Steven Roach (2:37)

I think there's a massive drop off with a lack of human intervention in a lot of these interactions or models that are being built out. So there's always a little bit of mistakes with the front end, the back end, overall security layers. If you are willing to risk 1% of the, you know, mistakes out there, you're more than likely going to have a larger percentage of damages that are being done to your organization without human intervention. An example of that is a tech entrepreneur. The story is very well known, written by multiple outlets out there. Tech entrepreneur, um, actually integrated ripple it AI agent into his overall development workflow. It started to do a little bit, you know, craziness happening. Some authorized unauthorized charges were happening, some crazy changes were happening to his database, actually removed his database, very live database, without, without any human interaction. So if you're allowing your overall AI agents to have the read ride, overall delete access without actually providing those guardrails, you're asking for a world of hurt. So providing those guardrails and human interventions at every step needs to be worked into your overall workflow. And I think that's a very damaging thing that know organizations are just saying we're going to build it, we're going to press go and then see what happens. Stages just like your regular, you know, development QA staging and then prod. Those stages still need to apply to AI intervention as well. So involving the human into that, then you're good.

Benjamin Shapiro (4:36)

So Stephen, what I'm hearing is there's a couple of different places where the big AI errors can happen. You mentioned sort of in the back end infrastructure, people are using agents that are deleting stuff and deleting databases for them, even making fraudulent charges. And I know that you mentioned replit, which is an AI agent or an AI builder tool. Obviously there's the horror stories of oh no, AI ate my lunch. What are the common mistakes? Is it people that are, you know, using AI to clean their database and they're getting bad data? Is it the front end development that they're using it for? Where is the most common breakage?

Steven Roach (5:21)

I think all of the above. Some of the success stories out there is automating certain workflows that are just heavily manual. When your project manager just needs to get through a process that can be automated based on. We're seeing a lot of note takers nowadays. We don't need manual note takers anymore. There are certain keywords or certain phrases that we may want to write down for our future reference. But every video streaming or video conferencing tool nowadays has their AI integration for note taking, which can be incredibly powerful. But how do you translate that? What are the next steps? Can you then say take all these notes and then assign assigned a task coming out of a meeting, you know, into jira? There's an automation process that can be orchestrated, you know, using AI if needed. And again, that should be in the hands of your, your subject matter, subject matter experts to be able to hand out certain tasks. That's, you know, AI is, you know, probably automating across the board human intervention. There's an automation piece of that, but there's human intervention as well.

Benjamin Shapiro (6:36)

So let's talk a little bit about that human intervention piece because I think that's where people are getting into problems. Using your replit example, somebody integrated repl it and said have at my code base delete my database. Sure, you can delete whatever you want. Obviously that is a bad idea. Yeah, right. You can't just let AI run wild and let it iterate on its previous mistakes. So where do you think about integrating human in the loop and where's human judgment important? What are the checkpoints?

Steven Roach (7:04)

So human in the loop should be fully integrated in your not only inputs, but in your outputs as well. And that very well could be from your automation of prompts and then really looking at a lot of your outputs. I think there's a lot of mishaps like, all right, certain use cases are great, but if say your UX team goes in and uses that same agent, their output may be completely different and it may not be company approved. So you need to test out agent for every single use case in your organization if that department is going to be using that specific agent. So monitoring every aspect of use cases from front to back across the board needs to be implemented. And again, all stages need to be implemented from human in the loop, making sure that there's check marks in every area.

Benjamin Shapiro (8:05)

I don't know why this is the metaphor that sticks into my head, but it needs to be a human that puts the stickers on the avocado that says whether they're ripe or not. Right. Like there's all of the sorting and processing that goes through and I'm sure there's a machine that does color analysis and firmness analysis and machines, machine machines and then somebody needs to be there watching them go down the assembly line before they're in the box, just giving the final stamp of approval and it's, it's the quality assurance before something is publicly shift. Honestly, we've made this mistake a bunch over the last couple of years, mostly with like copywriting and transcripts thinking like, well, if we write our prompts and keep improving and improving, we won't need. No, you just need somebody to check what's going out the door first. Like make sure your kids have shoes on before you send them to School. It's common at whatever phase of life you're in. You all know that there's the one check that you need to do before you leave and you have to do it yourself and it's before something gets shipped. To me, that's the most important part of putting some sort of human in the loop. Are marketers specifically just forgetting this? Are we just saying, look, we don't have to do the work because the intoxication of automation is there?

Steven Roach (9:21)

I think there's a massive push from leaders that are demanding automation, AI be integrated into each workflow. And there's a lack of knowledge or understanding of what is to be expected of the outputs of every single prompt. And to be frank, we probably have a massive resource issue when it comes to individuals that understand it and want to make sure that the quality check is there. So if you are individual that's in marketing and you're only focused in on just I want the export to be xyz and you're not double checking that like who's going to be the person within your organization to double check that?

Benjamin Shapiro (10:19)

Yeah. It's funny because from a leadership perspective, right. If I put my, my CEO or, or sort of head of content hat on for our business, I want to lower our margins. The more we can automate, the less I have to pay, the better our margins are. So take all the humans out of the loop. Let's automate, automate, automate, automate. And I never have to pay anything. And now all of a sudden my business is performing better. And I think that we've all forgot that there's a quality bar that balances that out. And whether it hits you now or over time, obvious mistakes. Replit deleted my database. Oh, you know we shipped bad copy for a social post. Right. Whatever it is like that human in the loop, that step where a person is needed. If you lower the quality so far down where it's all automation, it comes back to bite you. And we're trading this margin versus quality. How do you find the right balance there of. I've automated everything I can, but there are these checks which I will not cave on because that's what makes sure that we're getting a better quality output using artificial intelligence.

Steven Roach (11:24)

And I think that's. That has to be done. That process has to be done way in advance of you implementing a AI infrastructure. And it needs to be continuous, you know, reviewing those deterministic or non deterministic, non deterministic workflows across the board. You, you need to make sure. That we like from A to Z. This process needs to be X, Y, Z. We're not breaking that. That is our policy. That is how bank of America operates. That's how Chase operates or Google, whoever. Those are your steps that need to be a part of this across the board. And we are not, you know, substituting that for anything. There's other processes, like a brand new project needs to come into play. Same thing, non deterministic. We want to, you know, be a little bit more free with that, like involve AI into those processes, but don't have it completely run that process across the board, you know, so I think.

Benjamin Shapiro (12:23)

That'S a great framework. Deterministic versus non deterministic. And we're going to get into a little bit tech stack talking. But you know, as I've been thinking about building our workflows and our automation, it's like, all right, where do we need to stop the workflow? Have a human check it. Check off a box, you know, put the stamp on the avocado, I guess it's a sticker that you put on an avocado and then let it roll down into the box.

Steven Roach (12:49)

Yeah.

Benjamin Shapiro (12:50)

If it is a, this is a yes or no question. It is a deterministic output. You probably don't need a human. Yeah, right. If it's a subjective, non deterministic, then you probably need somebody checking what the output is.

Steven Roach (13:02)

Exactly. Across the board. And you can involve your communication applications right now. Now Slack teams, you know, I want this process to be approved. It does not move from this step unless it's approved and that it's, it's routing it to the right people.

Benjamin Shapiro (13:19)

All right. I've been trying to get into the like, let's talk strategic and. But I, all I want to talk to you about is the nerd stuff and about the actual building because it's all I've been doing late lately. We're basically taking our production infrastructure and getting it ready for third party producers to use. The way that we create the Martech podcast using our workflows and our automation and the AI integrations and a big thing that I've been doing, you mentioned Slack of the workflows is there is a trigger in Slack where you can basically say, okay, is this approved or rejected? So when we get to a certain stage you can qa. All right, hey, you know this publish. This post was published. Was it published? Can you verify it looks good or bad? And then a human comes in and gets a notification in Slack and says yes or no and it Finishes the workflow or moves it along. Yeah, I've been that to me unlocked everything. Right. Where's the human in the loop? How do you integrate it? It's like we get to a certain point, you have a check and then you check a box and then the AI and the workflow can continue.

Steven Roach (14:20)

Yes.

Benjamin Shapiro (14:21)

Now I'm in like nerd zone. Talk to me about the tech stack. Talk to me about the infrastructure you mentioned. Replit already. Like how are you thinking about building these automations and workflow at the sort of enterprise scale? What's the toolkit?

Steven Roach (14:39)

I think it does vary your overall infrastructure. Current infrastructure either be AWS or gcp. Those still matter. Those core infrastructures, data layers, you know, your, your AI layer and overall governance, they still matter. I think where it's really important is to involve the newer, newer style of thinking. Expanding that overall AI layer into that infrastructure. Excuse me, that interface layer, the orchestration, memory, tools and actions and then infrastructure as well, I think. So you need to do an evaluation of what tools or what models actually fit your organization. If it's ChatGPT because you like the creative writing or you like Claude because of code, which everyone typically does nowadays, or Gemini.

Benjamin Shapiro (15:39)

Hang on, you flipped how I think about this already. I think of Claude being the Creative Writer and ChatGPT being the one that's better for code. Am I getting that wrong?

Steven Roach (15:48)

A little bit.

Benjamin Shapiro (15:49)

Damn. A little bit.

Steven Roach (15:51)

It's Claude is really good at cloud, incredibly good at code. It is the go to language or go to model for a lot of these organizations. A lot of tools that are being used today is built on top of plot. I think their Overall, I think 40% of their revenue is through their API. And instead of this consistent like chatbots, like ChatGPT typically gets their revenue from. That's a major thing. I'm not saying that ChatGPT is hauling that code. There's a lot of instances where it is Incredibly good, especially GPT5 Pro. Incredibly good with thinking. It's incredibly good. But you, I think there's tools out there like a open router that gives you access to all of these models to fit into your overall infrastructure.

Benjamin Shapiro (16:45)

You said before, there's your, let's call it the database layer. Right. Whether you're on AWS or you're using Google Cloud. Right. Where am I storing all my bits? Where's all my information coming from? Then you have an AI layer which is interpreting that data. And then I think you said that there was an interaction layer. So in between there, there's a router that's saying, okay, I understand what you're trying to do, let me find the right model. Right. It seems like there is a bridge between your data to your AI level or layer. Yes, that's a router. Talk to me about what are the router solutions. Because everything that I'm building is I got to pick a model and then down the road I'm going to have to update that. I'm going to have to go back into our workflow orchestrator and say, you know what, when I built this workflow, Claude 4.0 was, you know, the bee's knees and now we're on Claude 6.0. So I got to go back into my workflows and update them every time there's a new model. How do I integrate a router to fix that for me?

Steven Roach (17:53)

So within that workflow you need a node and I'm speaking directly to N8N, which is something that a tool that we use internally. You need to build a router that is specifically good for either one code another, creative writing another for simple questions and you can have it being routed to a open source model that you know, can answer generic questions with Internet access. The way that we have it currently built out is in that specific format. Yes, there's always going to be a layer or a opportunity for us to upgrade the model. We're moving at lightning pace right now and I'm pretty sure Gemini 3 is going to knock everyone's socks off and we're going to have to update. It's always an iteration that we're going to have to update our models within our workflows across the board. But you do not need to be locked in to one specific model or organization. Like if you're or organization, you're choosing ChatGPT for everyone and leaving out your dev team front end or back end when they prefer CLAUDE for other things, you're locked in into a massive contract with them. How do you get out of that? How do you service your other teams? And that's where I do not want any parts of our organization at QT to be locked into one. We're moving too fast at this pace so we need to be very open minded.

Benjamin Shapiro (19:34)

So there's this scale of the me's of the world who are developing infrastructure and workflows by themselves, Vibe coding away, creating Python scripts that they don't know what they actually mean and you know, using sort of user friendly interface tools like Zapier and make. Honestly, I kicked the tires with N8N and I was like, I can't even get our web hooks to work right. Like the basic triggers I couldn't get to function because the platform is a little bit more technical that I could easily understand. So it's like, oh, I should use a router to understand what the models are. I don't know if Zapier can handle a model to choose what I should be using for each individual task. But there's the solution n8n that if you have a probably a developer, at least somebody a little smarter than I am can use to essentially take the model selection process out of the selection process. You don't have to choose it anymore. Yeah, all right, great hack. Hopefully we can figure that for some of the down market tools as well. Then we got into the human in the loop. Like how do you sort of build in these notifications and checkpoints to put the sticker on your avocado? What are some of the other hacks that you have when you're building your workflows and automation to make sure that your outputs are always the best they could possibly be?

Steven Roach (21:02)

Oh, there's I. We have a protective agent to review every log and that's, you know, a lot of the logs, especially for speed. You just have to get the outputs out there. But there's a security layer that reviews every prompt that was inputted and every output. So it gives us insight into the direction of organizations or direction of departments or employees are asking specific questions across the board. So we can modify that across the board. So we. Protective layer.

Benjamin Shapiro (21:42)

In plain English, you have a code police agent, right, that is looking at everything that is going out and saying, nope, this didn't pass our security standards. Or highlighting when something has a potential security risk.

Steven Roach (21:59)

Yes. And we have a hash layer on top of that as well. We don't want any client information or names to be inserted into any model. So it hashes any, it captures any name before it actually is inputted into the model and then it actually regenerates that name when it actually outputs as well.

Benjamin Shapiro (22:23)

Let's talk about the risk because I went to this entrepreneurs meeting yesterday and one of the tasks was, hey, you've got your P and L. Download it into a PDF and upload it into whatever AI model you want and then ask it financial questions so it can tell you what a CFO would say about your P and L. And I'm like, hold on, I'm just going to give my P and L to Claude or ChatGPT and then say, hey, just don't share this with anybody. Like, I've got a paid account, I've got all the privacy restrictions turned on. But like, come on, it's my P and L. It's pretty. God, I probably shouldn't even be telling this. And so like, what, what's the risk when you're uploading sensitive data? Client names, you know, any passwords, like, if it's not hashed, what can happen with certain models?

Steven Roach (23:16)

And, and this is not for all of them. I know Claude, with their teams and enterprise pricing structure, they do not train their models based on your inputs. I do think that's going to change over time. So you run the risk of all of your information being trained, your conversations being trained, just to improve their models over time. And the sensitivity or confidentiality of that information that you've uploaded could run your organization into massive risk and lawsuits. We're at the very beginning of all of this. We're trying to navigate where the legal implications are going to take place. And I'm pretty sure a lot of our legal teams are just like, I want to make sure that no information is being uploaded into any model. Just so we do not have or bury the risk of that. I'm not saying do not use any models, but make sure that security risk is there for your organization.

Benjamin Shapiro (24:26)

I understand what you're saying, which is you don't want to put sensitive information to a model that could be public. But it's not as simple as, here's my P and L. So anybody can go into ChatGPT and say, hey, can you show me? I hear everything's P and L, right? Like, there has to be some logic in between with the LLMs that's saying this is information that should not be shared. So like, you can't just go and ask for all of this information and assume that it's going to come out the other end when somebody asks for it. So how is it actually a security risk if the LLMs are taking this data, even if they're training on it? This is what a P and L looks like. I'm assuming that they're not just saying, well, I can spit out this information to anybody that asks for it.

Steven Roach (25:11)

I believe over time, once these, you know, models are being trained on our overall conversations, what's really going to happen is certain information can be leaked about your organization or your, your clients. And, and I think that's a major thing. Client has spent, you know, over, you know, $15 million on a certain campaign. Here's the ROI based on that and over time, as long as these models are being trained, it can spit out that information. And if, if there's enough of that information, it can relate. Like if, I'm pretty sure you asked for sources of that data and if, if it provides a source, what's it going to do? It's going to provide your information as a source of that.

Benjamin Shapiro (25:57)

When we upload the transcript of this interview into our LLM for processing, I just want to make it clear that I do not pay my utility bills with company credit card.

Steven Roach (26:08)

Got it?

Benjamin Shapiro (26:08)

Full stop. I just. Let's make that clear.

Steven Roach (26:11)

But there is one additional element to that. Claude has a service called financial Services. I do think organizations need to look a little bit deeper into like their accounting partner using those services because that is a service that would never be trained. Their models would never be trained on that particular data. So they need to be very cognizant. And this is common sense at this point. If you're going to upload your PNLs, do not do it in your personal or even your enterprise model. Use the cloud. Financial services. Yes, it's a little bit more, but that's an extra layer of protection for you.

Benjamin Shapiro (26:50)

Okay, so we talked about model selection. We talked about a little bit about orchestration. N8N Tell me N8N seems to be the sort of term du jour or the orchestration tool. Orchestration. Well, I don't even know what to call it.

Steven Roach (27:07)

It's an orchestration tool.

Benjamin Shapiro (27:09)

An orchestration tool. Langchan Make Zapier down market what are some of the other options? How do you think about the difference between them?

Steven Roach (27:19)

I limit my other options to really LangChain. I know make is another one. Rider is another popular one, especially in the healthcare sector. There's a lot of protection layers for those tools. But I love the openness and the capabilities that LangChain and N8N provides. We can automate anything, any department. There's no limitation. And that's the beauty of it. And that's why I chose it for our internal uses, but also recommending it for external uses as well. I want to make sure that it's secure but has the capabilities of the creativity that myself or my team can make anything possible. Can automate something as simple as a PM task that helps our PM team, but can help us as front end developers or rear end developers or just overall your AI department. I think that's the beauty of a lot of these tools. So zapier. I think over time they'll open up. But N8N right now the reason why it's so popular is you can create anything from the ground up.

Benjamin Shapiro (28:32)

Yeah, there's there. I kick the tires. We've been building workflows with Zapier for years and like I said, I tried to make the migration to N8N using Claude to take my existing JavaScript and just turn it into NN8N workflow. And all I found was, yes, it generally got the structure right when I tried to recreate what I already had. But none of the actual workflows worked right. Like the individual steps needed different code, different integrations, different step. It was just so complicated to get the individual steps to function that I'm like, okay, this is another level of sophistication who can do these automations? Like when you're at a smaller organization like mine, when should you be focused on learning and sort of biting the bullet to training yourself on getting up to speed with the technical knowledge to be able to use a tool like Langchan or NNN as opposed to hey, Zapier is more user friendly. Make is kind of in the middle. When should you bite the bullet and start to invest in learning how to use these tools?

Steven Roach (29:43)

I do think that is a downturn or downside of N8N is the user friendlyness of it. There's always an opportunity to continue, you know, educating yourself on, you know, a lot of the capabilities within these tools. Either it be JavaScript or Python, you can, you can definitely vibe code, a couple of things to insert into a specific, specific node. But I do think there's always a larger opportunity to, you know, integrate, you know, those skill sets into your workflows because they're incredibly useful and incredibly powerful. A lot of what we do is either in JavaScript or Python and it just gets the job done. So there are, you know, your backend developers who can actually do this. A lot of your newer AI engineers or researchers can definitely do this. But the typical day to day, you know, end user can learn it. I don't want to discourage anyone from not learning a language because it's only going to make us better.

Benjamin Shapiro (31:02)

And yeah, but I'll be honest, I'm not learning how to code. The coding language for me is English and, and that's where like the translation layer between like dummy podcast host and Python script is Claude. Claude, right. And I'm going in and being like, all right, here's the trigger in Monday and I need to update these boards and I want to call this webhook and you Python this stuff for me and it'll spit out some code, I don't know what the code means and putting it in and all of a sudden it works.

Steven Roach (31:35)

So a great example of that is using CLAUDE to build your workflows. You can vibe code something nn it uses JSON, you can vibe code a lot of your workflows using claude, using Gemini and export, inserting that, testing out your workflow. Modifications may be needed, but you can definitely automate that process. It gets you there. I say 80%, 90% of the way. Some configurations may be needed for like Slack integration authentication credentials and things like that. But as far as like the Python script, CLAUDE would be able to answer that for you. It gets you there a lot faster. I think that's the thinking that needs to be integrated into each step is if you just need something really quick, it'll create a nodes, it understands it. It may not be fully up to date on the new advancements or updates. With nan, that's when you direct it to the updates and then it would integrate that for you.

Benjamin Shapiro (32:39)

Yeah, it's a brave new world. And this whole concept of vibe coding is turning idiot podcasters into engineers, which I don't know if the world's going to be a better place if all of a sudden I'm responsible for writing code. It will. Our workflows work because we test them and because we have somebody that's smarter than I am that looks over and be like, yeah, this all checks out. But let's be honest, like anybody, we're in the age of personal software, right? Like anybody can go and create their own workflows. And whether using some of the user friendly tools or going up into like the Enterprise. N8N Langchan more sophisticated models, like in theory, anything is possible to automate now we have the tools and you don't necessarily need to have all the coding language fully baked. You also mentioned replit, which is my understanding. Like replit just builds it for you. You don't have to understand the code, it's going to create the code base for you. And again, it's another tool. How do you think about the difference between the workflow automation where you're going in and creating the actions and steps and workflows and even your scripts, as opposed to replit, which is just going and building software for you.

Steven Roach (33:49)

You control every aspect of it. When you build out each node, each workflow, you run the risk of having replit, unfortunately, like the tech entrepreneur just running the show for you, I want to make sure that each step is doing exactly what we, the agency actually wants it to do, I wanted to make sure that it's one converting the prompts input into the best case scenario of a prompt, engineering and automating that process for us and making sure the output is exactly the way that we frame it and inserting that into a communication or a HTML file. I want to make sure I'm orchestrating every step along the way and then having that just run seamlessly without any error.

Benjamin Shapiro (34:45)

Fundamentally we're moving, like I said, into the era of personal software where the understanding of what the code base is is becoming less important. Anybody can write code, can run automations, can build their own software. What advice do you have for the builders that are not engineers to make sure that they avoid getting themselves into trouble because we can't go through, read the code base and understand it. How do you put guardrails and protections into what you're building to make sure that it stays up to date, there's no security risks and that it's actually going to keep functioning?

Steven Roach (35:27)

I think there's two ways of going about it. You still need your know it alls, your front end and backend developers, your overall AI governance team, to be fully involved in those steps. I don't think we should ignore them, but I do think having a security layer for everything that you do is highly important. Matter of fact, just build another agent just to double check. If you want to build an agent in Ripple it, have another one. Just double checking every step, making sure that it's going through and making sure it's thinking correctly and you know, just making, just call it your security agent. You know, be very cognizant, very, very mindful of how you're actually going through and building out each steps. Because unfortunately there's a lot of individuals out there running into those steps or running those issues without checkpoints, without, you know, that extra layer of security built into their processes and they're running free.

Benjamin Shapiro (36:33)

You know, before I talked about how there's this desire to improve margins, to automate, automate, automate and make your business more profitable. And it comes at the risk of quality. Right. And deteriorating your brand over time by not having the human in the loop checks. And that's a problem us marketers are going to have to figure out and face. On the technical side, it's the same thing. We have this need to develop more personalized, sophisticated code without an understanding of how and why it works. And fundamentally what we're doing is we're improving our margins, we're giving ourselves Access to more sophisticated software. But we're running more risk. We're creating more opportunities for either code malfunctions, security risks, all these problems. It seems like there's two ways you can go. One, have a human in the loop. Have somebody who understands what the code base is to go through and vet and make sure that what you're creating checks out and you're not going to get yourself into problems. You're going to incur some more expenses, you're going to be able to sleep at night. Two, continue to go down the path, use artificial intelligence to do the checks for you. For you. Yes. And again, there's still some risks, but it's probably a better security future proofing option than just doing nothing at all.

Steven Roach (37:57)

Exactly.

Benjamin Shapiro (37:59)

All right, Stephen, I want to move on to our lightning round where I'm going to ask you a couple of Martech related questions related to human AI and oversight. Are you ready?

Steven Roach (38:07)

Definitely.

Benjamin Shapiro (38:08)

What's the biggest data visualization mistake you see Fortune 500 brands making that kills executive buy?

Steven Roach (38:14)

In burying the business impact metrics in a sea of charts and not really reviewing or getting to the impact insights. There's been often a lot of times where Tableau has an issue with getting what our executives are actually looking at into plain text. My executives only want what is impacting their business in plain text. Numbers are up, numbers are down. Here's the correlations to why those are the most important things. So we often bury that into fancy charts, which are really beautiful. I've created a ton of them. But we need to get to the impact of why that is. What's the correlation? Why is that correlation there? Like let's, let's really get to the weeds of it. And I think integrating AI components in those tools in the future is going to be a massive thing. You know, I wouldn't be surprised if there's a consolidation of platforms in the next five years. You know, because there are lack of input, lack of AI integration and things like that. We, we need to make sure AI inputs or components are actually within these tools to get to the deeper details that our executives may be looking for.

Benjamin Shapiro (39:37)

It used to be so hard to build these customized reports that you really had to know the number you wanted to look at. Right. I want to track this metric over time in a rolling, 12 month, whatever it was. Right. We had to be so sophisticated in our thinking to isolate this one number because it was hard to build the report. Now it's easy. And we were drowning in this sort of fog. This, this Sea of data bloat. And we essentially lost a little dog fighting skills to understand what metrics really matter. All right, let's move on to our next question. What's the first sign that a company's AI implementation tool is about to fail spectacularly?

Steven Roach (40:20)

That is a very loaded question because I think one of the first red flags is our leadership may treat AI like a magic wand. Let's throw some AI at it and it's just immediately going to fix itself. You have to be very strategic about how you implement AI into these, into these tools. And I think a lot of the warning signs have always been there. You just have to be able to integrate AI in a very strategic way and making sure that it's actually purposeful. Again, going back to deterministic and non deterministic factors. If you are, say, throw all deterministic factors into the non deterministic, you're going to run into issues, you're going to run into security issues. Definitely just logic that is built in. I actually saw a recent podcast from the CEO of Box. He made that very clear. Enterprise software is not going to go away because of it. And we need to make sure that it is fully integrated into each step. There's layers to this. So you need to be very mindful again in how you integrate your AI infrastructure and making sure you're not falling flat. And the most recent example of that is the MIT study. I believe 95% of AI agents have failed in organizations. In a test study of like more than 600 organizations or 300 organizations, that's a very telltale sign. And it really didn't come back to the models failing. It was just that gap of knowledge. The models are so good, the implementation of those agents is still, you know, a massive drawback for a lot of organizations. So again, making sure that you are mindful of how you're actually implementing those agents or the full AI implementation across.

Benjamin Shapiro (42:29)

The board, it comes back to deterministic versus non deterministic.

Steven Roach (42:33)

Very much right.

Benjamin Shapiro (42:34)

If, if there is a very binary output, a yes or no question, does this exist or does it not automate away AI the crap out of it? But if you're saying, hey, artificial intelligence, I need you to make a judgment call on something and you're expecting it to get it right all the time, you're setting yourself up for failure. That's where you need humans in the loop.

Steven Roach (42:57)

Without a doubt.

Benjamin Shapiro (42:58)

Let's move on. What Martech trends are you watching that most marketing leaders are missing?

Steven Roach (43:04)

I think the advancements of AI in our traditional martech phrases. Take mmm, for example, next gen neural networks, which accurately predicts or or measures your customers or your internal overall workflow. For mmm, it enhances what we currently have in place. So evaluation enhance and then we continue to progress over time. I think those are very powerful steps that are often overlooked. And I think that's. Those are the key things that takes us to another level. You want accuracy, you want speed, but how do you get there? I think those are great steps to support that argument. There's other things that are out there. I think one of the more major things and something that we touched on is does AI orchestration, Is it going to take over the full AI agent implementation? I don't think it's going to. I think there's advancements. But orchestration is a key part of how AI is involved in your organization in today's world. And having those guardrails is going to be a major piece to that as well. A more technical aspect of the things or. I can get really into the weeds when it comes to Transformers and how it's used in Mamba. Mamba is going to be the next wave of how data analytics is actually involved.

Benjamin Shapiro (44:47)

Mamba, like Kobe Bryant, the black Mamba?

Steven Roach (44:50)

Yes, sir.

Benjamin Shapiro (44:51)

Okay. Why the Snake? What is Mamba?

Steven Roach (44:54)

Mamba is a specific coding language. So what's currently happening during studies is they're replacing the codex of Mamba and replacing it or replacing transformers. Transformers allows machine learning to operate as best as possible. And a lot of things when it comes to overall analytics that are currently being done in Python, replacing it with Mamba, it speeds up every aspect of your model. There is a road. And I do believe that this is going to be a more hybrid way of moving forward is having both Mamba workflows and your traditional transformers moving forward. So having that Codex, whichever one that you choose to involve, is going to run a little bit more smoothly. With Mamba. Your more traditional analytics, or just AI in general, is going to be using Transformers, which is currently being used today.

Benjamin Shapiro (45:54)

There's a fundamental architecture change coming that's going to impact how all of the LLMs and all of your analytics operate. You're not going to be Transformer based. You're going to use this other, let's call it a language called Mama.

Steven Roach (46:07)

Yes.

Benjamin Shapiro (46:07)

Okay, Library, You're. You're speaking a language that is way too technical for me to understand. But I'm. I'm excited and fearful of Mamba now. All right, last question for you. Can you explain to a CMO why their AI integration is actually driving customers away.

Steven Roach (46:26)

I think one of the biggest issues that we have today is we are getting deeper into misunderstanding the human interaction. People want to interact with people.

Benjamin Shapiro (46:41)

You know, that's a novel concept.

Steven Roach (46:44)

Yes, but they do, you know, and you see it in today's. Not to get off on a small tangent, you see it in today's kids. My, My kids watch streamers. Very limited amount. They watch streamers. They're interacting, though. They're speaking to one another. And I think that's. We're missing that in business today. So there's, there's still a. A concept where, you know, when you call, you know, Verizon and you immediately get this AI voice and it's instantly redundant. It's like, I don't want to speak to any AI. I want to speak to a person. And I think that's. We're missing that. And I'm not saying the solution is to remove AI, is to design the experience to fit AI into it. You know, we need better handoff triggers, better overall contextual personalization. That fits me. My plan is very different than yours. Verbal could be the same, you know, but we're two different individuals. So having AI understand and help us understand is going to be a major, major step forward. And then, you know, making sure to enhance every step going forward. You know, we're. We're not. The use case is not to remove humans. It's to better our overall involvement and enhance our overall processes across the board.

Benjamin Shapiro (48:06)

You know, there's a consistent theme here where we talked about how marketers want to automate to improve margins and builders want to automate to not have to pay an engineer to build something for them. And in customer service, we want to integrate artificial intelligence.

Steven Roach (48:25)

But at.

Benjamin Shapiro (48:27)

Each one of those business functions, there has to be a balance between understanding where you're sacrificing quality. If you're a marketer and you think you can automate without having a human check what you're shipping, you're setting yourself up to fail. If you're building something and you think you can just vibe code without having some sort of sanity check, that what you're coding isn't going to create undue security risk, you're creating problems. If you're in customer service and you think that people want to engage with your AI integration, you're taking crazy pills, right? Humans want to interact with humans. Now, artificial intelligence can be human. Like it can create value, right? It definitely can lower margins. But if you sit down, you ask a person whether they want to integrate with your chatbot or whether they want to talk to a person, I think you know the answer. And fundamentally, we all, at every function of our business, need to think through where we want artificial intelligence to support what we're doing and where we actually still need the human logic, feel. And the term du jour is vibe, but just the interaction level between one person to another to make us feel like we're still human. And Steven, I appreciate you coming on the podcast and telling us a little bit more about where we can use artificial intelligence while remaining human and keeping that humanity in the loop.

Steven Roach (49:59)

I appreciate you. Thank you, man.

Benjamin Shapiro (50:01)

All right, that wraps up this episode of the Martech podcast, thanks to Steven Roach, the VP of Data Ecosystems at Qualified Digital. If you'd like to get in touch with Steven, you could find a link to his LinkedIn profile in our show notes or on martechpod.com or you can visit his company's website, which is qualifieddigital.com if you haven't subscribed yet and you want a daily stream of marketing and technology knowledge in your podcast feed, hit the subscribe button in your podcast app or Visit us on YouTube and we'll be back in your feed next week. All right, that's it for today, but until next time, my advice is to just focus on keeping your customers happy. Foreign.

Podcast Announcer (51:01)

Thanks for listening to the Martech podcast, and I hear everything. Production Looking to launch or scale a podcast like this one for your brand? Then visit iheareverything.com.