context and you're now giving that model a mission of like.

Who it is and what it cares about. You're basically giving it a character. So it will be better at predicting the next word in the sequence when it's writing that email. And to make it even better, you could give it samples of emails that you've written that are similar to what you're asking it.

And that will all be fed into the context and it will all be consulted. When it's formulating the reply to give you, a better quality reply. And it's really like, prompting is really an art, not a science. In the really early days, there were all these crazy studies showing that like, if you offered chat GPT a tip, it would do a better job.

I'll give you $5 if you do a good job. It was actually scientifically proven that it would be better, which is hilarious because like it can't use money. It's a large language model. Um, can't

Nic: Use money yet.  

Rachel: Money yet. Exactly. But the trained on all this content, like it knows, oh, if you're, if you're gonna get a tip, you need to do a really good job.

So as I said, it's an art, not a science. And I think the key is like experimenting with your sort of context engineering.  

Nic: That's something I've gotten really good at over the past couple years. And really learned how to build it out.

Like if you're not great with Google sheets and stuff like that and you really wanna split the numbers and understand the difference, that's great, but. You really need to give that prompt context. My prompt today was four paragraphs long breaking down piece by piece by piece on what I'm looking for, exactly how I want it, the tabs that I want below, the columns I want above.

And then I was continuously giving it context to refine that data as well. Give it as much context as possible. Talk to it like it's a 5-year-old. With a science degree. I find that works out very well for me.

Rachel: Also the phrase download your memory sounds insane if you're not familiar with what you're talking about.

But basically what that means, for people who aren't familiar is like you can go into these in your settings for these services and download your memories, which is like all of the context and things that it's learned about you and the way you like to work. So you can really see like what context it's working off of.

To specify example inputs and outputs. If you have certain types of emails that you're responding to for customers, like somebody who's interested in a roof replacement versus a repair versus like insurance work.

Like if you give examples of here's the input, here's an output I've done in the past. With a really clear format. Again, it's recognizing a pattern. That's all it's really doing. So the more you give it that pattern, it will hook into that and be able to give you much better outputs. So I think that's really key.

You can use projects in Claude. Or custom GPTs and just reuse them and keep refining them. It gets better and better as you give it feedback. Right.  

Pete: I think the biggest thing that I've found using it is like, it's essentially trained me on how to prompt it.

Right. Like it's taught. Yeah. Like I, I've found that being as specific as possible helps you to not have a million follow-up questions and constantly have to refine.

Rachel: If you are consistently adding more and more context to your prompt it's really good to a point. Like there is a hard limit. So I also recommend that everybody like, makes sure that you're sort of revisiting it so you're not just only adding, if you're adding, adding and never taking out the things that are less relevant.

Your mileage may vary.  

Nic: Every Monday we do a kickoff with all the weeks priors metrics and everything else. I was able to kind of slowly figure it out and then build an agent so that's scheduled and does all the work and pulls all the data on Monday morning at 8:00 AM so everything kind of flows through. Is there any other ways that you use agents in your day-to-day, Rachel, that you would recommend here?

Rachel: Yeah. So I guess just to like level set first, like what is the difference between an agent and an LLM too? 'cause I think mm-hmm. That's something that people are really confused about in general. Like, if you have played with chat GPT or Claude in the olden days, like a year ago you know, you would've chatted back and forth , it didn't really have a lot of capabilities.

Now it has all these connectors. You can connect it to Gmail, you can connect it to all these different software tools you use. And it really is like a proper agent. Now, an agent is basically just an LLM that has the ability to call preset tools. So the LLM is like the brain, and the tools are like the hands, you're basically giving it access to additional context that it can discover on its own through calling of these tools.

If you ask like, what's the weather? It'll just guess like what's the most likely weather. But if you give it access to a tool where it can call like the Weather Network or something, or, or get access to the intranet it would actually be able to Google what's the weather, where this person is located and give you a proper answer.

So in terms of the tools I use, like I've set up a ton of workflows, basically anything that I really hate doing, I have tried my best. To automate it with an agent because I really wanna spend my time doing stuff that I feel is valuable, that only I can do. You know, I've set up a bunch of things related to our development process with agents, but one agent I'm really, really excited about that we've built it's called Roofr voice, which is our AI receptionist at Roofr that's currently in beta. Roofr voice is an agent in that, you know, if a homeowner calls a Roofr, Roofr voice will answer. It's a agent that has voice capabilities and it will talk to your homeowner.

And actually it'll ask them questions about, for example, where they're located. So if the homeowner shares their address. That agent can call our address resolution tool and make sure that it's a real house. It'll find exactly where that is. It'll get the footprint of the roof. So that later if the homeowner asks on the phone like, Hey, I'd really love an estimate for a roof replacement, we can then call our estimate tool and take those measurements and actually give the homeowner an estimate right on the phone.

So, in the time of, you know, a three minute long phone conversation, we've collected the address information. Given an estimate and basically qualified that lead because we can have the LLM judge you know, how well the homeowner reacted to that estimate, right? So stuff like that is where I get really excited and I think the sky's the limit with like building agents and there's definitely a lot of cool stuff that we have in the works here.

Pete: Let's talk a little bit, Rachel, about let's say I haven't really employed it or, or put it into my business at all.

What are some ways that I could start using it? In regards to my roofing company?  

Rachel: If there's something that you find yourself doing, you're spending a ton of time on, like writing copy I think takes a lot of time, like answering emails. These are all things that can easily be automated with ai. Using chat GPT or using Claude. Like you can use projects or GPTs depending on uh, which tool you use, and you can save and reuse context so that you're basically developing almost like a knowledge base of your own

work projects or your work processes, and continuously leverage those. So for example, I would build up a process for if there's an inbound email basically classifying that type of email and feeding it. Examples of how I've responded to emails like that in the past to just remove the sort of critical thinking that's required.

Basically you wanna turn yourself from the person that is doing this really manual drudgery to the person that's just overseeing ai, doing that work and making sure that you know it did a good job and that you agree with the way it answered the email or whatever it is.

Nic: There's a lot of conversation out there for like, oh my God, AI is gonna come for your jobs. If you are just punching data in and out all day, maybe. But if you're in sales, if you're running a business, if you're a Roofr, it's not going to take your job. It's just gonna enhance that. It's gonna give you that time back for these meaningless parts of your day.

That you're running through and allow you to do more with that time. Those emails, lead scoring, content strategy, visual analysis, all this stuff, and really start to build out a process for it.

If you guys are in any type of commercial or any type of like bidding system in which you have to go through, you know, spec writers and stuff like that, using AI for that. Is a game changer. You can scour all the bidding sites out there, government sites and all that stuff, and put your name in for that.

Just using something like Claude and like the MCP for Chrome and training it to do so. There's great ways to get in front of people that you'll never get in front of before because it just took too much time and having it do that so you could focus on running the company and building more stuff in your CRM.

Is a huge win on that aspect. Mm-hmm. Nic's advice corner back with a vengeance, another  

Pete: appearance. I like it.  

Rachel: I think Joel mentioned this in the chat, brainstorming with AI is also awesome. So if you open up a new chat GPT conversation or a new Claude conversation and you say like, Hey look, I wanna find more time in my day.

Here are the top time sucks. It will suggest things that it can help you with, and even give you resources for the specific tools that you use and the processes that you use, which is, you know, pretty crazy.

You're getting like AI to build your own AI process, but it's honestly what works the best. Same with prompt writing. If you know that you wanna use AI to automate any given thing, you can give it information and say write a prompt that I can use to get you to do this thing, and it will suggest one.

It's feels really silly while you're doing it, for sure. But it's a great tip.  

Pete: Yeah, using AI to feed the AI is pretty effective actually.  

Rachel: And as you do that if you find that you're not getting good results, you can basically manually create what we call evals, which evaluate. The AI's performance based on a set of inputs.

So if you're vibe coding something or, or building an email you can give it examples that you've made up to try and throw it curve balls, and then see how it does. Get Claude or whatever to iterate on the prompt based on what it didn't do well.

Nic: That's crazy. One of the best analogies for anybody who's just getting into AI think of it as you just hired for $0, a personal assistant that is there to do whatever you want, whenever you want. That person's gonna give you all that time back in the day.

It's like, oh dude, I could sell a lot more if I'm not. Sifting through leads or looking for that, or lead scoring or whatever the case is.

My best advice for it is just be curious. If you don't know how to use it, just start using it and asking it questions like Rachel said. Try something out, see if it works, build this, take this data that I have from my CRM. Analyze it for me, and then ask it questions and see how it kind of builds out.

And the more you do it, the more you'll figure out.  

Rachel: Yeah, and you never know where the value will be. As we were working on building Roofr voice as we've onboarded more and more customers into the beta, a really great sort of outcome of it that we weren't even necessarily anticipating is like spam management.

Roofers get a ton of spam calls every day. With AI answering the phone you know, you're gonna get summaries when it is a real customer calling you. You're gonna get summaries about where they are and what they're looking for, but you also don't have to listen to people selling you X, Y, Z and I'm listening to all of the phone calls that are coming in through Roofr voice.

And the amount of like sales calls is absolutely crazy. Anything that's high touch that takes up time that you can't avoid is a great use case for ai for sure.  

Pete: I think back to seven or eight years ago when I first came to this industry, at that point aerial measurements were kind of a thing that people were skittish on.

Oh, they can't be accurate. Why would I use them? That's not how I do things. And now look at, here we are today, and if you're still potentially doing it manually, you're the odd man out at this point, right? That's become part of the everyday process for probably 99% of roofing companies.

This is the next kind of revolution of the industry and, it's gonna change the way that we do business. And so, like Nic said, just be curious. Get out in front of it and see what you can do to potentially leverage it in the business. I mean, there's a million different use cases.

The Roofr voice that Rachel's talking about is just the tip of the iceberg here for us.

Rachel: I saw some questions in the chat about how to use Roofr with ai.

Mm-hmm. It's something that we're focusing on a lot. My team is working on building tools to make it much easier to get. The benefits of AI with the data that you have in Roofr. So we're working on a lot of things related to actually being able to query your data in Roofr using ai.

Currently, we don't have an API sort of connector that allows you to pull your data. Into something like chat, GPT. The reason for that is there are security concerns. Like it's something that we've vetted and we've thought about, and we still may do it, but we wanna make sure that when we build it, it's built really safely because ultimately it's your customer's data, right?

Anytime we're sending that to any AI tool there's definitely privacy and security concerns there. In terms of what we have next up. Roofr voice is just gonna get better and better.

We also have Roofr sites that allows you to spin up a website so easily using AI and the information that you have inside Roofr. And then as I said, we're building up a sort of infrastructure so that when you're inside Roofr, you can chat with your data and you can do things like, ask what job is most likely to close?

Where should I spend my energy next? So that's all stuff that we're working on. 'Cause we know that it's incredibly important for all of our customers to be getting all of the great benefits of AI inside Roofr as well.  

Nic: For anybody who's on the scale plan, the performance dashboard is just

life changing. Being able to compare and contrast from different date periods and stuff. Imagine that with predictive analytics in there as well. That's coming. It's going to be really fantastic.  

Rachel: I will mention too, there's some connectors that you can use that sort of indirectly hook in with functionality that we already have in Roofr. For example, I'm a Claude advocate.

I love Claude. I haven't used chat GPT in a while, I'm assuming you can do this in chat GPT as well, but you can connect it with your Gmail account. So like if you have an automation set up within Roofr to send you an email when something happens, you could use that as a trigger as well within ai.

Sometimes you just need to work on some different workarounds if you're looking to build a custom trigger until we have these features live in the app.  

Nic: I think the next slide's all about you, Pete, because of, your name.  

Pete: Process is a big part of this, right? I literally was on a call not too long ago with someone who said like, oh, I took my workflow and plugged it into an LLM and let it audit my workflow, you know, to see whether I needed more stages, whether I had stages I could remove.

Thinking about the process as a whole, based on the information that he had fed it, he was able to clean up his process and organize himself a little bit better. Start small,

the email piece that Rachel brought up a couple times in the beginning, that's a great place to start, right? Like we could just get it, helping us writing emails, answering emails, that kind of stuff.  

Nic: It's just like building a roofing company. Start small, get organized, lean on partners. Find the CRM that you need and add it as you go. You do not have to be Rachel when you first get in. It's a process and each day you get better.

Rachel: The technology is also changing all the time. Like literally week to week, it's changing. I think the other thing that's important is you don't have to adapt every single time. The technology changes. You can start your process and gradually add on as things get easier and as things change.

It can feel so overwhelming to look at all the tools that are out there and even try and understand like which of them are legit and which aren't. You do have to be a little careful as well with data security.

It's not an exciting topic at all, but it is important. As Claude adds more connectors, you'll be able to just scale the process you build essentially. Which is really great. 'cause every time they add a new one, I'm like, oh, I can do this now.  

Nic: I think we got a couple more slides here. Rachel, can you dig into this a little bit?  

Rachel: Yeah. At the bottom there we have just manual processes. Writing out on a piece of paper, keeping a spreadsheet as your CRM, digitizing it is obviously using something more like Roofr to keep track of everything.

And then, above that we have actually automating it. So you're actually leveraging the software as like a power user. You're building workflows. 'Cause as much as we are continuously adding more AI features inside Roofr, we do have automations And, you know, you don't need to use AI for everything.

You can also automate things without ai. That's a feature that people get a ton of value out of. And then above that we have like predictions. So actually using AI to leverage the data you have, noticing patterns. Building workflows, and then predicting things based on that. So you can really.

Get the most leverage and understand where to focus your efforts.

Pete: Can you describe what would be some discerning factors of like, if I was gonna automate something versus potentially use AI for it?

Rachel: I think one sort of factor that people don't take into account a lot with LLMs and with AI is that you can't predict with a hundred percent accuracy that given these inputs, you'll get the same output. They're not deterministic. Which is fine, you know, when you're writing an email or when you are, having a conversation with somebody.

And it doesn't matter that you use specific exact words to reply to them because. There's not more than one right way to answer a phone call. And so using AI for something like that makes sense because again, you're predicting it.

Automation is deterministic. It means given this set of inputs, you're gonna get exactly this set of outputs. For example, inside Roofr, if you set up an automation rule saying like, okay, every job that's tagged with follow up should generate a task for Bob.

To follow up. You don't need AI for that because you know if there is a job tagged follow up, you want Bob to follow up. There is no reason to get a large language model involved in that unless you are getting it to, for example, write the follow up, write the email. If you want a hundred percent accuracy and you need to know that something will be responded to in exactly this specific way you should be using. Something like an automation rule not a large language model that is predicting.

And that's really important because I think even within like the tech industry software companies are building AI, everything and it's like really exciting. But there's also a lot of places where it's actually better to use the old way of doing things or plug AI in where it matters.

You don't want an AI running your entire production process. You just don't, because things will slip between the cracks. You can plug it in where it adds value versus dictating the whole thing.  

Nic: I just had this conversation with my parents too and like building out automations and stuff and they're like, well, we just wanna make sure it's our voice. 'cause they saw stuff on AI and they were worried about that. It's a great point. Like you don't wanna automate all your comms.

I've been using Claude or chat GPT at least for the better part of two years, and trained it out pretty well to my voice. And I still have to go write it better, more into my voice as an extra prompt afterwards to really make sure doesn't sound like a robot talking. When you're talking with customers, the more homey feels the better, right?

Especially on follow up, some sensitive stuff like payment follow ups as well. Invoice follow ups.

Rachel: I'm glad you mentioned that too, 'cause to give the example of payments, that's another example where I wouldn't necessarily use ai. Anything that deals with really sensitive information.

If you're taking credit card information for your customers, anything like that you don't really wanna be using like a homemade AI process to Handle that information. There's a reason why there's a ton of security standards that companies like ours have to live up to.

There's been tons of stories in the news of like people using Claude code to build things and it exposes API keys, it can expose sensitive information. So I'm not saying that's to make people scared, but like it is healthy to have.

Some amount of caution. the risk of sounding like an AI wrote your email is. It's there, but the risk of an AI taking your customer's credit card information is, is a lot more, more severe, obviously.

So you just wanna give

This has been realtalk with Rachel.,

Pete: You don't want AI to completely change the way you do business at this point. You want it more to supplement what you're doing. And so just playing around with it and seeing where you can fit it in is, is the ideal way to handle it at this point.  

Nic: And I think the second point there, understand buzz versus reality is super important.

There are so many people on social that go into like some crazy buzz stuff and I was like, dude, that's not feasible. People have all these ideas, manipulate the stock market

You gotta be able to have. Reality built into it. Systems that can scale with you, that can grow with you, that you can build on later.

Pete: Rachel, it looks like you have an invite to do a, in-person training.  

Rachel: Yeah, we can talk after what? We'll see, see what my fee is? I dunno.  

Nic: Thank you so much, Rachel, for everything. We got the spinning wheel left to do so Joel hit it.

Congrats on, I think this is the first like swag bag that we're giving away, not just the socks.  

Pete: Thank you Rachel. Thank you for jumping on. Tons of great information here. Obviously, this is an evolving landscape.

I think we've answered most of the questions right, Nic, along the way.  

Rachel: And I would love to come back and answer some of them in more detail once we have some more features that can hook in with AI. As you said, constantly evolving.

Pete: I'm sure we'll have Rachel back at some point to discuss, what her team's working on and what they've come up with.

Nic: Check out Roofr sites, check out Roofr voice. Those are some really awesome tools out there.

Rachel: Yeah, absolutely. Awesome. Thanks for having me.  

Pete: Thank you everybody for jumping on and we will see you next time on the Roofr Masterclass. Thanks everybody.