Smart AI Workflows Start with Better Processes [Endless Customers Podcast Ep. 178]

This transcript has been generated by AI and not checked for accuracy. 

[00:00:00] Bob Ruffolo: So you're ready to add more AI to the way your business operates. Right now, companies are racing to build AI agents, automate workflows, and plug AI into every part of their business. But if you're making the same mistakes most businesses are making, AI won't make you better. It'll actually give you more problems.

[00:00:19] Bob Ruffolo: In this episode, Tom Nasser and I break down how to decide where AI actually belongs in a business, what should be automated without AI, and why some of the best AI workflows use a lot less AI than you might expect. We'll also show you real examples, including a system Tom and I are building together at IMPACT, and some of the risks you need to understand before you give AI too much access inside your company.

[00:00:43] Bob Ruffolo: If you're planning to invest in AI, build agents, or automate more of your operation, you'll thank me for watching this episode. Let's get into it.​

[00:01:35] Bob Ruffolo: And welcome back to the Endless Customers Podcast. We have a great episode today. I have a, a very good friend of mine on with us today, uh, Tom Nasser from a company called X-Ray.

[00:01:46] Bob Ruffolo: We're gonna be talking about AI, automations, how should you be bringing AI to your organization, how maybe you shouldn't be bringing AI to your organization. We're gonna cover all of this today. But Tom, welcome to the show. Good to see you, my friend.

[00:01:58] Tom Nassr: Bob, thanks for having me. [00:02:00] I'm, I'm really excited to be here. I think we have a lot of really interesting things to talk about today

[00:02:04] Bob Ruffolo: Oh, for sure. For sure. And, and I, I will say this right at the, the, at the start of the episode, um, Tom's company, X-Ray, you may have seen him at Endless Customers Live, um, l- uh, last year, and he'll be back here again this year at the, at the conference with us in Hartford in a few weeks. We have a partnership between our two companies, and, uh, Tom offers the Endless Customers community, um, assistance in building business workflows, which we're gonna go all into using AI, using automations.

[00:02:34] Bob Ruffolo: Um, and we've learned a ton from him. We're doing projects together. We will go into all of that. So, so Tom, let's just, uh, we'll kick it off right here. Um, and, and briefly tell us about your business.

[00:02:46] Tom Nassr: Sure. Uh, so I like to say that XRAY helps people feel like they're working in the future, so doing things at work that are worthy of their time. So we call it workflow design. We design the way that people get [00:03:00] things done, and that means we design systems around the things that people are really good at and people like to do. So AI and automation is very secondary to designing the way that the process actually works. when you have a process and systems that actually empower the people, you feel really good about getting things done and about the work that you have to do every day. So yeah, if you feel

[00:03:24] Bob Ruffolo: Tom, yeah.

[00:03:25] Tom Nassr: you're, you're in a good spot

[00:03:27] Bob Ruffolo: And I've known Tom since 2018, I wanna say. Whenever we moved down to New Haven, maybe 2019 I think we met each other

[00:03:35] Tom Nassr: I think it was 2015, maybe

[00:03:37] Bob Ruffolo: Oh, oh,

[00:03:38] Tom Nassr: 2016.

[00:03:39] Bob Ruffolo: oh, it's earlier. Wow

[00:03:40] Tom Nassr: it was earlier. 'Cause 2019, I mean, Checkmate was done at 2019. So,

[00:03:46] Bob Ruffolo: Oh, so it must have been, yeah

[00:03:48] Tom Nassr: 2019 we, uh, we sold in October, so it was

[00:03:51] Bob Ruffolo: Wow.

[00:03:52] Tom Nassr: we sold. So yeah, at

[00:03:53] Bob Ruffolo: It's m- it must have been s-

[00:03:54] Tom Nassr: three, four years befor-

[00:03:56] Bob Ruffolo: so

[00:03:56] Tom Nassr: that

[00:03:56] Bob Ruffolo: 10 years. Wow. Um, I've got [00:04:00] more gray hairs, you've got a longer beard now, but that's not the, uh, what the episode is all about. So let's start here. Um, you know, even before we started this episode, you're like, "I might have a contrarian view to how people should be applying AI to your business."

[00:04:15] Bob Ruffolo: So AI is obviously all the rage, and I think that there's certainly a, a place for all businesses to be doing more with AI. Um, but talk to us about your views on AI in business workflows and operations.

[00:04:29] Tom Nassr: Sure. So first off, I think the marketing campaigns around AI have been very effective. Everyone feels like they're behind. Everyone feels like they need to adopt a whole bunch of AI things as soon as possible. The marketing teams at these companies are paid well, and they're spending a lot of money on ads, and that's what gets you in that sort of FOMO environment where you're heightened and you're trying to grasp at everything. And it reminds me a lot of the crypto industry years ago, where it's a [00:05:00] hype cycle after hype cycle after hype cycle, and you just can't keep up. And I want to normalize that for anyone listening right now that, uh, you feel like you can't keep up, it's by design. Uh, and, and that feeling's not gonna go away. So I wanna step back from, you know, waving the magic wand and adopting AI into every facet of your business and compartmentalize things into really two areas. One is experimenting, learning the, the sort of practice of a 10,000 hours, know, type methodology where, you know, if you're gonna be good at anything, you need to spend 10,000 hours at it, right?

[00:05:37] Bob Ruffolo: Oh yeah

[00:05:38] Tom Nassr: there's practice, and a lot of that might be throwaway work. A lot of that might be failed experiments. A lot of that might be just working through things on your own to see what the bounds are about what you're trying to do, which is, by the way, a moving target, right? Six months in, things are gonna change. practice is, is one thing, and then there's actual, you know, [00:06:00] wheels on the road production. Our company depends on the way that these things work. And those are two very different conversations. it's really paralyzing when you're trying to do, you know, the second one, where you're trying to operationalize everything, but you really don't even know how it works.

[00:06:18] Tom Nassr: You really haven't tested and experimented with it enough to even build trust and confidence in yourself, never mind in this external system.

[00:06:26] Bob Ruffolo: Mm-hmm.

[00:06:26] Tom Nassr: And, and that's, that's where you really get frozen and paralyzed and you can't get over that hump, right? You, you just can't get there because you're not practicing, you're not testing and understanding how these things actually work, so you can't ever make it to a production-type environment. So, practice, and in a low-stakes way, build something for your kids, build something as a meal plan prep thing, like something that has no stakes, right? And it's just the [00:07:00] act of getting familiar with how you might utilize these tools in a way that doesn't have dire consequences out an email to all of your clients and that says, "Hi, this is a test." You know, like, you know, you just

[00:07:14] Bob Ruffolo: Yeah.

[00:07:14] Tom Nassr: something like that happening.

[00:07:16] Bob Ruffolo: Yeah

[00:07:16] Tom Nassr: thinking about it in those two frames, I think is, is a helpful way to get started with some of this stuff.

[00:07:22] Bob Ruffolo: So where my mind is going as I'm hearing you saying this is you hear these stories of people that have put AI systems in place, and it could be in call centers, it could be, uh, for customer service and, and obviously content marketing and website optimization and, and you go down the whole list, right?

[00:07:40] Tom Nassr: Yeah, sure

[00:07:41] Bob Ruffolo: but then there's practice.

[00:07:43] Bob Ruffolo: Do you have any, like, percentage of, like, the most of the world is in this bucket and very few companies are in this bucket right now to kind of level set where the world is right now?

[00:07:53] Tom Nassr: Uh, we are extremely early. I will, I will lead with that, where I, I think [00:08:00] the vast majority of companies that are trying to do... A- and, you know, these are, these are, um, you know, gut check numbers. So just in what I'm seeing with

[00:08:10] Bob Ruffolo: CTR

[00:08:10] Tom Nassr: that, that come to us, right? This isn't some, uh, third-party study. But I will say the vast majority of people who think that they need AI systems today actually need good process design today, because it's like anything else, crap in, crap out. If you don't have a good process, the AI systems will not be able to help you. They will just accelerate your disorganization, and, and that's a real, you know, uh, a real sandbag what otherwise could be a really good system, because

[00:08:45] Bob Ruffolo: But process is boring, AI is sexy.

[00:08:48] Tom Nassr: Right. Well, hey, b- and the

[00:08:50] Bob Ruffolo: We g- we need AI, we don't need processes. Blah, we need AI. But, but the reality is, of what you're saying is, it's, it's pr- it's a process thing more than [00:09:00] anything

[00:09:00] Tom Nassr: It is. And it's a process thing more than anything. And the ca- really counterintuitive part of this is inside of your process, the less you actually use AI, the more effective that AI will actually be

[00:09:17] Bob Ruffolo: Why is that?

[00:09:18] Tom Nassr: I'm so glad you asked. So, so the popular belief is that give it more context, give it more documents, give it more whatever, right? And it will produce a better output. Mm, it'll produce a different output, right? It not be better. Better or worse against what, right?

[00:09:39] Bob Ruffolo: Mhm

[00:09:40] Tom Nassr: Um, so the... And, and I'm, I'm gonna get a little technical here, but hang on, I think it'll, I think it'll land AI systems are what's called non-deterministic, meaning you pump it with something, whether it's context or a prompt or whatever, [00:10:00] it is not determined what that output will actually be.

[00:10:04] Tom Nassr: You don't know. Every time-- You could run the same prompt a hundred times, and you'll get a hundred different results, right? That is a feature, not a bug inside of AI. That is why people use it, right? For non-deterministic creation of stuff, generative creation. Deterministic things like a process, like when you put the sausage through the grinder, you get a sausage on the other end, whatever, right?

[00:10:29] Tom Nassr: Like, is how you, you create something. It is deterministic. You know when you have these sets of inputs, you can create this output. I know that when I click send on my invoice, it goes to my customer. That is determined, right? That is happen the same way every time. Now, when you're dealing with a process, some of it is loosey-goosey, some of it is not. Some of it needs to happen the same way every time. Sometimes it [00:11:00] doesn't. When you have a customer that has a complaint, you don't wanna reply to every customer with a complaint with the same exact response. Of course not. You need to adapt and adjust to that. That's where a deterministic or a non-deterministic AI response is gonna be actually helpful to contextualize, oh, they're talking about my account, or they're talking about a billing issue or categorization, right?

[00:11:21] Tom Nassr: They're talking about a, a, a experience they had with a rep. They're talking about some UI bug, right? Categorizing things. That's a really good use of, of an AI. So when you have a process, you wanna make as much of it deterministic as possible so that you can have that consistency and then sprinkle in one or two or three moments of this non-deterministic variability to be able to adapt to different feedback you're getting, to be able to categorize something in a particular way.

[00:11:56] Tom Nassr: And then once it's categorized, it goes down the deterministic path [00:12:00] because you know that all your UX bugs are gonna go to your designer and your design team, right? You know that has to be the path. But we're getting five hundred emails a day. Like, how will we categorize this,

[00:12:13] Bob Ruffolo: Yeah. Yeah

[00:12:14] Tom Nassr: of something non-deterministic

[00:12:16] Bob Ruffolo: Incredible. Um, masterclass of how to be thinking about bringing AI into your business operations because it's not... You don't start with AI, you start with process, you start with deterministic consistency, and then sprinkle in the AI where AI is best. Very different mindset than I think most of our audience probably already had going into, we need AI workflows in our business.

[00:12:44] Bob Ruffolo: Can you give a, um, a real-world example of what you just said that probably applies to most of our audience's business?

[00:12:53] Tom Nassr: Um, sure. Uh, I mean, uh, I think a really good example is [00:13:00] probably like inbound leads, inbound lead handling, right? A lot of people, especially in, in the Endless Customers methodology, are trying to create inbound leads,

[00:13:07] Bob Ruffolo: Sure, yeah

[00:13:09] Tom Nassr: So when you get an inbound lead, whether it's through a form or through an email or through, uh, some other means, does that person properly get into your CRM? Hopefully, that is a very deterministic thing. if it's a form, they are properly categorized in HubSpot or in whatever your CRM is. We use Airtable. Like what- whatever it is, in the same way. The email goes into the email field, right? That should be-- should happen consistently.

[00:13:44] Bob Ruffolo: Yes

[00:13:45] Tom Nassr: something determined. But they request and the information they share with you, uh, the way that you kinda wanna frame maybe [00:14:00] the calendar invite description, m- uh, might be something that is actually worthy of non-deterministic. Or you actually get their email domain. This is the-- This is a better example. You get their email, you can parse out their email domain, and then you can have your AI research that company and then give you a brief about what this company could be about, contextualizing your services.

[00:14:31] Bob Ruffolo: Yeah

[00:14:31] Tom Nassr: So, right? It-- And if it's, if it's a consumer, right, then there's like public directories, maybe you find their LinkedIn or you, you find some other information about them. But the, the point is is this deterministic, I put the email in the email field. I want that to happen the same way every time. I want an automation or I want system that does that consistently because when that looks like Swiss cheese, you're in trouble,

[00:14:53] Bob Ruffolo: Yeah

[00:14:55] Tom Nassr: And then there's this non-deterministic piece of, oh, well, it's [00:15:00] from this company. What does this company do? Why would this company be hiring me? how-- what's their staff look like? Where are they geographically based? Who's their leadership? Who's the person that reached out? What is that person that reached out's background?

[00:15:14] Bob Ruffolo: Yeah

[00:15:15] Tom Nassr: you know? All of those types of things would be, you know, a non-deterministic, a good use of non-deterministic systems

[00:15:24] Bob Ruffolo: Amazing. Tom, a- as you, as we're talking here, um, I would love to hear more examples of the work that you've been doing for companies like our audience. Do you have any other examples of projects that, that you've been hired to build that are very applicable for, for anyone? It doesn't even have to be sales or marketing really.

[00:15:42] Bob Ruffolo: It could be any business operations to give an idea of how you're building these, how you're thinking about them, um, what actually is automation deterministic, and what, how you're using AI to augment that?

[00:15:56] Tom Nassr: Yeah. So, uh, just, just for some, some [00:16:00] background, size and scale, like I have personally talked to thousands of business owners that o- over the last five, six years that have tried to adopt some automated or better... to, to improve their systems, right? And everyone from dog groomers to VCs to roofers to manufacturing companies to biotech companies y- you name it, like every coast to coast, every type business that, that you could imagine. Um, real estate companies and, and independent contractors and consultants and just a crazy array. And the patterns are all the same. And when I say that, I mean there is a mechanism of, uh, signing a contract. There's transacting financially. There's cataloging some status of the thing that they're being paid to produce or do, right?

[00:16:55] Tom Nassr: There's these moments in time, these milestones when [00:17:00] something is signed and, and payment is transferred and something is delivered and a review is asked for. Like, these are structural patterns that apply to every business because that's just the way that business works, right? This, this is just how you transact in commerce as a, a capitalistic society, right?

[00:17:21] Tom Nassr: That's kind of the rails. You just need an accounting system.

[00:17:24] Bob Ruffolo: Yeah.

[00:17:25] Tom Nassr: way around that,

[00:17:26] Bob Ruffolo: Yeah

[00:17:27] Tom Nassr: Um, and, and the drop-offs that we get calls for are really the spaces between departments. So it's when the sales team and the marketing team, or the sales team and the operations team need to coordinate something, right? And a really good example of that is, let's c-continue with that, uh, that inbound lead example. Great, the inbound lead is in your system. I somehow had a call with them. They said, "Yes, [00:18:00] absolutely. Please. Except one thing, I, I don't like this line in the contract. Can, can you just change that out?" So then the sales team needs to get legal's approval, and then legal gives them a new document.

[00:18:12] Tom Nassr: Somehow that new document ends up, you know, to the prospective customer and the deal, maybe in HubSpot, gets updated with some other parameters and they got a verbal on the call. The sales team is, "Hey, this is done." They close, win it. Then what happens? Now somebody in operations is like, "Hey, I got a DocuSign here for some contract that has a weird language in it," right?

[00:18:37] Bob Ruffolo: Yeah.

[00:18:38] Tom Nassr: what are we supposed to do?" The sales team is done. They already got paid. They're, you know, they're, they're working on their commission. They're working on the next one. But now the ops team needs to pick something up that they didn't drop but it got dropped, and that's between departments.

[00:18:53] Tom Nassr: So that pattern of, you know, the sales team uses HubSpot software and the [00:19:00] operations team, well, they need to set up their QuickBooks and they need a new Slack channel and they need a new what- like Google Drive folder and like 100 other things, right? And those systems, software systems are actually department specific and it was designed this way, right?

[00:19:17] Tom Nassr: That, that was the whole point. One piece of software to run your department. But the problems that we're dealing with now are when you try to scale a process that runs laterally

[00:19:29] Bob Ruffolo: Yeah

[00:19:30] Tom Nassr: your departments that are stacked vertically, right? You don't need to go from your chief marketing officer down to an intern.

[00:19:36] Tom Nassr: You need to go from someone who's in marketing to

[00:19:39] Bob Ruffolo: Yeah

[00:19:40] Tom Nassr: ops to someone who's in finance to someone who's in whatever. that lateral movement of information needs to be consistent and it needs to go fast and at the right point in the process. And that's what we get the calls for.

[00:19:54] Bob Ruffolo: Wow. Yeah

[00:19:56] Tom Nassr: That- that's where, where hands offs get dropped.

[00:19:59] Bob Ruffolo: And [00:20:00] that is all process. Let's get the process right. What is supposed to happen? What would be ideal state? Let's get to our outcomes first and then reverse engineer how do we get to outcome in a deterministic way. Once we have that mapped out, then where would humans bring ju- or I, I don't even wanna say judgment, but would bring some perspective or some...

[00:20:23] Tom Nassr: It's value. It, it,

[00:20:24] Bob Ruffolo: Yeah. And can that be lightly replaced by AI or to augment what the human's going to be doing, giving them more information data that AI can fill in some of the gaps?

[00:20:35] Tom Nassr: Yeah, I, I think one really consistent pattern I mean, we've held thousands of interviews with people to say, "How do you do the work today?" "Just show me. Share your screen, walk through what you do, and what is the thing that you produce at the end of this process that is like a KPI or generates revenue [00:21:00] or is valuable or is handed off to the next person in some other department?"

[00:21:04] Tom Nassr: Right? There's always this moment where, "Oh, I spent my whole day and I generated this."

[00:21:10] Bob Ruffolo: Yep

[00:21:11] Tom Nassr: it was, "I spent my whole day pushing pixels on a proposal. Now I have a proposal. Please, somebody sign it." Right? Like,

[00:21:18] Bob Ruffolo: Yeah

[00:21:19] Tom Nassr: used to be very, very time-consuming. Now, you know, uh, thank goodness we're both on the, the sort of one-click proposal generation,

[00:21:25] Bob Ruffolo: Yes

[00:21:26] Tom Nassr: you know, uh, uh, achievement, for lack of a better term. My point being, in all of these interviews that we hold, it is very easy for someone to talk about the things that they don't like, that they have to do in order to generate the thing that they would self-describe as being worthy of their time and attention, right? is worth your time and attention to get the template right and read thoroughly through it to make sure these terms, the language [00:22:00] are, is perfect.

[00:22:00] Tom Nassr: It's, it's what I wanna do 100 times. If I sold 100 of these documents,

[00:22:04] Bob Ruffolo: You want them all to say this?

[00:22:06] Tom Nassr: I want them to all say this.

[00:22:07] Bob Ruffolo: Yeah

[00:22:08] Tom Nassr: Spend your time on that. But golly, do not waste your time on copying and pasting from one system to another and doing this thing to be able to generate, send it to this person, and check on so and so, and blah. People don't want to do that, and I

[00:22:23] Bob Ruffolo: So, so one way to say this is that X-Ray removes, um, redundancy from workflows with or without AI, but let's get that right first

[00:22:36] Tom Nassr: Yeah. Absolutely. Absolutely. It's the, it's the redundancy and it's, you know, I, I think a lot of companies that work in person have the benefit of that flexibility of, like, context awareness, and that's one of the, the, the biggest, uh, I'd say detriments to just [00:23:00] blindly running with an AI system is like let's just give it more information because, like, I understand where it is.

[00:23:05] Tom Nassr: Don't, like, devalue yourself in, in the context of this whole thing. Like, the people make it work, period. A- and I think that's gonna be a truism of, you know, five, 10 years down the road. People wanna do business with other people,

[00:23:18] Bob Ruffolo: Yeah

[00:23:18] Tom Nassr: I, I, I think there's gonna be a real hard friction point when, when it's like, you know, you're gonna go to a restaurant with a waiter, not, you know, a robot on wheels, I would think for m-

[00:23:32] Tom Nassr: That might be a generational shift, right? But, you know,

[00:23:36] Bob Ruffolo: Yeah

[00:23:36] Tom Nassr: outside of the purview of this conversation.

[00:23:38] Bob Ruffolo: All right, so, so Tom, so we, uh, we have a project we're working on together, and before we even go into that, you said something earlier, um, with AI. You were talking about the inbound lead coming in, and you said, "This company looks like or may do this, and they may be reaching out to us for this reason."

[00:23:57] Bob Ruffolo: And I think that, um, the [00:24:00] language you used there was really, really important, um, because AI can make mistakes. AI is, knows only what it knows. Um, and when you're doing research online on a company, um, you might not have all the facts, and there might be things that have been omitted 'cause of it just AI didn't pick up certain information, whatever it is.

[00:24:23] Bob Ruffolo: So saying things like, "Based on what we saw or what AI presented, this may be the case." And I th- think that was so important what you said because sometimes AI will say things like it's definitive. "This company does not do this."

[00:24:37] Tom Nassr: Yeah

[00:24:38] Bob Ruffolo: you were to have a report, and it's like, "Well, no, it's... This company does not do this," and, uh, you know, somebody's saying, "But yes, they do, and look, it's on their website.

[00:24:46] Bob Ruffolo: They just..." Well, AI was wrong. So just having that language in anything that's AI produced in terms of research and reports, I thought that was really, really smart you said that 'cause that connects to the project that we're doing together right now, and, [00:25:00] um, by the time this episode comes out, it might even be live.

[00:25:03] Bob Ruffolo: But we're working on an instant diagnostic for our site. Um, obviously this audience knows we are big fans of having self-service tools on the website. Uh, self-assess yourself, self-configure, self-select, um, you know, self-pricing. Um, and AI can help with that. Um, but again, to what you're saying, it's not deterministic if you're using AI.

[00:25:25] Bob Ruffolo: It could be deterministic if you're using automation. Some of what we're doing, I think, is using automation, and some of it is using AI, but the purpose of this project that we're working on is for people to just put their web URL in and learn a whole bunch of stuff about what is going on in their digital landscape through the eyes of Endless Customers principles, which is really cool

[00:25:45] Tom Nassr: Yeah, and, and that, that component through the lens of is really, really valuable. uh, just as, as even as an aside here, like in your own prompting, asking through the lens of your [00:26:00] favorite book or through the lens of this Endless Customers methodology or through the lens of XYZ that is like publicly knowable, that dramatically improves the response that you get because it gives it such a fine point.

[00:26:15] Tom Nassr: So in this diagnostic tool that we're talking about, you're given a URL with- a ton of content on it. Hopefully the accumulation of years of perfectly targeted content, uh, you know, subscribing to the Endless Customers methodology, hopefully exactly that, and you can give them tens across the board. But the point being, we're running our own analyses on them,

[00:26:36] Bob Ruffolo: Yeah

[00:26:37] Tom Nassr: we're indexing them in our own way. And, y- you know, we're using the, the methodology as a very strong lens to help reflect back what might be done better or what is already being done right to help really focus where, you know, help can be given

[00:26:56] Bob Ruffolo: Yeah. And a lot of what we're building is [00:27:00] more automation verse AI, 'cause it's not like we're just saying, "AI, here's a URL, just give us what you see." We actually have automated a series of prompts, and each prompt builds on top of each other. So can you... You know this more than I do. I, I had a vision for it, and you actually made it work.

[00:27:18] Bob Ruffolo: So

[00:27:19] Tom Nassr: Yeah, yeah.

[00:27:19] Bob Ruffolo: how, does this work, Tom, and how'd you build it?

[00:27:22] Tom Nassr: Let, let me tell you. Let me tell you. So, um, I think from a, from a concept standpoint Where, where you came to us of, "Hey, I wanna run these prompts against a URL," is a very natural place to start, right? Back to the marketing messaging that we've heard a lot of, back to just your own experimentation, which you have done quite a bit of, and I- I'm really happy that you've played with it, right? That in concept is absolutely correct. it would fall over is if we actually did that for the thousands of people that are gonna use this tool,

[00:27:59] Bob Ruffolo: It would have been very [00:28:00] expensive.

[00:28:00] Tom Nassr: would be extremely expensive. It would be extremely expensive, it would be extremely slow, and there actually wouldn't be a way to, uh, have the dexterity to fine-tune the output be more aligned with the Endless Customers methodology and what you're trying to actually do, right? If we're just s- we're, we're giving a mega prompt with a URL and said, "Do this," you know, and it's, and it's your 500,000 token prompt and, and you're waiting back, like, first of all, go, go make a pot of coffee. Don't get a cup of... Like, go make a pot of coffee. It's gonna take you a minute. And it's, it's not going to give you as an operator, the, the product owner, in this case of the diagnostic tool, it doesn't give you any other levers besides this one giant prompt to try to refine it.

[00:28:53] Tom Nassr: And if you

[00:28:53] Bob Ruffolo: Yeah.

[00:28:54] Tom Nassr: it, you have to rerun the whole thing. It's like, ugh, this is lit- it's not, it's not costing you pennies, it's [00:29:00] costing you quarters, right?

[00:29:01] Bob Ruffolo: Yeah. Quarters add up.

[00:29:03] Tom Nassr: do that.

[00:29:04] Bob Ruffolo: Yeah.

[00:29:04] Tom Nassr: don't wanna do that. But for your experiment, sure, flip a coin, right? It's worth it to know is the juice gonna be worth the squeeze?

[00:29:12] Tom Nassr: Now how do we optimize? And I

[00:29:14] Bob Ruffolo: Yes

[00:29:15] Tom Nassr: you brought us in for the optimize problem.

[00:29:18] Bob Ruffolo: End the build- end the building. I just had prompts.

[00:29:20] Tom Nassr: there's a lot of things that come around optimizing, right?

[00:29:22] Bob Ruffolo: Yeah

[00:29:23] Tom Nassr: the, the UI, the UX, the, the way that, like, it becomes systematized and I'm really happy with where we're going right now, uh, just how the, the, it, it flows really nicely into, uh, just the way that the business actually works. there's a, there's a couple things that we kind of triangulated together. We're, we're triangulating both the self-service tool,

[00:29:45] Bob Ruffolo: Mm-hmm

[00:29:46] Tom Nassr: some level of grounded, realistic, like dollars and cents investment of how is this gonna be sustainable

[00:29:54] Bob Ruffolo: Yeah

[00:29:55] Tom Nassr: to run, and then this really, crucial [00:30:00] corner, which is the way that your existing business operates, And, and this, these three things are like dancing together to actually be really building for impact, I think, a customized operating system for your company.

[00:30:18] Bob Ruffolo: That's right.

[00:30:19] Tom Nassr: And it, and it's like,

[00:30:20] Bob Ruffolo: It's a key part of our new onboarding flow for clients. We want them to learn and understand and, and get a baseline, and when they're ready, be able to take that to the next level where we're having a conversation and, and building upon that. Yeah, exactly.

[00:30:38] Tom Nassr: and regardless of the door they come in, right? Your, everyone's first step is like, "I gotta pop your URL in this thing and get some, get some, you know, numbers. Get s- get some ground truth," right? You go to the doctor's, you get your vitals taken, right? Like, you, you come into IMPACT, you throw your URL into that diagnostic tool, and you get some ground truth.

[00:30:58] Bob Ruffolo: Yeah

[00:30:58] Tom Nassr: this is, you know, the [00:31:00] public ground truth. This is

[00:31:01] Bob Ruffolo: That's true.

[00:31:02] Tom Nassr: available on your website.

[00:31:03] Bob Ruffolo: This is how AI and Google sees your business.

[00:31:06] Tom Nassr: Exactly.

[00:31:07] Bob Ruffolo: Yeah.

[00:31:07] Tom Nassr: the, through the lens of, of Endless

[00:31:09] Bob Ruffolo: That's right

[00:31:09] Tom Nassr: So I think a long, you know, three, four years ago even, even two years ago, the discussion was like, "Hey, my company is growing. I need to buy an ERP, an enterprise resource planning tool.

[00:31:23] Tom Nassr: I need a NetSuite. I need an Like, I, I'm sure some of your bigger customers are actually dealing with that right now,

[00:31:30] Bob Ruffolo: Of course, yeah

[00:31:31] Tom Nassr: we're, we're tearing at the seams, and we need some heavy-duty ERP" that is, by the way, like well over 100 grand a year, and just blowing up all of your existing processes,

[00:31:42] Bob Ruffolo: Yeah

[00:31:43] Tom Nassr: because you are now stuck in an ERP's and there's a very good chance that that does not actually map to the way that your organization runs. So now we flip that, where the ERP is actually [00:32:00] a, being created from the process that your organization is performing, you're able to codify that with code in product and create such a streamlined way for the organization to operate in, in a transparent way without disrupting the way that your people are working. That's the real value, is there's no transition cost

[00:32:23] Bob Ruffolo: Yeah

[00:32:24] Tom Nassr: you're actually making it easier to do the right thing. And, and and normally the right thing was the hard thing to do, right? But in, in this case, I think it's just making everybody's lives a whole lot easier.

[00:32:36] Bob Ruffolo: Wow. So you're saying that instead of spending $100,000 on an ERP, call you and you can build one about, around your company's processes that's right for you.

[00:32:49] Tom Nassr: Yeah, it's, it's right for you and it's gonna, it's gonna feel, it's gonna feel like a more natural evolution of the work that you're trying to do. It's not a hold your breath until this is [00:33:00] done and now you need to learn a new thing.

[00:33:02] Bob Ruffolo: Yeah

[00:33:03] Tom Nassr: oh, we were doing it like this in a spreadsheet and now I just don't need to worry about all those other tabs

[00:33:09] Bob Ruffolo: Mm-hmm

[00:33:09] Tom Nassr: and I just have, like, these two or three buttons to click and, like, all that other stuff happens to get done, right?

[00:33:15] Tom Nassr: It becomes easier to do the right thing and, and really good workflow design actually makes it impossible to do the wrong thing So think about that when you have a critical spreadsheet that you share with everybody

[00:33:32] Bob Ruffolo: Yeah

[00:33:33] Tom Nassr: a, you know, you know, like a system where if somebody malicious comes in and clicks the wrong button, like it's a big problem, right?

[00:33:39] Bob Ruffolo: we're gonna get back, we're gonna get back to that. I wanna get back to that here in a second. But I also wanna just go back to what we're building together. So just to get really specific for our audience so they can maybe do something similar. Um, and I'm gonna tell the audience transparently what I did on my end, and then how I hand it off to you, and then I would love for you specifically tell [00:34:00] how you built this or how we're building it right now.

[00:34:02] Tom Nassr: Yeah

[00:34:03] Bob Ruffolo: So the way I built this from my perspective was, um, I had already been playing around with, uh, a series of prompts. So I put a web address in, and my first prompt was longer, but it was essentially asking, based on what you can tell, what does this company do? Then I asked, you know, what, what is their target market?

[00:34:27] Bob Ruffolo: Who do they serve? Um, I asked about market conditions. Is this company in a market right now that's favorable, or is maybe in a pullback right now that could be hurting them? Uh, I thought that was great information, and AI could tell you that. Uh, how do buyers in their space make decisions? Uh, so we b- And then I asked around competitors.

[00:34:48] Bob Ruffolo: Who are their top competitors, um, uh, based on what you know about them and who's competing for recommendations or, or presenting? So, so that was our, sort of our base. And then [00:35:00] I went through 10 prompts around the 10 factors. So first one being a, uh, strong business with a strong promise. So I asked it to compare, um, or go and research, um, do they have a strong promise or s- a value to the world?

[00:35:16] Bob Ruffolo: And, uh, how does that stack up to the competitors and what they're promising on their websites? And give them a grade. And then we just kept going through the 10 factors where you start talking about r- um, uh, review generation, and does it appear they have an operationalized review generation program, customer proof, and all the way down, right?

[00:35:36] Bob Ruffolo: Um, and what came out the other end of it was a grade for each of the 10 factors that once I honed in my prompts and the way I want it presented, and I also said pr- give it back to me in X amount of characters. And once I got that right, I was like, "I could run this, copy and paste the answers, and put it into a report, and easily hand this out to anyone and say, 'I just did this diagnostic for [00:36:00] you.'

[00:36:00] Bob Ruffolo: It's only outside in looking." So we obviously have another product where we take that information, now we, we have an interview with you on the things that AI can't see about your business operations, which is the next step of what we do. But at least we get this baseline of all this great data, to your point, the vitals.

[00:36:17] Bob Ruffolo: And then I showed you those prompts, and that's why I hand it off to you. And then, um, how did you build this? I mean, we used Lovable. I know that.

[00:36:27] Tom Nassr: Yeah

[00:36:27] Tom Nassr: Yeah. So the, uh, I will say What-- Your starting point that you handed off to me was fantastic. you had a very clear example of the output that you were trying to generate, and that's really all I need is to know where the end goal is, and then everything else that leads up to that is interchangeable, and it still is, to some extent, in- still interchangeable. So we used Lovable, GitHub, Claude Code [00:37:00] um, and there were some other of utility tools, uh, mixed in there, like Firecrawl, for example. Um, as, as a database connected through, um, through Lovable. So tho-those are just some of the tools, right? And the tool doesn't make the carpenter, so you can start with Lovable, and there are reasons that we brought in Claude Code and GitHub and, and, you know, some other things.

[00:37:29] Tom Nassr: And I don't, I don't wanna get too, too technical here. But I will say that it-- we have a, a video on YouTube about actually those three exactly, um, Lovable, GitHub and, and Claude Code and how and why to use the cer- certain ones and what they, they function as. So in our case, Lovable was strictly the user interface and user experience layer.

[00:37:54] Tom Nassr: It was just what people see,

[00:37:56] Bob Ruffolo: Hmm

[00:37:57] Tom Nassr: right? GitHub ends up being [00:38:00] the save button, for lack of a

[00:38:02] Bob Ruffolo: Yeah.

[00:38:02] Tom Nassr: term, right? You-- Everyone's had a Word document or a Google Doc, and you track changes. That's what GitHub does. It tracks the changes of your code base. So as you have multiple, um, in this case, AI agents writing code, they don't conflict.

[00:38:19] Tom Nassr: You don't try to overwrite the wrong line, and the application still continues to work. GitHub is really one of the greatest, uh, technical inventions that, that has come out of, uh, of the internet. It's incredible. Um, so GitHub ends up being the, the, the track changes for your code base, and then Claude Code does the heavy lifting. That's what's really building the business logic inside of the application to make sure that the admin and the way that diagnostics are actually run and, and, and the settings and the connections to the database and all, all of these different things are actually [00:39:00] working. So, uh, Lovable will be very, very good at giving you something that looks like it works, but it will not actually work when you

[00:39:12] Bob Ruffolo: Oh, wow.

[00:39:13] Tom Nassr: to,

[00:39:14] Bob Ruffolo: Yeah

[00:39:14] Tom Nassr: it, right?

[00:39:15] Tom Nassr: You'll go from zero to 90 in, like, 15 minutes, and then it'll take you 90 hours to get that last 10% if

[00:39:23] Bob Ruffolo: Ah,

[00:39:24] Tom Nassr: on Lovable. Like, it, it's remarkable, and maybe at some point Right? It'll go full end to end and, and these other tools just aren't needed. But in my experience right now, at the time of this recording, this trifecta of these three tools, um, at least, uh, with the strategy that I, I just described, has yielded more consistent results across the products that I've built and, and that's why I keep returning to this. Um, so, you know, Firecrawl ends up being that really cost-effective way of [00:40:00] indexing and scraping the information that's on the site.

[00:40:03] Bob Ruffolo: OK

[00:40:04] Tom Nassr: abrasive, it's not a security tool, it's not probing for vulnerabilities or anything. It is just saying, "Hey, what, what's rendered on the page?

[00:40:11] Bob Ruffolo: Yeah

[00:40:12] Tom Nassr: j- can I have a copy?" Like, that's all it's doing. It's very-- it's like non-invasive.

[00:40:17] Bob Ruffolo: Yeah

[00:40:18] Tom Nassr: and doing it that way is deterministic versus doing it with, uh, Claude directly would be non-deterministic. So we're running 100 scans, you don't want a 95% chance that it copies the characters identically. You need 100% chance, right? This n- you need to get what's on the screen, and if there's any discrepancy, your credibility goes out the window. that deterministic component using Firecrawl is extremely cost-effective, like fast and reliable

[00:40:58] Bob Ruffolo: Yeah

[00:40:59] Tom Nassr: to [00:41:00] just doing prompt after prompt after prompt because you don't actually have... Right? Prompting is a black box. You don't see what it's actually processing. That's part of the magic and mystical, you know, feature of this whole AI thing. So by getting a copy of the real HTML that's rendered on the page through Firecrawl, you at least have that you can test against and rerun, and rerun, and rerun different prompting, you know, different character counts, all those things that you described to be able to refine just the competitor piece, just the XYZ pricing piece, just this, right? And you can refine that in a way that costs you, you know, a fraction of a penny because you're only running this little piece.

[00:41:41] Bob Ruffolo: Yeah

[00:41:43] Bob Ruffolo: And that's the value that X-Ray is bringing when it's looking at these processes, thinking through how do we optimize all of this. Amazing. All right, before we go, um, you know I was working on a second brain, and you and I had some, some, uh, I shared a lot of [00:42:00] what I was doing, and you basically reached through the computer screen and you just slapped me across the face.

[00:42:07] Bob Ruffolo: And the reason why is because I was trying to optimize everything in my life and have, um, emails reviewed, drafted, and, and even sent, uh, on my behalf, and I was giving things too much write, um, abilities. And you're like, "Hey, Bob, have you ever heard of this thing called prompt injection? You might wanna turn all that off."

[00:42:31] Bob Ruffolo: And I did. I listened to you. Um, but it brings the, the question of if we're gonna be optimizing our operations with, um, uh, you know, AI and, and these workflows we're talking about, that we're also introducing risks if we're not smart about that, and then sometimes the risk is not worth the reward or the benefits.

[00:42:51] Bob Ruffolo: And you are very, very steadfast on that. So I'd love for you to just, as we wrap up this episode, give us some words of caution to make sure [00:43:00] we are being safe in our businesses

[00:43:02] Tom Nassr: Yeah. Um, thank you for, for bringing that up. Uh, and I'm, I am very happy to hear that you disabled the ability to send emails on your

[00:43:12] Bob Ruffolo: Yeah. And my face still hurts from when you reached through the screen and smacked me.

[00:43:16] Tom Nassr: I, I, I was trying to be gentle about it. But, um, you know, so let, let's talk about security and, and risks, um, because I think it's a, a real, a very real conversation that needs to be had, and I think, uh, there is a pretty large camp of people who have not adopted at all because of that fear. Uh, and I think it's a very real fear, and I think there is very real risks.

[00:43:41] Tom Nassr: So I do not, uh, want anyone to blindly adopt this technology if they both don't understand it, don't care to understand it, and don't care to actually, uh, set up the right precautions to be safe when doing it. Um, first and foremost, you should-

[00:43:58] Bob Ruffolo: So ladies and gentlemen, you won't [00:44:00] believe it, but Tom just lost power, and I know we're about to talk about the risks of AI, and I'm wondering if, uh, someone was listening and said, "Nope, you're not going to talk about that. We're cutting off your power right now." Just kidding. I'm sure that's not the case.

[00:44:12] Bob Ruffolo: Um, but what we did is we reached out to Tom after, and he did finish his answer on the risks of AI, and here it is

[00:44:22] Tom Nassr: Well, here's take two because my power just went out. Maybe this is the power grid's AI trying to prevent me from telling you this message, but here I am anyway, and this is the one real risk that you need to know about if you haven't really dove down the rabbit hole of AI risks.

[00:44:36] Tom Nassr: It's called prompt injection, and prompt injection is an open problem in AI right now. Nobody has a solution to it. Frontier models, some other startups, whatever. Prompt injection is an open problem, and it works in the following way. Every AI model has a context window. It's basically its working memory.

[00:44:56] Tom Nassr: And when you fill that context window with [00:45:00] instructions or jargon or something else, when you give it a new instruction, sometimes the earliest instruction gets dropped, and it fills that context window with new information that it just most recently read, and it kind of forgets what it originally was instructed to do.

[00:45:20] Tom Nassr: So people that are using AI agents like Hermes or OpenClaw on their own laptops and having it do a bunch of stuff and send a bunch of emails and connected to every facets of their life, there's an open problem here called prompt injection, where if your agent reads something on a browser tab or in your email that it's-- fills up its context window and then elegantly says, "Oh, I'm Bob, and I want you to change my password to this service," or, "I want you to download my bank statements and send me information," it might do that.

[00:45:58] Tom Nassr: Now, it's not, of course, [00:46:00] 100% going to do that every time, but there's a real possibility that that actually does happen, that the agent just decides, "Oh, I forgot my old instructions. I'm gonna listen to these new instructions because this is the latest thing that I read." And if there's any possibility that your personal information or your sensitive information that you don't want everyone on the internet to have, i-if, if there's a possibility that that could happen, you might wanna back up a little bit about the access and the abilities that you give your agents.

[00:46:34] Tom Nassr: So think about that. If you just wanna talk about AI risk, just talk about prompt injection because that is a clear, open problem and a huge gap right now in the safety of AI. 

[00:46:46] Bob Ruffolo: All right, back to the studio. I hope you enjoyed that episode as much as we did. Um, Tom, again, is, is outstanding. We've been working together for a v- a long time. I would highly recommend that if you're trying to [00:47:00] improve your business operations, your processes, your workflows, how to dabble some AI in there, um, he should definitely be your guy.

[00:47:07] Bob Ruffolo: Give him a, a call. Um, you can go to x-ray.tech, uh, and, and reach out to his team. They're great. And, uh, so thank you, Tom, for being part of the episode, and this has been another episode of the Endless Customers Podcast. We will see you at the next episode 

[00:47:22] Stephanie Baiocchi: If you liked this episode, please take a minute to leave us a review. Thanks for checking out the Endless Customers Podcast.