AI Meets the Mid-Market: How PE-Backed Companies Are Leapfrogging with AI

00:00:10:19 - 00:00:32:27 Michael Welcome to disambiguation. I'm your host, Michael Fauscette. Each week we interview experts in artificial intelligence, generative AI, and business automation to help business leaders understand how to use these tools for the biggest business impacts. 00:00:33:00 - 00:00:44:13 Michael Our show today is AI meets the mid market.

How backed companies are leapfrogging with AI. I'm joined by Andrew Brooks, founder and CEO of contextualize. Andrew, welcome. 00:00:44:15 - 00:00:50:20 Andrew Great to be here.

I love the intro to the podcast. So excited for the conversation. 00:00:50:22 - 00:00:53:04 Michael At least it's lively. I like that part.

00:00:53:06 - 00:00:55:03 Andrew So there you go. 00:00:55:05 - 00:01:17:20 Michael So I you know, obviously we had a nice chat to set this up and in that you shared some of your background and I you know, it's pretty pretty amazing and entrepreneurial track record. I mean found in smart things which, you know, became this widely adopted smart home platform that Samsung acquired, bought, build and sold SMB live to reach local. 00:01:17:22 - 00:01:26:24 Michael Now, you know, contextualize.

Oh, I mean, what what took you from that smart home IoT space to AI solutions for the mid-market? 00:01:26:26 - 00:01:59:24 Andrew Yeah. I mean, there's a there's a common thread that is not always obvious in my entrepreneurial background. I graduated in the late 90s.

Everybody was starting companies then. So that's actually kind of what what kicked off my passion around being a creator. But the thread that kind of ties all of the entrepreneurial activities I've done together is one of there's a new technology, a new disruptive technology and underserved market who has not adopted those technologies historically. 00:01:59:24 - 00:02:23:01 Andrew And how do you create a platform and a capability that actually opens up that opportunity?

So with SMB live, it was very much around small businesses, hyper small businesses not yet being online, and how they should adopt that with a smart home. Of course it was. Smartphones came out in 2007, right? The real, you know, the iPhone in 2007.

00:02:23:03 - 00:02:49:08 Andrew So people had this new interface in their hand. All of these sensors were emerging. And we said, there needs to be this consolidating platform that would bring these capabilities together for the mass market consumer. And I think the the common thread with contextual is very much as we saw AI become API addressable and therefore able to be integrated into business systems.

00:02:49:08 - 00:03:08:19 Andrew We said there's an entire group, and for us it's this mid-market, you know, B2B services organizations largely who historically has been a laggard in technology adoption. And we felt like there was going to be a leapfrog moment for them to say, hey, we can do new and different things with this technology that otherwise we wouldn't have been able to. 00:03:08:21 - 00:03:12:14 Andrew And so we wanted to build a platform that made that possible. 00:03:12:16 - 00:03:12:28 Michael Yeah.

00:03:13:01 - 00:03:13:04 Andrew I. 00:03:13:04 - 00:03:37:25 Michael Mean, I definitely have seen this with some of the companies that I've worked with over the last couple of years that that for mid-market companies especially, there's some some real opportunity to, to, to use this technology and apply to, to their business and really leapfrog, as you've said. And you know, I, I'm curious from that perspective, can you paint a picture of, you know, what is that opportunity look like to you?

00:03:37:25 - 00:03:47:19 Michael And, and why is this moment different than some of the other technological advances in waves that we've seen in the past? 00:03:47:25 - 00:04:09:13 Andrew Yeah. I mean, we see AI as in two flavors, right? You have AI as a new form of electricity.

It can do things that historically you wouldn't have been able to do regardless of how much software you wrote. And then, of course, AI as a tool for accelerating the creation of software. And the combination of those creates an opportunity. 00:04:09:14 - 00:04:43:08 Andrew We think, where especially mid-market companies, they would have never had the appetite, the economics, the operational capabilities to invest in purpose built solutions.

Historically, they would they generally are going to have an ERP system. Maybe they have a CRM system, they've got their Microsoft suite. And all of a sudden, because AI is this new form of electricity and AI can accelerate development, they have the right to have a solution that is uniquely designed for their people, their data, their process, their way of work. 00:04:43:10 - 00:05:06:04 Andrew Potentially new human AI interfaces that historically, you know, they wouldn't have created.

And that that's the that's the magical moment here is is there right to it to to own their AI as what we would call it, own how these systems don't rent a feature but actually own a solution that is that is really specific to that business. 00:05:06:04 - 00:05:31:16 Andrew And, and, you know, when we look at these mid-market companies, they have 25 years of institutional knowledge and it's locked in various people's brains. And you would never again have been able to write the software that got to that intelligence.

But AI gives you the ability to, to, to extract and, and and leverage that intelligence. And so we certainly encourage our customers not to think of AI as just another technology. 00:05:31:16 - 00:05:40:16 Andrew It's not people process technology where AI are to that. It's people process technology and now AI.

00:05:40:19 - 00:05:59:19 Michael Yeah, I like that. I think, you know, one of the things I've talked to a lot of companies about is, is the mistake, I think, of trying to to focus on the technology and not focus on the workflow or the people. Right. Because that's actually the the what's really involved in this is a lot bigger than, oh, I'm just going to implement another cloud solution, right.

00:05:59:20 - 00:06:23:22 Michael I mean, it's it's it's much more fundamental in a lot of ways. And you know, I think to you, you guys are focused on private equity portfolio companies and, and and I, I know when you spoke at South by Southwest, you talked about how, you know, private equity drives AI adoption. And you made the point that, you know, you're not disrupting a private equity portfolio company from the outside. 00:06:23:22 - 00:06:46:26 Michael It comes from the inside out.

And which ties to my comments, I guess, to about about human focus and workflow focus. But what do you think from an inside out disruption standpoint, what does that look like in practice? And you know, when you're working with a company that's been operating the same way for years or decades? How does that transformation, you know, really take effect?

00:06:46:28 - 00:07:08:20 Andrew Yeah, it's you know, when you you see all the hype coming out of Silicon Valley and, hey, we're going to completely disrupt the the legal profession. We're going to completely disrupt the, you know, the accounting profession, those we would consider those kind of outside in disruption. Right. It's a new entity coming into existence that's going to disrupt a bunch of businesses who happen to do that.

00:07:08:21 - 00:07:32:08 Andrew When we look at our the portfolio companies that we work with, you are you're working from the inside out, meaning you're working, you're starting at how do they do business today? Why does something take three days? That could potentially take three hours? Because there's there's nuance to it.

There's there's there's maybe not mature data structures. There's manual processes. 00:07:32:12 - 00:07:37:27 Andrew You know, one of the the examples I would give is. 00:07:38:00 - 00:08:03:02 Andrew So one of our clients is a fleet management company.

And and by that I mean they manage like forklifts and pallet Jackson pallet movers in warehouses. They, they they do that through a vendor managed network. Those vendors are ranging from large scale fleet management companies like caterpillar down to, you know, Joe the mechanic who knows how to water the batteries on a, on a, on an electric forklift. 00:08:03:02 - 00:08:23:21 Andrew And so across 3000 vendors who might be interacting with them, they're getting 14,000 plus emails a month with work orders and invoices of different forms and different formats.

And you have to go into that organization and accept that that's the way business is done and has been done for a tremendous amount of time. And you couldn't just say, hey, we're going to stand up an API. 00:08:23:21 - 00:09:05:04 Andrew In all, 3000 of your vendors are going to adhere to the API. That's not a realistic, you know, way to transform that business, but that that example, 14,000 emails coming in with invoices and work orders being handled by 13 humans is a great example where AI can come in.

And I think the message to the team members there is we're trying to we're trying to take the most tedious part of this job, you know, extraction and coding stuff into a system out of your work path so that you can focus on the next layer of value, looking for cost savings, looking for opportunities for efficiency. 00:09:05:06 - 00:09:23:10 Andrew You know, it changes how they work. And so I think disruption internally is understanding the people, understanding why they do something this way.

And then looking for for, you know, kind of evolution, not necessarily complete revolution of that, of that experience. 00:09:23:13 - 00:09:53:09 Michael I mean, that hits on one of the things that I'd say we've been coming back to both from the shows perspective and then just in general, from from our research, is that it's easy for for people to get focused on the idea that AI is going to be a replacement for people. But the truth is, from all I've seen, at least for now, it's not that it's about building this hybrid collaborative workforce that combines digital with human in some unique ways.

00:09:53:09 - 00:10:20:07 Michael And and I know you talked about before that, like 90% of the solutions you build include some new human AI interface. I mean, can you could you talk us through what that looks like and, you know, a real engagement? And then how do you make sure that humans both trust that new workflow, but also adopt the system and incorporate that into their workflow? 00:10:20:09 - 00:10:48:18 Andrew Yeah, change management is absolutely critical.

And I'll hit on some examples that I think drive that home. So and everybody talks about human in the loop. The problem with human in the loop is it carries a little bit of I'm just checking the eyes job. And people don't love that.

So you know, you hear kind of human in the lead or expert in the lead as, as as I think more accurate phrases around, you know, the types of systems that we would build. 00:10:48:20 - 00:11:30:03 Andrew I'll give you a good example and the evolution of change management that a client goes through. So one of our customers manages about 3000 vacation rental properties down on the Gulf Coast of Alabama. And as part of that, they they they run 10 to 12,000 work orders a month, ranging from, we need to replace the batteries in the remote to a need to replace the light bulb in the in the ceiling to fix the leaking toilet or the fridge of making noise, because these are owned units, meaning there's an owner and they're providing a management service when a when a when a technician goes out and does work, they are having to report back to 00:11:30:03 - 00:11:54:21 Andrew the owner.

Hey, we have this issue. We resolve this issue. We believe that this issue has been satisfactory. Resolved.

Well, humans are fallible and in the field, humans might record bad notes in the ticket. They might say something that they don't want to present to the owner. The pictures that field services individuals might take are not always representative of the work that was completed. 00:11:54:22 - 00:12:28:04 Andrew You know, and so their their old process was we have to have a manager review very quickly the details, the summary of the work done and the pictures on this ticket before we will send it to the owner because we're concerned about presenting, you know, poor quality, you know, photos or description, you know, to these owners.

So the first step of that AI solution was a we can have AI look at the pictures and confirm these pictures represent, as described in the ticket of the work that needed to be completed. 00:12:28:04 - 00:12:48:12 Andrew These these pictures do an accurate job of representing that. And of course, it's like if the pictures of your feet and you were supposed to fix the refrigerator, it can it can identify that. Likewise, they have a set of rules that say, this is how we want to describe the work.

We completed things, you know, follow words as an example are not appropriate in the ticket. 00:12:48:12 - 00:13:04:14 Andrew We don't want to have that and present that to an owner. That's an obvious example. But they have a range of standards that they look to adhere to.

The human in the in the loop or in the lead experience. That new human AI interface that we initially created is what we would call a sidecar. It's a little dash. 00:13:04:15 - 00:13:23:07 Andrew It's a web based dashboard that sits on the side of the manager's computer.

And initially it is just evaluating each one of those work orders against that set of criteria and saying, do we think this is good or do we think this is bad? The AI was making that choice, but there was no automation actually, back to the owner yet. 00:13:23:07 - 00:13:50:08 Andrew It was more a very rapid tool for the managers to say, I agree with your decision or I disagree with your decision. I agree that this conclusion is right.

That was then feeding back into the system to make it better. So so so step one in change management is ownership of outcome, right? That the people who are using the tool feel like they have an ownership in the quality of the tool or the output of the tool. 00:13:50:09 - 00:14:18:13 Andrew Give you a great example.

There's standard requirements if a if a if equipment is replaced in a unit is, one must take a picture of the serial number of the new equipment. You know, that makes sense. If you're replacing a, you know, a pump on a fridge, it doesn't make a lot of sense when it's batteries. Right.

So but but the but the AI was being it was taking it literally and it was saying you replaced batteries. 00:14:18:13 - 00:14:40:06 Andrew I do not see a picture of the serial number of these batteries fail. Okay. That's a great example where the feedback of the humans says, wait a minute, we need to loosen these sorts of of boundaries.

Once they got comfortable, the next step of that process was to automate the approvals. They said, okay, we agree with these approvals. 00:14:40:06 - 00:15:03:06 Andrew The pictures are good, the description is good. Automate that approval.

Now that goes back to the owner automatically. So now you're suddenly saving me time. The initial step was you given me a new interface. Now the interface is just focused on where are the you know, where the problems manifesting.

Where does where does work need to be? Where does this this ticket need to be updated. 00:15:03:06 - 00:15:25:04 Andrew And so I think that kind of evolution of a new place where I can just see the speed with which this is analyzing, that I can participate in a feedback loop to make it better and get confident that it's doing a good job, and then ultimately it's saving me time. That's a transition that our clients love going on, and they feel an ownership through that process.

00:15:25:07 - 00:16:02:28 Michael I mean, that's a great story. And I know we talked about that story a little bit in the in the prep meeting. And, you know, one of the things you said about that, that I thought was interesting and I've seen this myself, that you can have some unintended consequences in a positive way in some of these implementations. But I know one of the things you said in this case was not only did they get the the improve greatly the way these reports looked and went to the owners, but they also started to find certain items that hadn't been invoiced or, you know, other ways in that to recover some lost revenue.

00:16:02:28 - 00:16:17:08 Michael And I mean that unexpected ROI, I'm sure obviously that was a big benefit to them. But is that common? And and how should you know? How should companies think about those AI projects sort of in that context?

00:16:17:13 - 00:16:47:15 Andrew Yeah. So what was funny about that situation was we we didn't really instruct the LM flag when units that were showing up as being used in replacement were not actually showing up on the invoice. We just gave it all the information and said, this is this is what you're doing. You're evaluating this ticket and this is the this comes back to you can't write all the software that an AI will automatically do this through its nature as understanding its job.

00:16:47:22 - 00:17:07:15 Andrew And in that case, as you described it started flagging. Hey, you replaced these the filter or these light bulbs or these batteries and they're not showing up in the invoice to the customer. You're leaving money on the table. And that was because it was easy for the field text to just grab stuff out of the supply closet, stick it in there and they're busy.

00:17:07:16 - 00:17:34:10 Andrew They're busy. So that was a great unintended consequence. But I would use that as a, as a there's a bigger pattern that we see, which is oftentimes our clients are going through a kind of a multi-stage AI evolution. It starts with an automation.

That's an easy thing for people to understand. They know that if humans are doing a high volume task that requires little intelligence, AI can potentially do that. 00:17:34:13 - 00:17:57:12 Andrew Almost universally, though, you're getting now a data set into a system that that either in that case, which was a little bit unexpected but beneficial. But certainly no matter what, you will have a new data set that you can you can apply some additional thoughts to that.

That same. 00:17:57:14 - 00:18:24:13 Andrew Forklift management company that I described earlier, you know, their data set was every single repair and fix and maintenance task on every single, you know, forklift or pallet jack or pallet mover across hundreds and thousands of vendors. Oh, that's a lot of data to understand. Are we being overcharged?

Are the are we getting hit for towels when the contract doesn't allow us to do that? 00:18:24:13 - 00:18:53:20 Andrew So generally we want to take clients on a journey of saying get the data into a system. You don't know yet what the unexpected opportunities of using that data in a new and unique way might be. It could be, you know, cost savings.

It could be revenue opportunities. We see a lot of times the more data we get into a system, sales team members get Ahold of that and say, gosh, if I could get a view of that data, I look like the smartest person in the room. 00:18:53:20 - 00:18:58:04 Andrew And so you suddenly have a revenue driving opportunity as well. Yeah.

00:18:58:07 - 00:19:21:26 Michael Yeah. I mean, that's a that's a great example to what we were talking about before about this hybrid workforce idea. Right? I mean, the the collaborative effort is producing things that are well beyond the original scope, because you do have the capability to look across a broader set of data and analyze it in much shorter periods of time and in context, and really learn from that.

00:19:21:26 - 00:19:29:06 Michael And and then it's a back and forth between, you know, human and agent as they, as they work through those problems. 00:19:29:10 - 00:20:06:15 Andrew Yeah, absolutely. And, you know, that can range from, hey, we want to we want to take a data set and do prediction. And we didn't really think about why some of the data might be moving.

One of our clients is a is a physical security services organization that literally staff unarmed and armed guards at your 7-Eleven or your bank, etc. and, you know, they were trying to look at what what were some of the trends between 2024 and 2025 that kind of drove security services. 00:20:06:15 - 00:20:29:22 Andrew And one of the big trends was there were a lot of hurricanes in 2024 that required emergency security services because, you know, power was out and facilities were damaged, etc.

, and there were not a lot of hurricanes last year in 2025. Well, so now they're starting to think about, okay, how do we predict into 2026. You know where I just saw there'll be six major Atlantic hurricanes. 00:20:29:22 - 00:20:47:27 Andrew This is what they're forecasting this year.

Who knows if they're right or not. But you know the point is now they are looking at things that go far beyond crime rates and and and employment rate statistics, which are kind of obvious to humans to some of these other things that might not have been obvious as major drivers of the business. 00:20:47:28 - 00:20:59:06 Andrew And I think that's exactly right. If you if you invite the AI to to look across a span of data in unique ways, you can be surprised in some of that, you know, dynamic thinking.

00:20:59:08 - 00:21:19:03 Michael Yeah, interesting. I think it is a good example of of the power of the combined effort. And, you know, the human doing what they do and the and the digital work are doing what they do. And bringing that together is really powerful for sure.

Yeah. You know, you talked about a little bit before when we were talking a little bit about human in the loop. 00:21:19:03 - 00:21:36:25 Michael And, and I have some very particular ideas around that to I've thought about this a lot and I think, you know, the, the in the proof of concept world human in the loop. That makes a lot of sense.

Right. Because I don't really have the trust yet. I'm trying something. I want to see if it works, that sort of thing.

00:21:36:25 - 00:22:09:01 Michael But when you start to scale systems up, having humans have to approve everything that a machine can do at scale, that this doesn't seem reasonable for most processes, right? Yeah. You know, once we start talking about that progression to human in the lead, you know, what does that progression look like for for the companies that you work with and, and what has to change organizationally to, to adapt to those stages as it moves through them. 00:22:09:03 - 00:22:42:28 Andrew Yeah, I think the important it's very important early on to lay out with a, with a customer who's going on this journey, why their what the role is that they're playing in the stage of the rollout of the system.

Right. There is a very big difference between early kind of almost pre-production checking the work of the system. Did we catch all of the, you know, the unexpected cases and, and kind of strengthening the borders of the system? 00:22:43:01 - 00:23:16:02 Andrew You know, one of the first invoices that we processed were in production for, for one of the clients was Canadian and had Canadian tax implications.

None of the test data set that we had received had Canadian invoices on it. And it's like, oh, okay, great. That's a that was an edge case that we didn't expect. And so so you might be starting with, you know, we're seeking to strengthen the edge cases of this of this solution because you can't know going in necessarily, especially a company that's processing a large volume of information.

00:23:16:04 - 00:23:46:16 Andrew You know what all those different edge cases are going to be. We work with customs brokerage company that helps import products from across the globe. You know, the quality of data, the the use of data, the consistency of data across tens of thousands of importers can be can vary. Right.

And so you're you might initially be around let's let's strengthen the edges of this system I think then you're moving to how quickly can we get comfortable with portion. 00:23:46:18 - 00:24:05:01 Andrew You know fractional automation as an example. These are the use cases that we are simply going to automate. We're comfortable.

We've seen enough data. We know that this is processing correctly. So then you're moving into a fractional automation. And then you know for us there might be an argument that says you're never going to get to 100%.

It's just not possible. 00:24:05:02 - 00:24:31:13 Andrew But if you can get to 90 or 95%, then the the the remainder exists because there's probably real decision making that has to go on there. And it's and it might be it might be deeply nuanced. I think the, the, the point that we try to make in this conversation with the customer is it is not an endless first period where you're helping tighten the edge cases.

00:24:31:13 - 00:24:59:12 Andrew That's a 3060. There's there's a defined period during which that is happening. And then you are saying, let's not chase edges at this point. Let's chase automation in large chunks.

This group of data can go through, this group of data can go through, you know, as an example with the with the commercial or the the real estate manager, the the approvals can go through. 00:24:59:14 - 00:25:18:24 Andrew They still want to touch the ones that get rejected because as an example, it might require somebody to go back out and and take new pictures. You know that's not something an AI. I mean, you could fabricate a picture, but that's that would be disingenuous in terms of the, the result.

And so I think it's it's it is you are helping us refine the system. 00:25:19:00 - 00:25:39:04 Andrew You are helping us determine where the system is eligible for automation now and then. Ultimately, you are looking for us. You're playing a role to help us add more insight into the system in a way that makes it even more powerful or more intelligent than it had been.

And I think that those stages work very well for for mid-market companies especially. 00:25:39:07 - 00:26:01:25 Michael Yeah, that makes sense. There is this, you know, there's certainly processes that are never going to fit underneath that fully autonomous umbrella because of risk or, you know, risk mitigation, that sort of thing. But a lot of, a lot of the things that are processes you undertake there can move to that automated phase at some point and provide additional benefit.

00:26:01:26 - 00:26:02:14 Michael Yeah. 00:26:02:20 - 00:26:26:12 Andrew We do. One of our clients is an engineering firm, and they have experts who go take pictures of, you know, the, the, the, the bowels of a building and the fire doors and the egress and how pipes go through walls. And the system does a really good job based on a significant amount of their historic proprietary data.

Right. 00:26:26:13 - 00:27:01:08 Andrew There's a ton of their inspection results over X years, you know, vectorized and and and pre-processed and made available to this system so the system can spit out a here's our from a bunch of pictures. Here's our analysis of building code violations. Condominium association as built inconsistencies, etc.

, etc.. You're still going to want somebody, I mean, an engineer with a certification who is putting their name on it is going to take a look at that information. 00:27:01:08 - 00:27:23:12 Andrew But you might dramatically accelerate the extraction, the analysis, the organization of that data in a way that their job gets a lot easier. I think in those cases we might eventually, you know, transition to to full automation.

But, you know, if somebody's got a liability that they are taking for for signing off on this, there's still a step. 00:27:23:14 - 00:27:48:10 Michael Yeah. No, that that makes sense. You know, one of the, one of the big topics this year related to that, that that we've been looking at is is governance.

And and you know, I've come to the conclusion that governance really does have to be built in from the front or governance by design, as I've called it, not slapped on at the back like perhaps we've done with, you know, a lot of governance in the past. 00:27:48:12 - 00:28:20:19 Michael And I know you've talked about hard constraints, soft constraints, gating, you know, different, different parts of the control architecture that you need to put in place from a governance perspective for AI, you know, to justify decisions before you put it in production.

How do you approach that when you're to point AI in a mid-market company that, you know, may not have really had a governance team, or definitely not a dedicated AI governance team in the past? 00:28:20:21 - 00:28:51:03 Andrew Yeah, I mean, there's there's a lot of ways that we seek to unpack, you know, where governance needs to layer in as part of our conversation. So, you know, one, of course, very early is like data, right? We need to the risk of using data in a certain way, the risk of touching data in a certain way, anything, you know, proprietary, confidential and how that is being is used is really important.

00:28:51:04 - 00:29:20:06 Andrew We will tend we tend to do, you know, some non-trivial pre-work on data transformation that might be in our system or it might be in a client's existing system in order to, you know, anonymize data where you don't need it. Right. It's easy to just pump anything into a system like. But but that's not necessarily what's needed for the system to be intelligent.

00:29:20:08 - 00:29:24:21 Andrew We will often. 00:29:24:24 - 00:29:50:00 Andrew When we're when we are writing to existing enterprise systems, you know, Salesforce as an example, it's very common that our solutions are writing to new objects that are not the core object of the of the, you know, the business itself, so that they can put an additional layer of checking in their core system before that enters, you know, a production workflow. 00:29:50:00 - 00:30:13:06 Andrew But I think it's it's, you know, the conversations that we tend to have are first, you know, we tend to use the hyperscalers as, as in our solutions.

Right? We believe there's a significant benefit to our customers in, in the fact that, you know, clouds making an announcement every single day. It seems like these days with new capabilities that add value in these solutions. 00:30:13:06 - 00:30:39:13 Andrew So we tend to use the first we have to get comfortable.

What does that mean from an API standpoint for your data being provided into, you know, that solution? If we are going to do some, you know, either fine tuning to a model or in context, learning to a model, what are we willing to give that solution versus what should be deterministic code after, you know, after the fact? 00:30:39:13 - 00:31:07:00 Andrew And I think all of that at that data layer is is perhaps the most foundational element of, of our governance approach. And then the hard constraints, soft constraints are going to come in where like you can't touch this source data, you're not even close to this source data versus, oh, you know, we're okay with you pulling that and wrapping it and maybe responding, you know, back and augmenting in a way.

00:31:07:02 - 00:31:33:14 Andrew You know, I think the our general message to especially mid-market companies is you have all of this proprietary knowledge that has huge value, but you're not getting the benefit from it now. So let's not over index on security. They're not to be we don't want to be. Of course we want to be secure, but we don't want to have a pattern, which is we're so tight with our with our data that we're not going to be willing to get the benefit from it.

00:31:33:14 - 00:31:36:19 Andrew I think that's actually a misstep for these companies right now. 00:31:36:24 - 00:31:57:09 Michael Yeah. I mean, I definitely like the idea that there there are things when you're defining the process and how the agents are going to, you know, work inside of this. It's it's it's almost like this three dimensional space at an airport.

When an airplane comes in, you're you've got an air traffic controller that keeps them in the zone. 00:31:57:09 - 00:32:13:07 Michael But if they're in the zone, you don't you just let them go, right? I mean, they're not they don't need anything. But if they get near the edge, you need to bring them back.

And then they're also have to be places that just simply can't. That's a boundary and you can't go beyond it. And if you make it so that it doesn't have access to that or you can't. 00:32:13:08 - 00:32:24:09 Michael It can't approve this, it can't do this, then that's that's a simple way to put governance in place that, you know, you can trust because it's it's predetermine some of those issues for you.

00:32:24:12 - 00:32:51:02 Andrew Yeah. And architectures are generally going to be very much separation of concerns architectures. Right. Heavily API driven across those separation of concerns.

So that, you know, this this agent as an example with these AI tools can touch, you know, this read only copy of our sandbox of our data. And that's all it can do to make an interpretation and pass that back to another, another system. 00:32:51:02 - 00:33:16:02 Andrew I think architecture decisions early can give you a lot of confidence and enforcement. And then, of course, you know, this can be an endless.

You know, turtles all the way down approach. But you can put, you know, kind of evaluators into the equation as well, you know, before an action gets taken, I'm going to evaluate this against a security stance that we have or a data integrity stance that we have before. 00:33:16:02 - 00:33:46:25 Andrew I allow it as an example. And again, for us that would be that might be a complete separation of concerns.

It's a you know, I want to make this decision. I'm going to pass it to an evaluator. And now the evaluator is going to say, yes, go execute this. The evaluator itself can't even execute it.

The you know, the final step in the in the equation can and you know, those sorts of expert or orchestrator expert models and evaluator models I think are are key to, you know, preventing an unexpected outcome. 00:33:46:27 - 00:33:56:26 Andrew You know, certainly our systems are not hey, one one agent can touch every one of your production systems and it's a single fire prompt. And good luck. That's not that's not how we are.

00:33:56:28 - 00:33:57:22 Michael Close your eyes. 00:33:57:22 - 00:33:59:20 Andrew And exactly, exactly. 00:33:59:21 - 00:34:22:25 Michael Yeah, yeah, yeah. You know, I like that.

I think one of the, one of the systems that I was looking at recently did something in governance that I really like, and I, and I think that applies to exactly what you're saying this idea of, I called it an interceptor. But, you know, it's a it's an agent that sits on top of a small model that's, you know, just for whatever that function might be. 00:34:22:25 - 00:34:49:06 Michael And, and you can build into that, that sort of singular mindset of, you know, here in our culture, we talk to our customers and do this for our customers, but then we don't do these things.

And you know, what's what's in, what's out. But, you know, things like that are very difficult. You can't just say to an AI agent, be nice to our customers because you're going to have some unintended consequences that you definitely didn't foresee, nor will. 00:34:49:08 - 00:34:51:10 Andrew Right?

Everyone gets a refund. 00:34:51:13 - 00:35:12:00 Michael Exactly. And you don't have to return a thing to us. We're happy to your money.

Yeah. So I mean, that really makes sense to me, I think to have these constraints as you design the system versus, oh, now I have a problem that happened at machine speed with 10,000 agents. I can try to fix it. But the problem is already so big that, you know, who knows?

00:35:12:01 - 00:35:40:10 Andrew It's yeah, I and I think as, as part of that this we talk about you know, I joke about single shots but people just try to, you know, cram a gazillion things into the, into the prompts for somebody to evaluate a lot of different factors. And we're much more we like to tune that down. Right. You know, we think we think there's there is certainly a sweet spot of giving us context.

00:35:40:10 - 00:36:01:09 Andrew So decision making can happen. But if you're doing an evaluation like these are these are our standards for what a photo must look like to demonstrate work in a, in a, you know, in a, in a home, you know, that's a very prescriptive, you know, step for us to break apart and just have that system be responsible for that. 00:36:01:12 - 00:36:33:12 Michael Yeah. That makes sense.

We were talking about one of your agents you called Digital Greg, which I like, by the way. I don't know why, but that's attractive. But anyway, the best commercial refrigerator estimator you'll ever find. You say, and it, you know, challenges the capturing of decades of all this institutional knowledge in a system.

But there are things that you just like these are our human expert gut feelings about things, too. 00:36:33:13 - 00:36:47:18 Michael I mean, there are some things that make sense to go in the in the agent and some things that don't. I mean, what's the what's the line between what you think AI can absorb versus what really does need that human touch when we're making decisions? 00:36:47:25 - 00:37:15:28 Andrew Yeah.

I mean, I think the, the, the this is very much the inside out disruption that we talked about. Right. The Greg that has been doing commercial refrigeration estimation for 25 years has a tremendous amount of knowledge. Oh, you never use that coupling with that, you know, pipe valve size.

It just, you know, you know, whatever. Oh, you know, we never do half inch pipe on ceilings because guess what? 00:37:16:00 - 00:37:53:15 Andrew People grab onto it and hold on to it. And you would not expect somebody to put 190 pounds on it, but they'll do it.

And therefore, you know, that's a learned thing. And that's the sort of like nuance and, and depth that that person learns over over 25 years. I think we're on we're kind of on a little bit of the bleeding edge right now with these sorts of projects is and our strategy right now is deep, long interview day in the life sitting, looking at, you know, many different examples of course, pumping examples into AI and having it do its analysis, but a lot of interview based discovery.

00:37:53:15 - 00:38:24:27 Andrew And you're seeing this, you know, more and more in the media, right, that the technology is less the challenge and the domain knowledge. And so how you capture domain knowledge is, is, is critical. What you're never going to know in a system like that is, oh, I actually happen to know that Susan, our engineer it's, it's, it's little league season and she really likes to get to her boys games on Tuesday and Thursday. 00:38:24:27 - 00:38:42:18 Andrew And so I'm going to bake that into the estimate in terms of time, because we're not going to be able to work overtime on those days.

That's stuff where I just I don't I don't see it right now. And I don't know that there's a mechanism for getting to that. And so I think you want to take the tedium out this valve, this joint, this, you know, coupling. 00:38:42:25 - 00:39:09:20 Andrew You want to get the rules that we use half inch pipe when it's going to be on a ceiling.

But you've got to allow for a human to say, I happen to know this team. And, and so-and-so has been sick for a couple of weeks. And we're going to we're going to account for that, or, you know, even some things of, you know, we were looking at examples where we got to bring the crane in and lift the condenser off the roof, but there's this fixed thing here that's going to make the crane a little hard to position. 00:39:09:20 - 00:39:20:27 Andrew And so we're going to actually take a little bit longer.

That comes from experience. And I'm not positive that we're we're at a place right now that AI can can interpret that. 00:39:21:00 - 00:39:42:18 Michael And yeah, that that to me that does seem like a logical line for now. And it's not a fixed line, right.

It's a moving line. Because certainly as systems learn more, your system will actually pick up some of those things over time and context and memory. Memory, obviously the important piece there. But at the same time, that's a long process.

00:39:42:18 - 00:39:48:13 Michael That's not a you're going to go in and interview everybody and capture everything they know tomorrow kind of a thing. 00:39:48:13 - 00:40:15:25 Andrew So and I think as part of that, that, you know, the message to companies who might think, oh, well, we want to wait until it's possible to do all that. That's a mistake. Because as you just described, there are tools for learning.

There are tools for managing memory and context that that system has where the earlier you get people using them and providing that feedback and clarifying why a decision was made in this way. 00:40:15:27 - 00:40:23:02 Andrew You know, the the, the, the, the better off you are for having, you know, that come along in the evolution of the system. 00:40:23:04 - 00:40:44:16 Michael Yeah. That that makes sense.

It certainly embeds itself. And you collect that knowledge over time. It makes it makes sense that you're growing that knowledge base. You know for mid-market companies a lot of times when they as a growth strategy, they do acquisitions and, you know, roll ups and that sort of thing.

And obviously private equity involved in that too. 00:40:44:19 - 00:41:03:22 Michael But you end up with a whole bunch of systems. I know I was talking to a CIO recently about his CRM systems and, and how they could get a complete picture of the customer. And pretty much at the end of the conversation, we decided that they couldn't because they had 33 different systems that were complete silos.

Right? Yeah. 00:41:03:27 - 00:41:26:03 Michael And it's and it's a nightmare. But and historically, I mean, that would be a in some cases a project you couldn't even really complete.

Right. But how does they change that? I mean, what does it what is the bridge look like now that you can build this, you know, agenda layer that can help you with that problem? 00:41:26:06 - 00:41:48:18 Andrew Yeah.

I mean, and you nailed it, right. Especially in PE backed businesses where P is buying a platform and then adding on to it over time, you're going to run into system spread and system system scope. In fact, I was just before this on a call and new acquisition. We've got our CRM here.

This company is on HubSpot. 00:41:48:19 - 00:42:13:26 Andrew How are we going to how are we going to rationalize that? Are we doing a migration or can we create a common layer? And I think we would argue that an agent layer can normalize relatively well across disparate, you know, formats of data that an agent layer can normalize.

This client is the same as this client, even if the names are different. 00:42:13:26 - 00:42:28:13 Andrew And then it really is, you know, again, kind of breaking into a separation of concerns is, hey, your first job is to at least figure out what's the how do I how do I know that this customer is this customer is this customer across these, these different platforms? And then what am I trying to get out of it? 00:42:28:13 - 00:42:53:10 Andrew I'm trying to get, you know, a common view of, of of growth or history or contact or contact movement, whatever that might, might be.

And so I we're big believers in don't be shy about dumping a lot of different data and a lot of different data formats into an AI system. You can test that stuff very fast to figure out if it can, you know, you know, sort it out. 00:42:53:10 - 00:43:10:27 Andrew And and then it's a question of like, well, so what are you trying to get? Are you trying to get historic reporting?

Fine. Maybe then you have an agent pipeline that's just doing your transformation to get it into a data set, into a, you know, a data lake or a data warehouse or a BI tool, and you can do your reporting or you looking to do something else. 00:43:10:28 - 00:43:26:25 Andrew A lot of times we would argue what you're really looking to do is take the data you have across these different systems, enrich it with other data that you might never, never have had a chance to pull in and make predictions or see around the corner a little bit more than historically you could.

00:43:26:27 - 00:43:51:24 Michael Yeah. I mean, it is it is an argument for how much automation can help in areas that we aren't really maybe looking at. When you think of bringing it into the business, you're thinking about the, you know, those those cool workflows that you can do. But but underneath that, the automation around the data layer can have tremendous effect for the business.

00:43:52:00 - 00:44:00:09 Michael And even though it's kind of under the covers, that could be a real transformation for them in and of itself. 00:44:00:12 - 00:44:27:13 Andrew It reminds me of a client that I was working with years ago. This is a little bit before GPT was available. It was not AI addressable.

And they're a field services, snow removal, landscaping for commercial real estate company, and they receive work orders from a gazillion different systems that their clients use carry for work order management or service channel for work motor management. 00:44:27:13 - 00:44:52:26 Andrew These third party systems that your big you know Lowe's keys in the request to have something done at their facility. And the challenge then was this massive reject attempt to figure out what are they asking to have done, and how does that correlate to a team and a task for the company?

Because, you know, I said, hey, I want you to remove the shrubs, okay? 00:44:52:27 - 00:45:26:00 Andrew That's a landscaping request. That's this. Or, you know, hey, the branches are a problem.

Okay? Branches. That's, you know, kind of the same thing. And you had all these different variations with no guarantee that you were going to get the information in a way that you could translate to a readily understood task.

And as I think about the effort to do that five years ago, that effort is would be solved in an instant today because you would just simply say, what we care about is snow removal or rice removal or landscaping or, you know, sweeping of of parking lots. 00:45:26:02 - 00:45:44:13 Andrew Tell us which this is. Right. And, and the LM would have no trouble doing that interpretation.

And so I think you're right, that data that data normalization is put aside. Workflow automation. Data normalization is is a huge value that people can get immediately in right now. 00:45:44:15 - 00:46:08:08 Michael Yeah.

And just being able to tie all those layers together after all these years of, you know, SaaS solutions and credit cards, it's right. Managing all that is complicated. Well, you know, for, for for a executive CEO in a mid-market company or private equity operating partner that might be listening. And, you know, there's a lot of pressure to, to to act.

00:46:08:10 - 00:46:30:25 Michael The boards are pushing. I mean, I'm on a couple of boards. We're talking about it at every meeting, you know, how do we get more of this into the business. So there's a lot of pressure.

But what advice would you have for, you know, for, for that executive about taking their first step without, you know, trying to boil the ocean and it's and scariness I guess. 00:46:30:27 - 00:47:03:10 Andrew Yeah. I think I'm going to answer in a couple of ways. So certainly if, if within your organization you see multiple humans doing the same information task, that's an opportunity.

Go, go go tackle that right. If you see a super high volume task where speed to respond with intelligence is is a differentiation for you, go tackle that. And one of our conversations was with a company that does disaster recovery in homes. 00:47:03:10 - 00:47:32:08 Andrew You had a flood or a fire or something like that.

The faster they can respond with a quote, the more likely they are going to win that business. And so that would be kind of a high velocity task for us or a data insight. If you felt like if you had just a little more, a different view of your data, maybe that data enriched with a little bit more information, you could make better targeting decisions or staffing decisions or resourcing decisions. 00:47:32:10 - 00:47:55:00 Andrew You know, those are those would be like where I would try to start.

It is very common that it's an automation, like I said, automating that, you know, high volume information task or that high velocity tasked with the data assembly to feed that. I would say if you have a bunch of employees who are like, I've spent four hours a week doing this and eight hours a month doing that, those are tough. 00:47:55:01 - 00:48:23:09 Andrew Those are those are going to be tough to build there. There are simple workflow tools that you can use to try to tackle those, but you're ROI is going to be a little less meaningful.

And so you're not going to drive Alpha if you're a portfolio of a PE firm necessarily for that. So I think I think that don't be afraid to take your business and jam it into an LM and say, you know, where should we start these these tools will give you a lot of good insight. 00:48:23:09 - 00:48:49:13 Andrew And certainly we have tools that are kind of what we think of as our key themes, and we do that all the time. I would choose a champion within the organization, somebody who is going to be the the internal cheerleader for that, for that solution.

Because especially if you try to roll something out to employees who are feeling a little threatened or feeling that this is this is this is not where I want to go. 00:48:49:20 - 00:49:01:10 Andrew They're going to find reasons to hate it and dislike it and undermine it. And so you got to have that internal champion to, to to see exactly what's happening and, and turn that into a success. 00:49:01:13 - 00:49:33:12 Michael Yeah.

That makes sense. Well, so really interesting conversation and very actionable, which, you know, I think is is probably the best benefit of any podcast you could do is you can learn something you can actually do something with that's, you know, serious value. But but before I let you go, one thing I'd like to ask at the end, you know, just for the for the benefit of the audience, could you recommend somebody, you know, thought leader and author or somebody you follow, somebody that you think the audience would enjoy and and learn from?

00:49:33:14 - 00:49:39:22 Andrew Well, I was going to pull up your book, but I felt like that would seem a little too staged. 00:49:39:25 - 00:49:41:27 Michael It might have seemed a little stage. I appreciate. 00:49:42:04 - 00:50:07:06 Andrew I know you and I talked about this.

I have actually recently been reading Walden by Thoreau, and the reason I have been doing that is, I think now with our post Social media advent of AI, the lessons that you can take from that book are very different than when we all read it when we were, I don't know, 15 and in, in, in middle school, right when you hadn't lived life. 00:50:07:07 - 00:50:29:14 Andrew And certainly for us it was pretty a lot of these technology advancements. It is there's a lot of fear that is happening in the industry right now.

And I think if you read something written in the mid 1800s and and see some of the fears and concerns he was expressing, but realize that we're here, we are almost 200 years later and it's okay. 00:50:29:15 - 00:50:56:03 Andrew The world has continued. We have evolved. Certainly there are challenges, but there are, you know, there are there are great benefits as well.

I mean, he died of tuberculosis at like 49. So you could you could point to some advantages of, of of technology. But I do think the piece in that book, it's okay. Calm down.

Take a take a beat in our modern world is helpful to me now at 51 years old. 00:50:56:03 - 00:51:02:16 Andrew But, you know, maybe it wasn't when I was 15, as I said. So. And it's an easy read.

You can break it up and have it over your coffee. 00:51:02:21 - 00:51:14:22 Michael Yeah. No. That's great, a great book.

I should reread that. It's been quite a while. Although I will say when you when you first said it, the first thought that came into my mind is there is some argument for going to live somewhere on a remote lake. 00:51:14:28 - 00:51:41:12 Andrew But I well and of course I am being a being a child of the current age.

I actually have a long running GPT thread where I'm talking to it as if it's my book club about it. And you know, the GPT is pretty aware that current, you know, the current stress of social media and and our global order is can be stressful and that we might all want to go to a lake. 00:51:41:12 - 00:51:44:02 Andrew So it has that that instinct. 00:51:44:07 - 00:51:53:18 Michael I like it, I like it.

Well Andrew, thanks again. So we appreciate it. Really interesting conversation and I know the audience got a lot of benefit from it. 00:51:53:25 - 00:51:58:06 Andrew Absolutely.

Thanks for having me on. 00:51:58:08 - 00:52:24:04 Michael And that's the show for this week. Thank you all for joining us. Remember to like, share and subscribe to the show.

If you enjoy the show, please leave us a review to help others find us. For more research on AI and other software, check out arionresearch.com. If you're an expert in AI, generative AI, or business automation, either as a provider or an end user, email your information to disambiguation at arionresearch.

com and don't forget to join us next week. Disambiguation is an Arion Research production. I'm Michael Fauscette and this is the disambiguation podcast.