Governance Is Functions: Why Your AI Won't Scale Without Discipline by Design

00:00:11:06 - 00:00:32:09 Michael Welcome to this ambiguity. I'm your host, Michael Fossett. 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. 00:00:32:12 - 00:00:43:24 Michael I show today is governance is functions.

While your AI won't scale without discipline by design. I'm joined by Chris Morancie, fractional CTO and founder of Digital Operations Factory. Chris, welcome. 00:00:43:26 - 00:00:46:20 Chris Oh, thanks, Michael.

Thanks for having me on. 00:00:46:22 - 00:01:10:18 Michael So just to get us started, can you tell us a little bit about yourself? I know you have a really interesting background, you know, from computer information systems, MBA and business strategy. You know, now you're working towards data science and you built workforce development programs, digital operations factory, fractional CTO, helping companies get AI in production.

So, you know interesting path. 00:01:10:18 - 00:01:15:24 Michael And so how did that unfold for you. And you know, just tell us a little bit about kind of what you're doing. 00:01:15:27 - 00:01:33:15 Chris Yeah, I mean I grew up so I grew up in the Caribbean, I grew up I grew up in a really, you know, poor family.

And my dad dropped for the school at 12, got into business and, you know, had uncles and aunts. They got into business. And business is what got us out of poverty. So I kind of grew up kind of going like, business is what actually do does it.

00:01:33:15 - 00:01:50:06 Chris Right. I mean, no disrespect to anyone, but it's maybe not 9 to 5. Most of the time it's going to be, you know, you're going to launch something, you're going to do business. So I kind of followed in the family path, which was kind of like really natural to just be in business.

But I kind of like, you know, became a little black sheep because I love tech at that time. 00:01:50:07 - 00:02:16:04 Chris Right. It's like early it's the early 90s. I was born in 1981, so early 90s, you know, computers, all.

I'm like, I want to be. I want to get into tech. Right. So really got into that.

When did my bachelor's in computer information systems, you know, but again at heart entrepreneurial because just grew up in that environment. So one of the really interesting things I understand from a really, really early age that it's just tools, everything around you is just a tool. 00:02:16:06 - 00:02:35:06 Chris Right? So, you know, computers, information systems, right?

AI right. No AI, just a tool. Right. So I have really always kind of kept this lens that it's all about the business.

It's all about doing things. So value and it's all about what is the best tool to kind of get you that value. And that's kind of like how I operate from. 00:02:35:06 - 00:02:58:28 Chris So I purposely did this by bachelors in computer information systems, went back and did my MBA, went back and worked my my master's in data science.

And throughout my career I've always done like Google Pops, right? I've always I've always had one thing business really going on in terms of my learning and my practice, and then one part that's always been technical and quality preparation called it luck, but I believe not when 2026. 00:02:58:28 - 00:03:15:01 Chris It's the age of the builder now where you can combine technical with business and you can compress those two sort of like roles that used to be so dissimilar, and it works out really, really well.

So, you know, again, it's I would I would like to say it's been a purposeful path, but it's almost like it's been a calling. 00:03:15:02 - 00:03:17:28 Chris I've been I think I've been made for this time. 00:03:18:01 - 00:03:21:27 Michael I mean, in the end it's, you know, certainly better to be lucky, right? So either.

00:03:21:27 - 00:03:24:13 Chris Way, yes, you did good. 00:03:24:14 - 00:03:26:13 Speaker 3 Better be lucky in good. 00:03:26:15 - 00:03:48:03 Michael So I know you're in the middle of writing what you've called the definitive AI and production framework. And, you know, one of the design decisions you made was that scalability shouldn't be standalone in the framework, but it still it should be the DNA from day one.

So, you know, you had a you had a great line when we were talking about the episode design for scale. 00:03:48:04 - 00:04:03:10 Michael Make sure it doesn't break at scale and don't go broke it scale. So can you walk us through that thinking and you know and why? That third piece, especially the cost dimension, is one that I most companies just overlook today.

00:04:03:13 - 00:04:19:01 Chris Yeah. It's again, you going back to this business background, my dad was just telling me that people, you know, at the core, you know, most people are pessimistic, right. So you do something and you're like, what if it doesn't work? My dad just say, well, what if it does?

Like, are you ready? Are you ready for the success that if it does. 00:04:19:02 - 00:04:40:29 Chris And part of that is so part of that in the in the framework is born from that sort of like line of thinking is what happens if it works. Right.

So if I, if we go from like the reverse, let's take the last part as if it works, right. Are you going to go broke at scale? And what that means it's, you know, token, you know, like token responsibility, cost responsibility is a really big deal. 00:04:41:00 - 00:04:58:02 Chris It's something that we have to handle from step one.

Right? And what does that look like? That looks like hey, I we selecting the right model for the job. Sometimes it looks like hey, should we actually bombed the LMS for example, in AI meaning that, you know, what should they do.

Right? We know LMS are good at reasoning, right? 00:04:58:02 - 00:05:15:04 Chris So if we have a reasoning task, yes. But if it's a nonresident task and if it's a deterministic task, that's rule based.

And while we burn in tokens. Right. Well, we can run a simple, you know, sort of like deterministic function. Right.

The simple if else statement could do right. A simple matrix would do. Then why are we wasting tokens on an LLM. 00:05:15:04 - 00:05:36:04 Chris And all those decisions themselves will help you at skill not go broke.

Right. And one of the biggest things that you kind of align sort of like, you know, pricing strategies from, you know, in terms of what your AI supports in terms of a service, which as you line it to cost. One of the biggest things that you have to yourself is, what if we grow this thing where we have a million users? 00:05:36:04 - 00:05:55:09 Chris Will we be able to even afford to actually deliver the service?

So when you look at something like this, really interesting questions that are again, Jamin, to just having a really good, I should say, business hygiene, you start to understand that it's your responsibility as a leader to understand that when we bring in AI, can we afford it if it actually works? 00:05:55:10 - 00:06:13:00 Chris Right. And you kind of like work your what you know from okay, what happens if it's broke? Then you have to understand like what happens, right?

If it breaks right. As we scale, you know the classic skill in question, as we get on more users and more people actually start depending on it, is this thing built so that it doesn't break under the underneath the load. 00:06:13:00 - 00:06:33:14 Chris So yeah, you're going to design for it not to break. You're going to actually design for if it doesn't break and everybody's happy with it that you don't go broke, you know.

So those two things are like really, really important. And these things are so like central to a business running that I didn't think it was it didn't have justice to have it as a specific pillar. 00:06:33:15 - 00:06:54:26 Chris It had to be embodied throughout the DNA of AI. Because again, as I said, it's always business first, right?

It's all about value delivery first and the sustainability of the business first. So this cannot be like a second class citizen. It has to be just woven into the DNA of any sort of like, you know, architecture that you put into your company for AI. 00:06:54:28 - 00:07:17:12 Michael Yeah, I mean, that makes sense.

And, you know, we talked a lot. Obviously enterprise scale, you can absorb more mistakes, but we talk a lot about how this is a great equalizer for mid-market companies, small companies. And so I just wanted to take in that last comment. You know, the idea of, you know, will it scale and will it cost me too much?

00:07:17:13 - 00:07:37:01 Michael I mean, what do you see? You know, what are the most common mistakes? You know, like like maybe, you know, you're overengineered at you're running a model that's five times more expensive than you need, when you could have just run a small local model and would have done the job just fine, right? I mean, how should companies be benchmarking model selection from the start?

00:07:37:08 - 00:07:56:06 Chris So actually that's a really good question because actually we're looking at right now, I'm actually looking several engagements where that's actually a really big deal. Right. So one of the things that we kind of look at is number one and this is what I advise on, I say number one, go out and get the best model that gives you the best output right baseline. 00:07:56:07 - 00:08:09:22 Chris You go like okay cool.

What's this scenario? We know we have something that actually works. Then what you actually do is that you work backwards. And I like to call like find the Goldilocks zone.

Right? You kind of go like, all right, I'm going to go back to a to a cheaper model, a cheaper model, a cheaper model. 00:08:09:22 - 00:08:32:10 Chris And then you're trying to figure out is the policy still acceptable? Is the quality still acceptable.

And the the magic here is in the in how you answer the question is the quality is still acceptable. Most companies or most right implementers, what they're going to do is they're going to do assertion testing as, as i.e. binary good or bad, true or false good.

00:08:32:12 - 00:08:50:16 Chris Right? Like accepted or not accepted. But it's different. Right.

AI for the most part it's replacing. You're not well it's all in your your your knowledge workers. So what you need to do is do true evals. You have to put them in front of people and say, would you have accepted that response?

Would you have generated that response? 00:08:50:18 - 00:09:11:16 Chris Right. So I think a really strong grasp of the difference between assertion testing and evaluation testing is a really, really big deal. So you do different things like Elo testing and so forth and so on in front of users and we're like, hey, here's here's a input.

Here's an output out of these five outputs. If you're testing five models, which one seems the most natural and the most correct to you. 00:09:11:18 - 00:09:28:02 Chris Right. And that's a really, really simple way to kind of do it.

So again the lesson here is you're doing a lot of Black or Goldilocks testing, right. Trying to get to the to the right combination of, you know, affordable and acceptable. So that's that's a nice cheap and quick way to actually go about doing that. 00:09:28:03 - 00:09:46:07 Michael Yeah I mean that's interesting because I know, you know, when you go into a proof of concept and especially if it's if you're in a consulting environment and you're working with a company, you go to the proof of concept, you're really just trying to prove that it'll work.

And so the tendency is to, to, to, you know, kill, kill fleas with a sledgehammer. 00:09:46:08 - 00:09:46:13 Michael Right. 00:09:46:14 - 00:09:47:22 Speaker 3 You over engineer. 00:09:47:22 - 00:10:02:01 Michael It to the point that you know it's going to work.

But I guess the problem is, from what you just said, that it's that next step, right? The step of working backwards to what's the right size of all of the component parts to make sure that not only does it work, but it scales. 00:10:02:06 - 00:10:17:25 Chris Exactly. Yeah.

Because scales are only a function of capability. Scale is also a function of survivability, right? You know you're going to launch this thing. Yes, it's going to work.

But can your company survive it actually working? So you actually have to start answering this question somewhere from a very, very early standpoint. 00:10:17:26 - 00:10:19:20 Michael Yeah. Victim of your own success, I guess.

00:10:19:25 - 00:10:25:06 Chris Yes, yes. It's a problem that we all like to have, but you actually have to actually prepare for it just in case. 00:10:25:08 - 00:10:25:27 Speaker 3 Yeah, yeah. 00:10:25:28 - 00:11:03:20 Michael No, that makes sense.

Well, you know, over the last couple of years, I spent a lot of time talking about the digital workforce and digital employees. But but this year, because a lot more companies are rolling things out into production, I've been focusing a lot more on governance and what I've called governance by design. And I know when we chatted before, we seem to be in violent agreement that that's, you know, that's a really important way to think about it, especially, you know, designing governance in from from the start, you know, and I know in the past we kind of thought about in business, we sort of thought of governance is more like documentation or 00:11:03:20 - 00:11:25:10 Michael logs even.

Right. But AI governance is different, right? It's it has to execute. So can you, can you can you unpack that a little bit for us?

I mean, how do you see that and what does that mean for companies when they need to architect AI systems with governance from the start? 00:11:25:13 - 00:11:45:07 Chris Yeah, actually. So one of the one of the lines that I really, really love, and I hope, you know, the listeners actually kind of like remember this line and have a conversation with the team, you know, on Monday morning. On Monday morning conversation is if your governance cannot stop your model from doing something wrong in real time, then it's not governance, it's just documentation.

00:11:45:13 - 00:12:03:01 Chris Right, right. So and if you really unpack that statement, what that statement means is that it then means that your governance has to be baked in into your infrastructure. Right? It has to be baked into the architecture of what you're doing.

So again, not to put in the need for documentation, you know, that's obviously what it's going to start. 00:12:03:02 - 00:12:20:16 Chris You want to document things so that obviously you can hold people and systems accountable. Right. So the number one thing is obviously you want to have your policies right.

But one of the one of the coolest things about the policies is that they're written in such a way that they're not just necessarily written for humans, they are written in a way that they're also written for the agents. 00:12:20:18 - 00:12:41:13 Chris Right. So an example of what you can do is that you can have a policy agent that serves as a firewall right within your system, right. So whenever, you know, whenever an agent request to perform, you know, a refund on a customer because it has access to, you know, to stripe or whatever to actually do that, the policy agent goes like, no, no, no.

00:12:41:16 - 00:13:04:22 Chris Like, you know, you you will not collect to do that right now. You have other things itself like you can have actually you can build an approval workflows. You can build in like decision login. Right.

Traceability of observability is a really very big, big deal. I do a lot of work on the AWS platform. So there's some amazing services built in me like CloudWatch, you know, and different things like as you can see, what you're actually doing. 00:13:04:25 - 00:13:26:04 Chris You know, so again, there's certain things that you can actually do.

And as I said one of my favorite patterns is the firewall pattern. Right. Your understanding is that for every agent that you have, you have ingress. You know what's coming into it.

You have egress, what's going on over it. Right to the tools. Right. And to, you know, and different things if you need you want to put a firewall between all of these.

00:13:26:04 - 00:13:45:18 Chris And that firewall itself is the embodiment of your policy, right. So you can build these things in and, you know, just and just as a general rule, when it comes when it comes to governance, it's if if your agent needs to do something, it needs to assume a role. And that role needs to be a very short term role, you know, it only assumes that role when it needs to do it. 00:13:45:18 - 00:14:02:07 Chris And then there needs to be something that governs what actually approves that role, that role in real time, like through a short term token or something like that, you know, so those are the kind of things you have to build into infrastructure, build into your architecture, built into your design.

They can just be documents because, you know, the agents don't respect your documents. 00:14:02:07 - 00:14:05:26 Chris They don't know about them. If you if you set it up. 00:14:05:28 - 00:14:26:21 Michael Yeah.

I mean, yeah, that's interesting. And I think that's important because a lot of times we don't really think of it that way. Again, because we're thinking more after the factor. Reactionary.

Right. And what you're saying is you can't react to something that moves at machine speed. You have to have that engineered into the system itself so that you systematically limit behaviors that you don't want. 00:14:26:22 - 00:14:28:00 Speaker 3 Right, exactly.

I mean. 00:14:28:01 - 00:14:45:13 Michael I was I was in the event this week we were talking about it, and somebody gave this a story I thought was interesting. It's like, send your kids out in the backyard to play you don't, you know, you did put a fence around your backyard because you don't want them to run out in the street, but at the same time, you they may need to go out of the backyard. 00:14:45:13 - 00:14:51:15 Michael So you build a gate there, you close the gate and you lock it, but you've got an agent sitting in their hand in the key when they need.

00:14:51:15 - 00:14:52:08 Speaker 3 To go out. 00:14:52:09 - 00:14:57:02 Michael I need always need access, right? They just need access when they need it and they don't need it at other times. 00:14:57:02 - 00:14:57:09 Speaker 3 Right?

00:14:57:16 - 00:15:19:04 Chris You you you hit it on the head, you know, and you know, we can go so deep into this, right? That's why I am actually a really big fan of, you know, role based authentication. Right. And with access control, not just simple, just rule based.

I'm authentication. It's with access control and access control down to the point where you can have an agent assume a role just temporarily with a short term token. 00:15:19:04 - 00:15:34:28 Chris And then that token expires. But you have another agent or maybe a more deterministic function within your within your architecture that hands the the tier point, hands the key right and say, hey, hey kid, you can go outside, go open the door, but you can do this for 30s.

And I'm looking at you, right? 00:15:35:00 - 00:15:35:15 Speaker 3 I'm looking. 00:15:35:15 - 00:15:50:22 Chris At you, but I 40s is done. I'm going to invalidate that key.

You know what I'm saying? So that is really the right way to do things. And it's so Michael, this is stuff is so surprising because I, you know, I talk on LinkedIn about this all the time. Somebody is just as good old fashioned the software development hygiene.

00:15:50:24 - 00:16:09:03 Chris Right. Like good and just this good principles. But for some reason. Right.

AI has gotten a lot of us drunk, right where it's like, oh my gosh, we can do this. Don't forget this concept, this concept. They've existed for quite a while. You just need to apply them.

You know, that's that's kind of like the approach I take to it. 00:16:09:04 - 00:16:31:16 Michael Yeah. I mean, just because you can mean you should, right? I mean, that's a, that's a, that's a kind of an old same but a good one.

And you know, when we first started talking about MCP, you were talking about giving this, you know, continuous connection to the agent to do a thing or use a tool. But, but the fact is now when you think about laying out your system, designing the system, they don't always need access to that cool tool. 00:16:31:16 - 00:16:40:25 Michael They need access to that tool when they need to use it in the context of something you've already predefined that they could do right. But it's not a it's not an all always thing.

It's when you need. 00:16:40:25 - 00:16:41:08 Speaker 3 It thing. 00:16:41:09 - 00:17:00:13 Chris Is when you need it, it's a it's at least it's at least privilege just in time access. And I'm going to revoke you as soon as I know that you're done.

What should be done, you know, so you kind of, you know, prevent that. And I've become a really big fan of using the concept of like preventing replay attacks and all that kind of stuff, because that stuff can actually happen to as well. 00:17:00:13 - 00:17:28:24 Chris So, you know, all these things again, these are just really good principles. For example, from API development, right, that you can take and you can apply to agents.

You know, I tend to see agents from a really microservices lens myself, right? So a lot of the concepts from here kind of take them, you know, you kind of take them across because, you know, API security has been such a big thing for such a long time that people have done some good work on, then why not assume, you know, some of these security safeguards and policies and frameworks and apply them? 00:17:28:27 - 00:17:48:25 Michael Yeah, that makes sense to me. You know, for for the last couple of years, the biggest use case are the one we've talked about the most has been in customer facing roles like customer service, particularly.

Right. And and it seems to me that that's one of the areas where there's great return or great value, but there's also the greatest risk, right. 00:17:48:25 - 00:18:10:07 Michael Because you have all this brand risk around mistreating your customer in some way, or over serving your customer or, or even just representing your brand in a way that you wouldn't want that to be. And and when we were, we were talking before and I've been playing around with this and written a couple of things about it, and you mentioned it to this idea of real time semantic interceptors.

00:18:10:09 - 00:18:31:15 Michael You know, the idea that a secondary agent could be a policy layer and check every exchange to make sure that you're not violating whatever that, you know, whatever my culture is my my way of interacting with the customers and, and that, that that cost is, is negligible in the whole process. Right. Because you're, you're running a small model. 00:18:31:15 - 00:18:34:00 Speaker 3 You.

Exactly. Yeah. So. 00:18:34:01 - 00:18:43:09 Michael I mean, for leaders who haven't thought about that from an architectural pattern, what does that look like?

And, and why do you think that it's that important when you think about production deployments? 00:18:43:10 - 00:19:02:09 Chris Yeah, I think that's such a good question. And I think you're kind of touched on it. Right.

So again, a small language, an SLM, a small language model is very perfect for this role. It's a very specific role, extremely near policies. It's trained on, you know, on your do's and your don'ts. Right.

And you kind of like you just it just monitors and, you know, sort of like put governs. 00:19:02:10 - 00:19:39:27 Chris Right, right. What, what your main agents are actually doing, you know, at the end of the day, and I think we talked about it a little bit early on. Right.

The idea that you want to have a list privilege for this agents. And I think one of the things I really want to stick in the minds of everyone listening is this we have put so much emphasis on prompt writing, and we believe that if we write a prompt and the promises do not issue a customer a refund right of over $50, we want to assume that the agent will not ever do that. 00:19:40:01 - 00:20:01:16 Chris That is wrong, because sometimes that's going to go in in a direct competition to the we might have we might have developed the agent to do so.

If we tell the agent, hey, your customer service representative, your your goal is to keep your customers happy. It goes like, well, customers are happy for $50 refund, right? But at the same thing you do not do it. 00:20:01:21 - 00:20:10:06 Chris But the person is mad.

They're saying they're still mad, right? And I have the access to give them the refund. It's going to give them the refund. 00:20:10:08 - 00:20:10:26 Speaker 3 Right.

00:20:10:27 - 00:20:35:12 Chris So you know, something about, you know, promising to say you can do is ensuring that the architecture, the infrastructure does not allow them to write. And so part of that architecture and that design is something else, is maybe another agent that's a little bit more focused, a little bit more hyper focus that has a different role. That agent has a role which is don't violate the policy. 00:20:35:13 - 00:20:53:02 Chris It doesn't care about the about the customer happiness.

So now you kind of look at it and you're like, okay, cool, right? No. We actually having someone manage and overlook, you know, that the central agent so that it cannot cross the line. So those are the things that you have to think about.

And as you said, I mean that that costs nothing. 00:20:53:03 - 00:21:09:19 Chris A small language model running on an AWS server, right? It's going to cost you like 30 bucks a month, and that's going to be able to serve for and of transactions, right, every couple of hours. And again, nothing to be safe.

So I like to tell especially CEOs I like to tell them nothing to make you sleep at night, you know. 00:21:09:19 - 00:21:15:06 Chris So. And you can sleep more at night. Go ahead.

Let's do it. Right. So that's kind of like oh I see that. 00:21:15:08 - 00:21:15:26 Speaker 3 Yeah.

00:21:15:27 - 00:21:35:02 Michael I mean that makes sense to me because, you know, it's like it's almost like we say goal. You know agents are goal based. But the problem is if you set goals and they're competing goals, which one takes precedence. And unless you've already defined that in some way or in your, you know, in your example there, you've said, okay, you can't do this thing.

00:21:35:06 - 00:21:39:07 Michael So you know, yes, keep the customer happy, but not at all cost. 00:21:39:08 - 00:21:40:00 Speaker 3 Right. 00:21:40:04 - 00:21:51:18 Michael That's a that's a scale issue and a customer service issue. But you don't want to go too far to to get to any one goal.

If there are other competing priorities involved in that. 00:21:51:18 - 00:21:52:02 Speaker 3 That makes sense. 00:21:52:02 - 00:22:08:12 Chris And that's the thing right? I'm of the of the of natural language is is at fault here.

Right. Because the way that we communicate. Right. I have a nine year and I have a five year old and I believe that I communicate with them very, very clearly.

But apparently I do not, because sometimes the outcome I get is not what I anticipated. 00:22:08:12 - 00:22:10:15 Chris But when they explain to me what they heard, I go like. 00:22:10:16 - 00:22:14:00 Speaker 3 Oh, oh, oh, right. 00:22:14:02 - 00:22:32:01 Chris Maybe I did give you conflict and, you know, sort of like directions.

And that's where the mom comes in, right? Was the who's the firewall? Like, I don't think that's what your dad meant. Like, you know, so it's kind of like that same concept.

A lot of times I talk about AI like your kids, right? I mean, it's really, really, really, really smart kids that need some governance, right? 00:22:32:01 - 00:22:52:16 Chris Because sometimes they just what we think that we're saying is not what we're saying, at least from the receiving standpoint. You know, because agents are going to optimize for goals.

That's what they're going to do. That's what they've been built to do. They're going to reason themselves to optimize for goals. Right.

And I'm wondering about this. Michael is really quick, and I've realized this because I've had conversations. 00:22:52:19 - 00:23:03:02 Chris If you if that point doesn't hit you like a ton of bricks, it's me. Because you have you are not implementing.

00:23:03:04 - 00:23:20:22 Chris Agents that are central to your business yet. As soon as you start doing it, that statement is going to hit you like a ton of bricks. You're going to go like, oh yeah, that already happened. And because again, this advice is not for if you're writing a simple agent says, hey, review my review my grandma, and see if you know this capitalizations.

00:23:20:22 - 00:23:42:19 Chris We're talking about things where there's multiple steps in the workflow and multiple ways that an outcome can be satisfied. This is where the powerful reasoning of AI agents come in. But it's also a danger comes in because once you introduce reasoning, you introduce variance, or at least the possibility of different interpretation of a rule. 00:23:42:21 - 00:24:00:21 Michael Yeah, I mean that's in in the functions.

You know, we've thought of software systems as deterministic forever. And yet what we're saying is we have agents now that are probabilistic. So you've got to have deterministic guardrails. Or the outcomes aren't necessarily going to be what you intended.

00:24:00:24 - 00:24:01:23 Speaker 3 Yeah. Yeah. 00:24:01:26 - 00:24:22:00 Michael But what about security. So I mean you know we think of agents you know, it's just sort of human nature.

You humanize things. And so we start to think of having a conversation. You know, I noticed this morning I was working on something in Cloud Coworker. And at the end of it I said, thank you.

And I'm like, oh God, I'm being awfully polite to the machine. 00:24:22:00 - 00:24:24:20 Michael But maybe it doesn't really care if I said thank. 00:24:24:20 - 00:24:25:04 Speaker 3 You. 00:24:25:04 - 00:24:42:10 Michael But but it's easy to fall into that, right?

So, I mean, how do you deal with this when you're thinking about treating security boundaries like it's not just identity management, right. It's there are other things involved in this. So how do you how do you set these boundaries in a systems when you think of security. 00:24:42:16 - 00:25:08:18 Chris So that's a good question right.

So one of the things I'm going to just going to drive off what you said here, the starting point for most people, right, when they start thinking about agents is to treat as a person right. And that mental model, it's useful. Right? It's very useful.

But if you just stop via, you've already inherited all the wrong assumptions about speed, manipulation, resistance and judgment. 00:25:08:19 - 00:25:32:25 Chris Right? So that's again, and I know something maybe going to sound like a broken record here on this. Right.

But again, it's come down to a few things. Right? A human, for example, knows exactly when to to email a whole customer database. An agent will do that even as access.

And the prompt is ambiguous because the matter what people think, agents do not have really good judgment, right? 00:25:32:27 - 00:25:55:00 Chris Like they don't. Right? Or maybe not at all.

Reasoning is not judgment. Right? So again, so whenever you look at these, if again if you, you kind of like imagine it, right. You look at anytime for an agent to do its job, it needs to have memory long term short term working memory.

Right. You need to have access to tools. 00:25:55:01 - 00:26:15:16 Chris Right. You know it's going to reason, right.

So what you're doing is that you understand that at every point, at every one of these points is a security concern, right? And for example, explain to people like this, you know, like sometimes an agent needs information on your company to, to particular do a particular job. So we do rag we say okay cool. 00:26:15:18 - 00:26:38:20 Chris Going to bring in rag.

We're going to bring documents in so we can actually do a better job. Well someone can actually inject you know, someone can poison your rag. And when the rag information comes in, you didn't you didn't have you didn't have an injection. Right?

So even from your knowledge bases that actually feed into your agents, you know, that's that's a that's a security vector, right? 00:26:38:20 - 00:26:57:28 Chris So all I'm saying is that you have to look at every ingress and egress point of your agent, and you have to secure that because and then you also have to assume that something is going to go wrong. You have to assume that something will get in. Just go to security practices.

Right. So again, let's go back to the beginning of this question. 00:26:57:28 - 00:27:15:03 Chris I said we want to we want to treat agents as, as as as as humans. We need to expand that to understand, for example, that if I'm a human working in a in a high security capacity, guess what they're going to do to me?

They're going to look at every part of my life as to where I can be compromised. 00:27:15:06 - 00:27:33:14 Chris So I need to look at every part of an agent's being to understand where it's going to be compromised, where it goes, what touches it, what comes into it, who it talks to, write everything. So all I'm saying is expand the model, expand the model to understand so you can fully, truly understand how you you can be possibly compromised. 00:27:33:16 - 00:27:55:07 Michael Yeah.

I mean, I guess you don't really think about it until you start to think about, you know, risk surface area, you know. Where because you're you're talking about something that has this, you know, capability to have agency and to make decisions and, and be goal based. But that also opens up all these different attack vectors that perhaps we haven't really thought of in the past. 00:27:55:08 - 00:28:17:20 Chris And one of the ones that was really interesting that I saw, like, like a month ago.

And we're having a conversation, you know, in working with for health, with a clinic, it's one of the people don't even think about it's you give an agent a tool like, let's say, an API call to get data. What comes back? What if someone sends you back something that can be injected into the prompt, right. 00:28:17:21 - 00:28:39:16 Chris You're trusting a downstream provider, and if you just take in the data that they give you back and you're sticking it back to that agent, that then has to access a false and tools into your entire company.

That's a that's that. Again, the attack surface for this is just ridiculously wide. Right. So again you do need these firewalls every ingress every every egress point.

00:28:39:18 - 00:28:40:03 Speaker 3 Yeah. 00:28:40:04 - 00:29:00:16 Michael Yeah. That makes that makes sense. Well you mentioned something before we were talking about the episode about operational sovereignty.

And you know, there's been a lot of talk about data sovereignty, but I haven't really thought about it in the, in the under the guise of operational sovereignty. You know, if I'm dependent on an external model, I've, as you said, tools. 00:29:00:16 - 00:29:24:16 Michael I've got certain security issues. Countries think about national security and supply chain risk a lot.

Right. Well, what about businesses? How do you how do you evaluate the vendors for all of these different pieces of the model, the the MCP server, the downstream tools? What should I be thinking about?

I mean, what's what risk do I inherit? What kind of SLA. 00:29:24:19 - 00:29:32:15 Michael Do I need to put in place? And how do how do I, as a company, maintain control over my own operations or get this operation?

00:29:32:16 - 00:29:33:22 Speaker 3 Oh man, I. 00:29:33:25 - 00:29:58:15 Chris I love that question. I love the question. I love questions that are so current.

I just sent out an SLA, an email yesterday to a vendor. Right. So, you know, I'm working I'm consulting with a particular company. Great company, brilliant people love just love and working with them.

And one of the things I kind of stepped in and I looked at was, hey, they're bringing a QA. 00:29:58:16 - 00:30:21:08 Chris They bring they bring in an AI power tool right into their system. And I'm looking at it like, wait a minute. If this tool goes down, then what?

Right. So he ask yourself, you have a bunch of questions to ask yourself about it. That's really specific to AI, right? So different things like right now we know a lot of the model costs.

00:30:21:09 - 00:30:23:26 Chris A lot of the token cost is pretty much subsidized. 00:30:23:28 - 00:30:24:18 Speaker 3 Right. 00:30:24:19 - 00:30:50:00 Chris What happened when when when that when if that changes then all of your providers price is going to go up. How is it?

How does it affect your pricing. You're promising 99 point. That's 11 911 nines uptime right? I did promising the same thing.

What if right. Because if they don't then you're inheriting their uptime their uptime promise. 00:30:50:01 - 00:31:17:10 Chris Right. So you look at all those different things.

Another one that's really interesting from that I sent that email yesterday was every industry, especially education, health and law, you know, just regulations that govern them. I think, for example, education, you have further. Right. If you're if you're dependent on a on a provider.

Right. And they're trained up models to provide you data idea abiding by virtue. 00:31:17:10 - 00:31:44:00 Chris And if they don't with the data that dissent you then are you exposed. Is someone actually complains who gets sued.

Right. So you're looking at things like compliance regulations, like what are they doing? What are they doing. Up time.

You know what. What is their what is their recovery policies if something actually goes down right. You know, there's so many different things in here that you have to you have to ask because you're going to be assuming a lot of risk when AI is in place. 00:31:44:01 - 00:32:05:18 Chris Because again, this is pretty much a still nascent technology, right?

So I think the number one, if you're a leader here, listening to the number one question you need to ask is what if something goes wrong? Especially what if something goes wrong that's very specific to AI, like the model or the model access or you know, again we talked about attack vectors and the surface area. 00:32:05:18 - 00:32:26:01 Chris What did they get attacked. Right.

Are you inheriting those risk. And and the number one thing obviously is if I'm working with a, with a vendor who's providing an AI powered service or just the model itself, what? What am I assuming the uptime promises and all that kind of stuff? And does it matter what I'm promising my my users long stream?

00:32:26:01 - 00:32:47:07 Chris So again, this could be a very, very long form of operational sovereignty. Has everything done to the line of workforce development and upskilling. Like, do you want to be beholden to a consultant or a consulting team for the rest of your life? Are you developing the resources and how so you can actually be independent, like you know, you are you developing small language models and all that kind of things for just for privacy and security concerns?

00:32:47:08 - 00:33:01:01 Chris Like there's so many things that goes into operational severity. It is it is pillar ten of my framework, and it is easily my favorite pillar because it's where tech meets business. And so it's an intersection that most people don't actually think about. 00:33:01:03 - 00:33:33:14 Michael Yeah.

And I mean, it is different too, because we've we've had enterprise applications for a long time. And certainly there are some operational, you know, issues there. Right. There are some SLAs that I want in place.

But but it's very minimum when you think about the, the the risk. And then when you're talking about agents and multi-agent systems and the fact that they're all across your business and that they have agency, that changes that conversation a lot and makes it a much bigger risk and, and something that has to be dealt with. 00:33:33:16 - 00:33:53:27 Chris Yeah. You do.

And again, one of the things that you can do to arm yourself against that, and especially let's just say if you're looking at just from a pure connectivity standpoint, you're going to be working with a model, you're going to be working with, with a vendor who uses a model to pull a service, always put a line of abstraction right between you and that service, right. 00:33:53:28 - 00:34:16:00 Chris So in case you have to switch out, your downstream applications don't have to change everything the like that ripple is stopped at that firewall at that boundary.

So I always do that right. Are you always going with the assumption that this is going to fail? And I might have to switch this out really quick, right. So I don't want to tie coupling between an external AI powered service and, and my system.

00:34:16:00 - 00:34:26:06 Chris That depends on it. So that's a really just if you're looking for some tactical advice, if you're listening, that's one of the really easy things that you can actually do to actually keep you a little bit safe. 00:34:26:08 - 00:34:30:16 Michael Yeah, there's a lot more interdependency because of all the different types of connections to. 00:34:30:18 - 00:34:33:01 Speaker 3 So yeah, yeah.

I will say. 00:34:33:03 - 00:34:56:14 Chris I'll let me say one other thing, Michael, that I'm so sorry. Yeah, yeah. I remember the email that I sent yesterday to one of the concerns that I had was, hey, because we were providing we are providing data back to that provider so they can, you know, they can make the what they provide to us higher quality.

The question was, am I trading your model for you so you can actually provide that to other users? 00:34:56:15 - 00:35:19:07 Chris Like because part of the feedback that we give, we consider that to be IP. So is that going into your model, is that going to be baked into the weights of your model. And once it's baked into the weights of your model, you just have some of my IP that my competitors are going to benefit from.

So and then if you talk about operational sovereignty, you talk about survivability. 00:35:19:07 - 00:35:47:06 Chris And if your IP and the way that you do things is special, that makes you special and separate differs you from your competitors. Now they have your intelligence. Now, this service that's looking much, much better because they're using the same vendor like so we're not thinking about these things, right.

So you know, you have to kind of really go down into understanding, you know, that one of the one of the challenges of LMS is that once you train them with proprietary information, they are baked into the weights, and anyone who has access has access to it. 00:35:47:06 - 00:35:56:16 Chris So if you're using a vendor who is using like, you know, like a single tenant or like, you know, that kind of service, your competitors are using them to then going to get some of your intelligence.

00:35:56:19 - 00:35:57:24 Speaker 3 Yeah. 00:35:57:26 - 00:36:16:21 Michael You know, I think about it because, I mean, I mentioned before that I like to think of of agents as digital workers and, you know, our new world as a hybrid workforce or work environment. And if you think about agencies, workers, you know that that's it's different. And and yet, you know, you you still have to manage them.

00:36:16:21 - 00:36:38:18 Michael And, and I know when we were we were talking before I love this distinction that you made between a good manager and a bad manager. Bad manager, you know, tends to manage everything the same. But a good manager knows how to, you know, tailor that to each individual and, and their, their abilities or different abilities. But I mean, the same thing applies to agents, right?

00:36:38:19 - 00:36:47:25 Michael I mean, how should organizations think about how they build out this hybrid workforce of humans and agents and what's good management look like in a blended team? 00:36:48:01 - 00:37:13:02 Chris Yeah, that's a good conversation. But I can give a very simple way to kind of look at this. Right.

Everybody likes a good old fashioned quadrant, right. Like you know. So I'll use that. You can look at agents according to capability and risk.

Right. So and then what I like to say is risk is equal to trust. So how much do I trust the agent to do something. 00:37:13:02 - 00:37:32:24 Chris And then capability is tied to reasoning like the you know does the agent high reasoning or low reasoning.

So really this is all so you're going to put them in for buckets right. You go like do I trust the engine that I have to trust at a very high level to do its job? I have no I have I have no choice. 00:37:32:26 - 00:37:46:22 Chris Right.

But if I'm going to trust you to do this job and I have to trust you, there has to be a real reason why I have to trust you. It's because I need your reasoning capability. So what I usually like to see is if you're trusting an agent at a very high level, I want to see that. 00:37:46:22 - 00:38:04:15 Chris I want to see you at the top right upper right hand quadrant, because you have to give me a reason why you're going to give that agent that trust.

What are you getting back for, that trust that you're placing in. Right. And that's just like that in everyday life, you and I, as humans, we do this right. We place trust in people that are going to provide us something back.

00:38:04:15 - 00:38:22:26 Chris I trust my wife with my life, but she's my wife. She gives me joy, right? She, you know, you know, she's the mother of my kids, right? There's a reason that she has my trust.

But when I see things like why you're giving this agent all access to this tools and all this trust, but the agent is only doing a little bit of reasoning. 00:38:22:26 - 00:38:31:12 Chris I go like, what's the payoff? What are you getting in return for that? Right.

So it's almost like a it's almost like a capability ROI. Right? 00:38:31:13 - 00:38:32:10 Speaker 3 It's like I need. 00:38:32:10 - 00:38:50:08 Chris To see a high ROI if I'm going to give you a high trust.

Right. So I can you can put them in that fort, but in those four buckets. Right. You know, high trust.

Hi, hi, hi, hi. Reasoning right. High reasoning is low trust and so forth. Everyone everyone knows how to do a portrait.

Right? So I kind of look at these and I question anytime I see high trust. 00:38:50:13 - 00:39:00:27 Chris But the need for reasoning I gotta go like why? Why are we doing this.

We need to switch this into something else. So that's like a really easy mental model to kind of use. 00:39:01:00 - 00:39:20:09 Michael Yeah, I mean a risk reward kind of look, think of that in the relationship. What you know, I give you access to these things.

I get this output. But if but if that if they aren't, if the risk over weighs the outcome or the output, then why would I do it right. It's just it's too much risk. Yeah, that makes sense.

00:39:20:15 - 00:39:22:10 Speaker 3 And I think you know. 00:39:22:13 - 00:39:23:12 Chris Oh go ahead Michael. 00:39:23:19 - 00:39:24:15 Michael No no no it's okay. 00:39:24:15 - 00:39:25:03 Speaker 3 Good.

00:39:25:10 - 00:39:43:26 Chris I think the magic to to understanding this is to understand that the reason why you use an LLM, an AI, it's for its reasoning capability. And I think people like we really need to kind of get that. Like, if you ask me to say Chris in like 2 seconds or 5 seconds, tell me why you use an LM, I just say reasoning. 00:39:43:28 - 00:40:05:25 Chris That's it.

That's that's the Super Bowl. You know, I don't want it for anything else. I want it towards reasoning capability. All the things about connecting to tooling and APIs.

That's just what old fashioned integration. You know, I want something that reasons before they use the tool. And if I if I need that I need an LM. If I do not need reasoning, it's a simple if else logic, then I do not need an LM.

00:40:05:26 - 00:40:15:07 Chris Right. And that's just simply all you put it. So if you look at capability for the eyes of reasoning, it's a very simple thing to actually kind of understand. 00:40:15:08 - 00:40:15:24 Speaker 3 Yeah.

00:40:15:25 - 00:40:36:16 Michael I mean that's really been the shift to if you think about the when I first started using generative AI three years or so ago, it was very much a I put in something, I get something out right. There wasn't a lot of there wasn't a lot of reasoning involved in that. But now that we've moved to reasoning models and you can, you can you can even watch the thought process, right? 00:40:36:18 - 00:40:55:14 Michael I mean, it the outputs I get are considerably different.

And I'm going to give it a lot more things to do than I would have done in the past. So it makes sense if it's something that's grounded in the ability to reason. I want the LM involved, but if it's not, I need a deterministic system. I need to be able to just build the if thens and have them do it.

00:40:55:16 - 00:41:09:10 Chris It's so funny, as all you said that you said, I am giving my LMS more to do over time. What that is, is that you're trusting it more and you trust me more because you need more reasoning from it. So you kind of like when that upper right hand quadrant like, yeah, I trust you because I need more of your reasoning capability. 00:41:09:10 - 00:41:13:15 Chris But if that didn't happen, you wouldn't go there.

There would be no need to do it. 00:41:13:16 - 00:41:14:18 Speaker 3 No, no. 00:41:14:18 - 00:41:19:16 Michael And that's the capability of improved. But also my trust in their capabilities is approved.

Yeah. 00:41:19:18 - 00:41:20:07 Speaker 3 That's a. 00:41:20:07 - 00:41:37:06 Michael That's a fair point. You know if we if we look ahead a little bit and it's always kind of fun at the end of the show to think about, you know, what kind of advice should we give CIOs, technology leaders, business leaders who, you know, and in this case, they're trying to move from pilot to production.

What do you think? 00:41:37:12 - 00:41:43:15 Michael What do they need to focus on? What's the most important thing they need to get right first? 00:41:43:18 - 00:42:01:22 Chris Oh, that's a really good question.

I this is going to sound like a cop out. But I think it's a little bit different for every different situation. And I think you kind of alluded to it a little bit early on. You had a line that I really liked, and I'm going to I'm going to start using.

00:42:01:24 - 00:42:32:26 Chris Right. It's something to the effect of, you know, enterprise companies, they can they can stand sometimes to lose a few bucks to figure stuff out. Right. A meteor company cannot.

Right. So if I'm looking at a meteor company. Right. And don't get me wrong about what to say, security is always a big deal.

But if you are a company where a $300,000 investment on an AI initiative is going to ruin you and ruin your balance sheet, then the number one thing you start looking at is the the economics of what you're trying to do. 00:42:32:27 - 00:42:55:18 Chris Right? I would focus on survivability, right. You know, sustainability, right?

I would focus on, you know, the operational sovereignty because I cannot afford right to fail at a lot like that, that at that level, if I'm an enterprise, obviously, you know, security, I think it's going to matter a lot more. Right? I have a little bit of a buffer on the money side. 00:42:55:19 - 00:43:07:07 Chris Right?

I can I can experiment, right. I can try different things. Right. What I kind of feel at is I cannot feel that security because that's what's going to do.

That's what's going to destroy my stock price. That's what it's going to destroy public perception. 00:43:07:08 - 00:43:08:21 Speaker 3 Of me. Right.

00:43:08:24 - 00:43:34:02 Chris So that's kind of how I look at it, I guess what I say in different situations, and then I also deal with SMEs, like I did a really small companies to as well. And for them, what I really what I focus on for them is kind of like I kind of shepherd them through to ensuring that that the AI that they believe that the AI experiments that they do, I like to call the AI investments that they do are really aligned strategically with their grand goals.

00:43:34:02 - 00:43:52:07 Chris And the way that I do that is I say to them, you want to prioritize pinpoints UCI to solve your pain points first, right? And then the second thing that you want to do is that you want to use AI to actually increase your sort of like economic metrics. And the one that I actually look at for SMEs is revenue per employee. 00:43:52:08 - 00:44:07:01 Chris And so you can talk about everything that you want.

But if you're a small company, what you want to bring AI for info, if you're talking about revenue is revenue per employee. Can you go from $100,000 to employee to 200,000 and 300,000 and in the number the third thing I want them to do again, this is SMEs, right? 00:44:07:02 - 00:44:27:04 Chris So one is pinpoints to increase revenue. But number three is tell me what your grand goal is because no SMB wants to no small business wants to remain a small business.

They want to go into, you know, mid-market and they want to go bigger. And I see let's talk aspirational. Where do you want to go? How do you how do you increase the tent for your services.

00:44:27:06 - 00:44:41:21 Chris Right. So we can actually talk about what what AI can do to help you grow the tent for who can afford your services and where you can actually go. So it's a classic. What's that answer matrix.

Right. Like can you create new markets. Right. Can you go deeper in an existing market.

So that's when you put the MBA hat on. 00:44:41:22 - 00:44:57:20 Chris You go like what is your strategy for growing your business and where is going to be used. Right. Is it market training or is it creating new markets.

Is it creating your distribution. So again different rules for different market segments. And it just has to be that way. 00:44:57:22 - 00:45:25:02 Michael Yeah.

No that makes sense. It is contextual. And the matter you know what your use case or what your business focus is right. Well I mean great conversation Chris.

And unfortunately we're, you know, getting close to the end of the time together here. But but one of the things I always like to ask at the end is, is for a recommendation, you know, thought leader, an author, podcaster, something that you think would resonate and help the audience. 00:45:25:04 - 00:45:47:00 Chris Yeah, for me, that's not without a question. I like to give credit where credit is due.

Dell Research Labs criminally underrated, right? If you have time and you're in the car, you want to listen to something, go to YouTube. You know, search up the Dell Research labs, you know, channel and just listen, you know, it's, you know, great conversation there. 00:45:47:00 - 00:45:59:24 Chris Research into how you should really deploy AI in production is the best I've seen.

Just the best. And it doesn't hurt it. It's pretty pretty personable right. You know, for deep research for a deep research.

Right. You know. 00:45:59:25 - 00:46:01:04 Michael It's funny. 00:46:01:06 - 00:46:11:03 Chris It's fun conversations.

It might just be us, right. Like we might just like this stuff. But in my opinion, let us put it. You know how I measure this.

My wife doesn't ask me to put on the put off the television in. 00:46:11:03 - 00:46:15:01 Chris The room. Exactly. That's a good measure.

Make sense? Right. 00:46:15:03 - 00:46:33:06 Chris You just like this stuff is boring. How can you listen to this, like 24 over seven?

She doesn't say that for the daily research people because they're funny and they they make the information really digestible. So again, if you if you, if you're a leader listening to this and you want to, you know, you have a 15, 20 minute drive, you know, you know, you know, pop this stuff on and listen to it. 00:46:33:06 - 00:46:45:04 Chris It's really easy to consume. And if you're super technical, they do a good job of bringing it to as well.

So I want to give them a lot of credit to kind of shaping a lot of my thinking and some of the stuff that I actually take out in the while. 00:46:45:07 - 00:46:51:02 Michael That's great. Thanks, Chris. I really appreciate it.

And thanks for joining today. Very, very interesting conversation. 00:46:51:04 - 00:46:58:16 Chris No, I mean, this is the second time we've spoken. I enjoyed it both times.

And we could have we could do this for four hours and you'll have more stuff. 00:46:58:19 - 00:47:04:26 Michael I we probably lose the audience. I don't think we can probably hold the Joe Rogan three our conversation, but. 00:47:05:00 - 00:47:08:01 Chris But yeah, I.

00:47:08:03 - 00:47:08:21 Michael Thanks Chris. 00:47:08:21 - 00:47:09:19 Chris I appreciate it. 00:47:09:21 - 00:47:12:27 Chris Thanks, Michael. 00:47:13:00 - 00:47:37:04 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 Arion Research.

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 00:47:37:07 - 00:47:45:22 Michael Don't forget to join us next week. Disambiguation is an arion research production. I'm Michael Fauscette and this is the disambiguation podcast.