October 22, 2025
TL;DR: Swept AI’s founder/CEO Shane explains why AI deployments fail without active supervision, how a “trust layer” prevents brand-damaging mistakes, and why every enterprise will soon need an AI trust score—especially in high-risk domains like digital health, finance, manufacturing, and insurance.
On this episode of the Futureproof podcast, host Alex Meisnik sits down with Shane, founder and CEO of Swept AI, to unpack one of the toughest challenges in tech right now: trust in AI agents.
What Swept AI does: Think of Swept AI as a supervision layer for AI agents. Instead of benchmarking the underlying models (ChatGPT, Claude, Gemini), Swept AI plugs into where teams build and deploy agents, evaluates how they perform in real contexts, and actively supervises and monitors them over time. The aim: prevent teams from “babysitting AI” and let them get back to running the business.
Why trust breaks: Shane recalls a public failure—when a major nonprofit replaced human counselors with a chatbot that optimized for shorter, cheaper conversations and ended up giving harmful guidance. The lesson: passive checkbox security isn’t enough. AI systems are probabilistic, evolve with every interaction, and can move faster than humans can notice mistakes. Supervision has to be active and continuous to preserve user trust.
Blue-collar AI to enterprise scale: Many early wins came from “blue-collar AI” (small manufacturers, restoration firms, and mom-and-pop SaaS). Teams used chat tools for emails and quoting, but struggled to convert experiments into repeatable, monitored workflows. Swept AI helps codify those workflows and flags failures—e.g., inconsistent quotes sent to large customers—so staff can run the business, not the bot.
Security parallels: Just as cybersecurity evolved beyond “prevent every breach” to detect, contain, and respond, AI trust requires limiting blast radius. Expect threats like prompt injection, jailbreaks, data leakage, and unexpected tool calls. The job isn’t only prevention; it’s rapid detection and mitigation.
A nine-pillar trust framework: Developed from decades in ML plus real-world partnerships (e.g., large nonprofits and universities), Swept AI combines operational lessons with best practices (SOC 2, NIST, etc.). The focus: transparent, active, always-on oversight because agents morph with context.
Toward AI trust scores: Shane predicts enterprises will evaluate AI more like hiring people: third-party vetting/certification (think “credentials”) plus ongoing performance checks. Today’s AI is like an eager intern—capable, but inexperienced—so certification alone isn’t enough; supervision in production is crucial.
Advice for founders:
Who should call Swept AI: High-risk sectors—digital health, finance, insurance, manufacturing, oil & gas, and enterprise SaaS—where AI mistakes can harm people, revenue, or reputation. Lower-risk use cases often have “good enough” tools; Swept AI is for when failure has real stakes.
One thing to make AI safer today: Evaluate statistically before launch (not just a vibe check with a few hundred cases). Test for the happy path and the failure modes. Also test with non-experts—real users won’t prompt like your in-house power users.
What’s next: Teams are moving past “chatbots that help helpdesks” to automating hard, repetitive work across the org. Shane wants software—not humans—to supervise AI, freeing people to focus on creative, high-value problems.
Alex: Thanks everyone for joining the Futureproof podcast. My name is Alex Meisnik. I am joined here from Shane. He is the founder and CEO of Swept AI. We’re going to go a little bit about to into what his business is, what they do, go a little bit about industry insights, and then we’ll just go from there. So Shane, thanks for joining me.
Shane: Of course. Great to be here.
Alex: Absolutely. Well, uh, why don’t you introduce Swept AI? Give us a little bit of a context about what what you guys do and and why you started the company.
Shane: Sure. So, Swept AI, you can think of it as trust layer for all things AI agents and Aenics. So, we’re not going to go on to the model layer and try to tell you how well ChatgPT or Claude or Gemini do, but instead we’re going to we plug in to wherever you’re building agents or potentially you’re deploying agents inside of your company and we’re going to evaluate them to see how well they actually perform in your context and then in the long term supervise monitor them to make sure that they stick together. And this really comes for me personally. I’ve been doing uh AI for 20ome years at this point with my masters in AI from years ago when we were doing AAR algorithms and like just primordial neural networks at that point. But I was really focused in what do you when you have systems that behave unpredictably or essentially probabilistically is really what we were looking at and how do you manage those? done that for 10 years in the data science world and then been in startups for the last 15 years as a CTO deploying various systems and just saw the opportunity in 2023 when people started deploying their first chat bots into systems and you know immediately had problems which were predictable if you were a data scientist but maybe not if you were a programmer or business person and so this technology has just really impressive potential here but it’s a tool and you screw up with tools if you don’t know how to use them. And so we come in there to really help you make sure you can use that tool and take it to its best ability inside your company.
Alex: So, you know, from someone coming in that really doesn’t know a lot of background about, you know, you know, bots that you businesses install like chat bots and any kind of other context of it, what is a really good like use case scenario for you um for any type of a scenario for like a small business if they’re running any type of like chatbot or anything that you guys can help with? give me like kind of like a a quick overview of like what you guys do, how you come in there, and how you help them.
Shane: Yeah. So, you could almost think of it as we really help operationalize often those first workflows. So, a lot of our early customers were what I’d call kind of bluecollar AI. They’re in manufacturing, small manufacturing or even just small like restoration companies, all kinds of things like that. Um, and some small mom and pop SAS. and they had some early successes using chat GPT or something to help them write emails or start with some workflow on quotes, but they weren’t really able to do much other than go put something in chat GPT and have it come out. And so we can help turn those into repeatable workflows where they don’t think about it anymore. It’s, you know, there are plenty of agent tools that help them do that. But what we really do is we help them evaluate and monitor it because what we do find a lot of times here is people once they start using AI, they stop doing the fun work or the new work and they start babysitting it. And that’s really where we want to take over. So we evaluate their AI, make sure that if it’s running quotes for them and sending it off to a large automaker that it’s consistent and flags it when it does poorly. And that just allows them then to actually run their business instead of just become a manager of AI because that’s I think that’s personally a biggest fear for a lot of folks is you just become a babysitter of AI instead of being able to enhance your business in other ways.
Alex: Got it. Yeah. I think that’s super valuable. I mean, especially from someone like you said like blue collar really just wants to run their business efficiently and you know get do what they know best and obviously that’s not you know AI if they’re you know running like a you know a house services company or anything like that you know their focus is not that they want to run it as efficiently as possible so you can kind of come in there and make sure that they’re doing it the correct way precisely so with swept AI is tackling one of the hardest challenges in tech the trust part so what inspired you guys to you know focus on AI integrity, safety, and compliance.
Shane: Yeah. So, our very first like recognition of it was when the National Eating Disorder Association had a labor dispute and so they put a chatbot in place to replace their humans. And uh it just had a poor reward mechanism and immediately started telling teenagers that were on this chat, hey, you’re struggling with bulimia. Let’s measure your belly fat or let’s talk about anorexia instead. which did exactly what it was meant to do, which was keep the conversation short and spend less money on tokens, but it completely eroded the trust of a pretty reputable and worthy organization. And so that’s where we see it’s key is a lot of times people uh they’re skeptical with AI and then they start trusting it too much and it goes off the deep end very quickly. And it’s just because again, we’re not cut out for predicting uh the way these things can be, but we’re also not cut out for having constant attention on our AI systems. And so if we want to actually trust systems that aren’t going to behave consistently, we have to have some mechanism in place. And it can’t be passive. It has to be active, supervised, and really informing us.
Alex: Yeah. I mean, like you said, it it probably thought it was doing the correct thing by giving other examples or or uh things to ways to help that, you know, that patient, but obviously it’s not trained in, you know, psychology or knowing like what that suggestion was going to do to somebody that had, you know, an issue like that. So, that’s kind of where you guys can come in and almost train it from the back end and kind of tweak it a little bit.
Shane: We retrain, we make sure that it’s aligned on what are the actual outcomes that you want. And then there’s a lot of times uh there’s great Donald Rumsfeld quote of uh from the you know there are things that you know you know things you know you don’t know and things you don’t know you don’t know. And that’s a big part with AI and what we you do with something like active supervision is you’re not going to be able to predict every way that it’s going to do something silly. doing a couple hundred emails manually yourself is just not sufficient to really know the bounds of a system. And so having something like supervision lets you go, okay, there are things we don’t know, we don’t know here, and we’re going to use statistics to be able to uncover them before they go and severely harm, you know, a person or your financial parts of your business or, you know, if you’re manufacturing something, it orders the wrong parts or too many or whatever the case may be. They’re all large impacts to your business and your livelihood. And those are just things that you want to catch as uh quickly as possible.
Alex: Yeah. Well, where my mind goes cuz I talk to a lot of like IT and like cyber security companies and that’s where my kind of main focus is and their trust is obviously in like cyber security attacks and things like that where your reputation is hurt. There’s like some quote or not quote but like a statistic that like after a cyber security attack on a company on a small business that they’re out of business within like 6 months after that cuz their trust is completely gone. So now you’re just taking that differently. you’re doing it on a not a cyber security threat level. You’re just doing it on something that it’s your businesses or your trusts message is getting, you know, taken the wrong way by a chatbot or, you know, some type of AI personality that maybe not relaying what you guys are trying to do the correct way. So,
Shane: exactly. and kind of your example. You know, in the past, it’s been very hard for a singular human in your workflow to completely take down your business without it being recognized because humans can only work so fast and only have so large of an impact. But when you’re handing over the reigns to some key things on your business to an AI system, it moves fast and it can just make these decisions and impact you a lot quicker than you could ever predict. And so getting in there and I think it also though goes to the cyber security of you’ve now got a a digital version of social engineering where people can attack these agents and just like cyber security today we know that we can’t train humans to buffer every single cyber you know every time a social engineering or fishing attack happens to them. We can train them to reduce it, but we have to be able to also understand a lot of cyber security is how do you respond? How do you limit the blast radius of this? And the same thing happens in this agent layer with us where it’s just as much about detecting how likely is this, you know, jailbreaking attempts, leaks at private information, prompt tools, tool calls that you weren’t expecting and catching them quickly enough to recognize there’s a pattern here and stopping it versus assuming that you can stop everything because we learned that I think a little bit the hard way in cyber security over the last 25 30 years that you can’t prevent every security vulnerability that’s out there and so you have to be as much about fire, you know, you got your fire prevention, but what’s also your fire response and your fire fighting.
Alex: Yeah, absolutely. I think that’s a really good insight on it, especially without, you know, bringing those two worlds or what you’re trying to do together. So, your nine pillar trust framework sounds like the foundation of like safe AI. How did you develop it?
Shane: So that was in code development with I mean ourselves of just having done this for about 20 years of knowing the places that machine learning models already have problems and then layering that with generative AI but that’s also in partnership with some of our earliest customers like the United Way and the University of Michigan where we’re going into the real world settings and figuring out where does this have real impacts on you know the customers or the patients inside of these systems. and to make sure that we’re getting everything. And then you also layer on, you know, best practices from SOCK 2 and your NISTs and the various things to say here’s what it is. And this is really just us trying to lead the way and push beyond passivity inside of these of a lot of security around AI and that is very passive checkbox based. And we really want this to be trustworthy, transparent, and just continuously active because these systems morph every single time they have a new interaction, a new piece of context. So, we have to treat them differently.
Alex: Yeah. Yeah. I I definitely agree. Um, so with you guys right now, do you think we’re headed into like toward a future where every enterprise AI system will need like a trust score kind of like what cyber security certifications are like today?
Shane: Yeah, it’s if we really push it out and think about how we talk about AI in the future, we talk about it on something that’s way closer to the way we hire people. And that’s where you are. You’re going to want some trusted place that says, “Hey, we’ve vetted and certified.” It’s the equivalent of having, you know, your PhD in the back or your AWS certification that says you do this, but then also you’re going to need some way to actually know that they keep on the job. It’s uh I remember giving a talk not too long ago where you have to think of AI today until we have some major breakthrough like a really eager intern. It’s going to do everything. It has zero experience and can just completely screw up. And so we do need these ideas of trust scores and assessments so we have some idea how they do but also we can’t just back off and say hey once it’s been certified trustworthy we can let it go. We really have to dig in with it. Um, all right. So, I kind of wanted to go maybe into something a little nuggets you can help other, you know, founders or CEOs or people that are starting, you know, a company like yours or in the same realm. You know, when you guys are going through this process of starting Swept AI, I know you guys are in the infancy stages or just the starting part of it, but where what are your biggest challenges or kind of what were the things that you know, roadblocks that you hit that you would probably change or not do again or maybe just do a different way?
Shane: Sure. I think some of the biggest roadblocks are chasing venture right away before you have customer proof points. I would strongly push ourselves and do it again. And still to this day, we do it of getting closer to bootstrapping of don’t expect somebody to bail you out with venture because they believe in your idea. They want it’s um a lot of it if it’s just betting on the idea. hopefully you have a great network than somebody that truly already will bet on you for an idea, but to think that they’re going to bet on you with no pedigree. Probably not. So, we spent far too much time chasing venture at the beginning once we thought we had a solid couple pieces of information. As soon as we had a handful of customers, the story completely changed and we were much more chased ourselves. And I think that’s where just take looking back there, I’d say focus on bootstrapping. Get, you know, can you do a runway for a year by yourself or doing it off the side of your desk versus just hoping somebody’s going to provide money cuz the one thing we saw is goalpost change all the time. Right now, if you have the word AI, maybe you have it a little bit easier, but that could change tomorrow for any reason. And you just need to still have good fundamentals. And if this business doesn’t fly without venture funding, then it’s probably not a solid business to begin with. You need something that can move in case the capital markets dry up on you.
Alex: Yeah. And yeah, I think like you just said, I mean, going after venture capitalists and investments, it must be really hard because I mean, obviously you guys when you’re starting the business, you completely 100%, you know, um, believe in it and believe in what you do and know that it’s going to help, you know, a bunch of companies. you know, pitching that to someone that really either doesn’t know you or um you’re just, you know, talking to for the first couple times. Yeah. It’s probably a little bit harder to get that message across to them without any kind of social proof or um you know, exact scenarios. So, basically what you guys did is you just slowed you slowed down a little bit, got some social proof, got some uh examples of like how you helped and then kind of the people actually started reaching out to you or wanting to invest.
Shane: Exactly. It’s you know we had a really big blitz uh when we had our first pilots and a lot of okay well pretty large numbers that you’d look for in revenue which we haven’t hit yet hit even today. Um but suddenly once we said okay we’re going to stop there and start focusing on going past these pilots to let’s get some recurring customers and we had six or seven small but uh good recurring customers then it became hey wait tell me more about what you’re doing and getting pulled into those conversations and the same is happening today where we’re having some more successes again and instead of having to go out and constantly I mean we probably had 150 nos on venture capital to start and so hey you got to be really resilient to that. But B, it’s like make sure that you have more than a story. What is it that’s pulling you pulling you there? And it suddenly became let me introduce you to this group. Let me introduce you here. I want to have this validated. And that’s when you know that’s there’s been a peaked interest, but just showing up and pitching pitching pitching. It can work, but it can you have to pretty much get struck by lightning at that point.
Alex: Yeah. So, are you guys are you guys focused on any type of one company or industry or niche? or you basically are trying to be that one, you know, stop shop for every business or every, you know, whatever. So, do you have an ICP or people that really should be reaching out to you that, you know, you can really help truly?
Shane: Yeah. So, we focus what we would say on much more high-risk areas. So, TR typically that’s going to be health, uh, finance, insurance, manufacturing, oil, gas, big enterprise SAS where you have some pretty big ramifications to companies. Not to say that we don’t work with the others when they have a high risk, you know, when it the risk to the reward inside of their maybe smaller company is outsized, but it’s really there are good enough tools for people who are doing things where the fallout to a failed AI system is, you know, is not going to kill their business or it’s not going to severely put them behind. There are good enough ways to handle that today. But when you have true risk to your system or your customers or your patients or whatever the case may be that where SWE falls in. So again uh you know I’d say our biggest place is digital health in all places inside of health but then it’s followed quickly up by that manufacturing and insurance.
Alex: I mean because digital health is huge. I mean, I even a couple years ago, even before AI, I mean, you have like the the bot where not the bot, but the chat thing where you can kind of go on there and put your symptoms in and it kind of gives you recommendations and things like that. And yeah, especially Yeah, that’s probably a big one to make sure that that recommendation or whatever that bot or that chat bot is actually giving them the right. I think there’s back then I think there was a little bit of a and it might be different now. they might actually be able to, you know, give a prescription or tell you what to do. But I back then, I think the I think their safety net was I think everything had to be reviewed by like the doctor and like the doctor prescribed you. So, you know, there probably wasn’t a ton of risk there, but maybe a little bit of if you didn’t make it all the way to that point, you know, the repetition of like, you know, what is this thing telling me, you know, turn it off or whatever.
Shane: Yeah. And you’ve got even, you know, the doctors, we’re humans. We’re we’re lazy by nature. And so you get enough AI recommendations that look good and the doctor just starts, you know, signing off on it. That’s where you get into trouble with these systems and why they’re needed today. It’s not the doctor’s fault. We’re we all become lax when we think that a systems running great. And so that’s where you really need that this kind of thing just to to make sure because our best work is not done with attention to details of super repetitive work where one out of every thousand is the problem. That’s just really hard for a human to catch.
Alex: Yeah. That’s actually a really good point and I mean yeah you get so maybe a doctor just gets so trustworthy on that one recommendation or something like that and then you know the one recommendation that was wrong is that’s the killer to like their reputation. So no I think that’s I think that’s really good insight.
Shane: Yeah, we see that. We call it the trust spectrum internally and we watch it with our own customers and like our platform and you can watch where they they tail off from superdetailed views of everything to eventually they never even log in anymore and they’re just relying on it working or getting alerts which is fine. It’s, you know, it’s a psych psychological phenomena we monitor, but knowing that is probably the most important thing for most businesses of it’s insufficient and there’s better uses of your time anyway than looking for needles inside of needle stacks.
Alex: True. Um, all right. So, if every company right now is deploying AI, if they could do one thing right now to make it safer, what would you tell them to do?
Shane: Evaluate it fully to a statistical level before you put it out there. So vibe checking it, putting a couple hundred, even a couple hundred emails is insufficient. Do the work to prove that it does what it says. And probably the other one I’d put on the other side is do a little bit of work to prove that it doesn’t do the other stuff. Uh because there’s a lot of times where we’re only looking at the happy path of how well does it, I don’t know, write quotes for me. Uh but not often looking at like well how much does it do these other things that I don’t want it to do. We just don’t often measure that. I’d say the final one that a lot of people just aren’t thinking about today when they do tests is they test it with themselves in mind which is usually a near expert of the system of like I’m going to go and talk to this AI to have it I don’t know diagnose my engine and they forget that the general public has no idea about engines and how they work and so when they come to utilize your system they are several levels below you and so making sure you’re testing with those long tales in mind of expertise. If you have a general purpose product like that, that is something people almost always forget and the agents get far worse often when they encounter people who are non-experts utilizing them because they’re not providing the context that you would expect. So, you need to put that into your system. How do you drag out more information? When do you stop? When do you elevate? People don’t do that today. That’s probably the one thing I would do is just be more robust. go get your, you know, you know, fourth grade nephew to try using the system, too, just because they’ll use it in a completely different way.
Alex: Well, yeah. I mean, because it’s, it even boils down to one person using, you know, one of the LLMs and getting a response completely different than somebody else that knows how to prompt it well, like, you know, how to frame the problem, tell it what to do, tell it how to act. you know, if you have that person obviously coming in that’s an expert in what they’re doing, you’re going to get a much different response than, like you said, your four-year-old uh nephew coming in and prompting it. So, yeah, it’s really good insight. All right, cool. Um, well, what excites you guys about the next chapter for Swept and kind of basically AI in general?
Shane: I mean, for AI in general, I think we’re finally cresting the ridge of people going and seeing beyond chat bots to what if we have AI performing tasks that were either hard or so repetitive that we didn’t do them. Like I feel like we’re finally seeing that in our own clients where they’re just going beyond easy stuff to well, wait a second, what are about these things that I do that are very repetitive? they they’re actually not high value in terms of work, but they’re hard work. That’s great. We’re now finally seeing, I would say, real value unlocked, which just means for me, I think there’s excitement of what does that unlock for those people in terms of what else they’re going to do and accomplish. That’s AI’s real potential. And that that’s where I also, you know, it’s great for swept because again, I don’t want people to supervise AI. I think that’s an utterly boring, useless task. I want I want software that does that. And I want humans to apply their ingenuity to new things. It’s every time we see a technological revolution, we unlock new capacity and things to do well beyond what we used to. And so that’s where I finally see like the industry has moved over away from just cool this chatbot can make help desks easier to wow, we can truly automate huge swaths of busy work inside of our company. What more can we do?
Alex: Yeah, that’s awesome. Well, awesome, man. Well, how can So, if the business was intrigued on kind of what they were hearing today, wants to reach out to Sweptai or you, how can they find you? What’s the easiest way to kind of get in touch?
Shane: Yeah, easiest ways.ai, our website, or contact me, Shane swept.ai or hello.ai. There’s a phone number if you’re old school and want to talk on the phone. We know that helps with some of the bluecollar relationship work there. So, that’s on the website as well. And yeah, just go in there, check it out, and we’re always happy to talk and assess where you are as well and point you into different directions or implementation partners if you need uh help there. Visit the website or send me an email.
Alex: Awesome. Well, thank you so much for joining me today.
Shane: Right on. Thanks, Alex.
Alex: Thank you. All right.
