

"The things that cost $100 million at GE 15 years ago can now be done at a 15-person manufacturing business."
Garret's approach to value creation in lower middle market companies starts with upgrading the basics: Microsoft licensing, cloud storage, Azure services. But with AI tools like Claude, he can now deploy enterprise-grade solutions, like configuring Intune for a 30-person company—in two days instead of two months. The foundation has never been cheaper or faster to build.
"You need your editorial AI agents, your fact-checking AI agents. Break down roles and tasks into multiple agents."
The mental model that AI is a "magic black box" producing perfect outputs from generic inputs is the most overrated aspect of AI today. Garret argues that implementing AI agents should follow the same structures and processes used for humans. News organizations don't let reporters publish directly to the Wall Street Journal: there's an editorial process. AI systems need the same rigor.
"Every company needs an open claw strategy. Not tomorrow, but you need to be experimenting now."
Within a few years, Garret predicts it'll be normalized to have group chats where 60% of participants are agents. But that requires solving identity management first. Agents need distinct identities: their own Notion accounts, Gmail addresses, Slack profiles, not fake personas pretending to be human. The security infrastructure for this doesn't fully exist yet, but it's coming.
"Focus too much on ROI per use case, and you're missing the forest for the trees."
Garret separates AI investments into two categories: (1) LLM-based process automation, where token spend vs. ROI matters immediately, and (2) personal productivity and general-purpose agents, where you need a longer time horizon. The second category requires investment like any training program: you won't see quarterly returns, but year-over-year you'll double your business with flat headcount.
"People go, 'I'll just put the LLM in, it'll be better than a rules engine.' But you need to break that mental model."
One of Garret's most practical insights: use AI to build comprehensive rules-based systems, not to replace them with LLMs running on every transaction. LLMs can design, build, update, and monitor deterministic processes, but they don't need to be in the execution path burning tokens. The old mental model that rules engines are hard to build no longer applies.
"In a year, the distinction between CTO, CIO, and Chief Digital Officer doesn't matter. The job's always been the same."
As AI democratizes technical capability, leadership roles converge. The "ivory tower" IT leader is gone. You have to be both a great leader and an active builder, using AI tools at a granular level, not just strategically. If you're a CFO who's never done a VLOOKUP, how can you lead a finance team? If you're a CTO who hasn't done the prompting, how can you lead technology?
Stay tuned for more conversations on AI transformation, organizational evolution, and practical implementation strategies on CTO2CTO.
Host: Lucas Hendrich, CTO at Forte Group
Guest: Garret Fitzgerald, CTO at Envision Capital; incoming CIO at Salute
Lucas Hendrich: Welcome to the CTO to CTO podcast. I'm your host, Lucas Hendrich, CTO of Forte Group. Today we have the pleasure of having Garret Fitzgerald with us. He's currently CTO of Envision Capital, though I understand you're moving to another role soon. To give us some context — context potentially being the word of this decade — tell us a bit about what you do at Envision Capital and what you'll be doing very soon.
Garret Fitzgerald: Sure, thanks. By the time this goes out, I'll be the CIO at Salute, a data center services company owned by New Mountain Capital. So I'm still in the private equity and PE-backed company space.
For the past year, I've been CTO at Envision Capital, a private equity firm that invests in the lower middle market. I like to describe them as some of the more traditional, gritty businesses: manufacturing, B2B services. I own technology for the firm and work with all of its portfolio companies.
Lucas Hendrich: Mm-hmm.
Garret Fitzgerald: A huge part of my career was at GE, early on in IT roles and later running some of the software and electronics businesses there. For about the past five years, I've been in the private equity space, either with portfolio companies or with PE firms.
Lucas Hendrich: Great. I want to start with low- to mid-market companies that may not come from the most mature technology standpoint. They have solid product-market fit, whatever it is, but they're not digital. This is an expansive question, but I think it's prescient. There's a perception that everyone is running very fast with AI adoption across the enterprise, and there are all these ways to do it — every foundation model has its enterprise suite. But these aren't enterprises. They're smaller businesses, maybe even family-run. How do you approach AI enablement? Where are the use cases, what's the easiest path to value, and how do you track what good looks like at the portfolio level?
Garret Fitzgerald: I think the foundation for lower-middle-market, low-tech companies right now is the technology we've taken for granted for a decade in the enterprise: basic things like Microsoft licensing, email, cloud storage, Azure services. Those have never been cheaper or easier to implement than today. Things we spent $100 million on at GE 15 years ago can now be done at a 15-person, family-owned manufacturing business.
That's the first step. Much of what I look for in value creation at these smaller companies is to come in and very quickly upgrade the basics to create that foundation. Forgive the building analogy, but you have to pour a new footer and rebuild the foundation before you start talking about the bathroom faucets.
Lucas Hendrich: Mm-hmm.
Garret Fitzgerald: That's now really fast and easy to do, and AI actually has a role in it. For instance, Claude Code is very good at configuring Microsoft services. In some of these businesses, security and compliance concerns are realistically much lower risk, so you can use these newer tools to quickly understand their licensing, reconfigure it, and deploy new services.
To close this train of thought with an anecdote: the other day we were deploying Intune at a company. The typical approach — finding a tech firm or managed service provider, planning it out, and spending two months executing — isn't financially feasible for a 20- or 30-person lower-middle-market company. But when I can come in with Claude and do it myself in two days, it becomes very economical, and you get the same enterprise result for the smaller company. So a lot of it is building the right tech foundation.
Lucas Hendrich: I'm going to challenge this a bit. I totally agree and can see the value, especially if you're starting from nothing. That gives you a cleaner starting point than 20 or 30 years of technical debt and legacy architecture you have to change while the business keeps running. But do you see cases where it creates a comprehension gap? I know we're talking about configuring services, not writing thousands of lines of code, so it's a different situation. But it's a risk I see: the total cost of ownership may be higher over time than you think. Do you see that?
Garret Fitzgerald: Part of your point I agree with: the mental models for using the latest technology and AI aren't there yet in smaller companies. It's not even the technical know-how to configure and set things up. It's in the functional roles. Whether you're the accountant or the sales leader, building the mental model for how to move what you do today into an AI-enabled system is a really big leap.
I've found there's no shortcut. You have to train people on a new mental model and show them the art of the possible so they can build that framework themselves. For instance, there are many great wins in customer service and sales that have been proven over the past couple of years. But until you build the sales leader's mental model around a new way to deploy a CRM and gather customer intelligence — until you show them and get them to that "aha" moment — you can set up the whole system and they still won't know how to use it. It's always been that way in IT: you can build the whole technical solution, and people still aren't sure how to use it.
Lucas Hendrich: Right.
Garret Fitzgerald: So that's definitely the biggest barrier. That said, many of these smaller, family-owned businesses have succeeded because they're nimble. They've weathered the last 20 years. And I've seen that once the aha moment clicks, many people jump in, using AI to better themselves, learn, and implement new processes at their companies.
Lucas Hendrich: I think there are at least two layers to the aha moment. Setting technology aside, why invest in a company? You want to take it to another level of scale, and often leadership doesn't have a mental model for how that happens. So that's one mental model you're forming. It's interesting how that dovetails with the digitally native mental model: "All of this was manual, and now we're building infrastructure to automate it" — that's pre-AI. And now we're adding another layer, with AI potentially accelerating adoption of all three, if that makes sense.
Garret Fitzgerald: Right. Two anecdotes from a company I worked with about a year and a half ago, a small construction business with 20 or 30 employees.
One was a newly minted project controller or project accountant with no accounting background and no formal project management training. He leaned heavily on ChatGPT, not only to learn the skills he needed but also to automate some daily tasks, and we helped him with that. We'd also implemented a new ERP system, so we gave him an expert reference resource to ask how to use it more effectively.
The other was essentially an accounting clerk who handled a lot of AP, AR, and invoicing, again without an accounting degree. Within 60 days, she was fully fluent in the new ERP system and had an accounting education — without spending six months in school.
So in many smaller enterprises, there's a huge return on enabling users within their existing job functions. That alone gets you a big return. I always advise starting there, before moving into larger investments in agentic process automation and the like, which I do still think are huge value creators.
Lucas Hendrich: They are. Although to reach that level of maturity, I think you need a certain level of process maturity. And I assume reaching it is part of scaling these companies — a starting point, right?
Garret Fitzgerald: I'd like to believe AI is making Lean Six Sigma great again.
Lucas Hendrich: Mm-hmm.
Garret Fitzgerald: I grew up in that at GE, and somewhere over the last ten years I feel it's really been lost: the rigor of mapping out what's actually happening in the real world, where the delays and gaps are.
What's different about doing that today versus 15 or 20 years ago is that the AI tools for running the process are excellent. You can sit with someone and talk through a process while running Granola, and focus on the conversation — the back-and-forth and the learning — without worrying about taking and structuring notes. At the end of a two-hour conversation, Granola does that for you, then builds a process map that gets you 80% or 90% of the way there. You'll have to tweak it and go back, but that used to take a two-day workshop. So a really powerful part of AI is helping with process understanding and mapping.
Lucas Hendrich: Totally agree. Although when I hear Six Sigma, I think of one defect per million, and I wonder when we'll get to one hallucination per million — choose your unit.
Garret Fitzgerald: Right.
Lucas Hendrich: There are existential risks in using AI in highly regulated settings. You definitely don't want a Six Sigma event. But I think there's also a growing tolerance for anything that isn't high-risk, and there's a history of that. It started with all of us being Gmail's testers for the past couple of decades.
Garret Fitzgerald: Right.
Lucas Hendrich: We're all used to some quirkiness you could call a bug in non-essential things. That's where the power of synthesizing information and producing artifacts comes in. I can't imagine what it'll be like in a few years, looking back: "You spent how many hours on that deck?"
Garret Fitzgerald: On hallucination and reliability, I've formed a mental model: implementing AI agents and using AI tools should follow the same structures and processes we've put in place for humans.
I think too many people see AI as a magic box: give it a generic input and get back exactly what you want, accurately. But no process in all of human history has ever looked like that. In the news industry, reporters don't generally just write an article and instantly publish it in The Wall Street Journal. There's an editorial process.
You need to think about the same systems with AI agents. You need editorial agents and fact-checking agents with more specific purposes, running through cycles to work out the defects and errors. There's a cost aspect — how many tokens you use and so on. But generally, if you apply the same models you've used for humans, break roles and tasks down across multiple agents, and experiment between general-purpose and specific agents, you can tune out those errors.
When I write a report and use AI heavily for research, it's really important to go through five, six, or seven fact-checking cycles using different models and tools. By the end, you get a pretty much perfect report with a reliable set of references.
Lucas Hendrich: I agree that if you take the time to do that, you're also cutting down the slop factor.
Garret Fitzgerald: Certainly.
Lucas Hendrich: I don't know if it's because I've been traveling a lot this week or if it's a broader symptom, but I feel a general AI fog setting in, where I can't tell whether I'm consuming something or generating it. It's the language itself, the tone. It's hard to find my footing — to stand on something and say, "This statement makes total sense; it's concrete." It's hard to describe, but I think there will be more of it. My only concern in a world where you can go from thought to PowerPoint presentation is that we'll have to consume so many of them.
Garret Fitzgerald: Right. Something I've talked about a lot recently is the concept of one-time use: one-time-use applications, content, and code bases. I think it's largely positive. In the context of social media feeds, LinkedIn, and so on, the generated content is becoming overwhelming slop. But in a business and work context, it's different.
Take data conversion during an ERP or CRM implementation. In the past, it was very manual. You'd avoid writing extensive custom code for it, because that ended up more expensive and slower than having a bunch of people convert the data by hand. Today, with AI, the ability to create a one-time, throwaway custom application to convert data from your old system to your new one is incredible.
People still think about building things with permanence. In the work environment, in some cases I'm encouraging people to think about how to quickly execute these one-, two-, or three-time tasks and then throw the tool away, because you don't need it anymore.
Lucas Hendrich: Spot on. And I think the impact there is probably greater in traditional IT than in software development, because so much of traditional IT is "Who cares how ugly the script is? It works and gets the job done," or some skunkworks thing you do once.
It reminds me of something I read recently that you may have seen too. If you're generating general-purpose code with no language preference, the LLMs — in this case Claude — have dominant languages of choice: TypeScript on one hand and Bash on the other. Who likes writing Bash?
Garret Fitzgerald: Exactly. For some people it's like writing hieroglyphics.
Lucas Hendrich: Yet it's a Swiss Army knife and so effective. So for that kind of work, it will be extremely enabling. The precision will come from what you tell it to do.
Garret Fitzgerald: You're right. Normally you wouldn't even pick TypeScript for a very simple, quick application, because getting all your types right is more work. But now you can use it even for simple builds, because AI is doing the work. And it's actually a good check on errors and slop: when the AI doesn't get the types right, it's caught. So it's an additional check.
Garret Fitzgerald: One thought on data analysis: I'm not anti–Power BI at this point. Ask me again in six months; maybe I will be. The reason is that Claude Code is actually pretty good at building out Power BI data streams and dashboards, so you can use AI to build traditional Power BI dashboards.
That said, to your point, I find it even easier and faster to build custom TypeScript and React dashboards that can be modified very quickly and easily, with all the same checkboxes, filters, and everything else. That's really fast and easy to deploy right now. In the enterprise, giving some power users that ability for their data needs is really interesting.
Lucas Hendrich: It's super interesting. I like to think about these things as an evolution over a wider timescale. Twenty years ago, you'd get a report generated in Crystal Reports — remember that? — turned into a PDF and sent out.
Garret Fitzgerald: Yes — BusinessObjects.
Lucas Hendrich: Right. That's the opposite of self-service. Then came the idea that users could run reports themselves: you'd go in, set the parameters, and generate something, or build Excel on top of a data source. Then the dashboard: standardizing the KPIs and calculations so people could just refresh their browser to see the information.
And I assume you're thinking the same as me: the next leap is just the question. It doesn't really matter what the interface looks like. Who cares? It's served to you in whatever way makes the most intuitive sense, so you can make a business decision.
Garret Fitzgerald: Just this week, Claude released a feature where, in the Claude desktop app — and I assume on the web too — if you say "Show me" followed by your query, it produces charts and visuals as part of the response.
How do you enable that from an IT perspective? My biggest fear about releasing data to users to self-build, in any form, was always their context and understanding of the data tables to begin with. I don't think that's gone away.
But ideally, IT's focus can shift from spending hours building front-end dashboards and tweaking them to exactly what users want, to putting context around the data, living alongside it. Then you just enable a Claude desktop plugin for that data set. Now the data has its context right there, and Claude has the plugin — or markdown, whatever — to know how to reference and use it. Boom. Users can ask, "Show me our EBITDA over the last three years."
That's how we enable users, within a conversation or by building more permanent dashboards or Cowork automations. To your point about evolution: in the early days, IT's focus was data. Then you really had to focus on everything from the data through the front end. Now, what's old is new again. Maybe we can go back to really focusing on the data and its context, and just give these AI engines and tools a view into it.
Lucas Hendrich: I think the important guardrail there is making sure we agree on what terms mean. It sounds so simple, yet I realize more as I get older that when we talk to anyone about anything, we all assume we're on the same page, and often we're not. We may use a financial term loosely — we all think "derivative" means the same thing, but it doesn't, and there's a calculation behind it.
So the guardrails will be more around the semantic layer and how things connect. That's another way of saying context, but it includes relationships — if this rises, that falls — that aren't intuitive to LLMs but are essential for interpreting data reliably enough to make decisions.
Garret Fitzgerald: Right. With the NVIDIA conference going on right now, you see a lot of hardware releases putting GPUs right next to the data. It's similar to what we saw 20 years ago, when CPUs started being put right next to the data. Now GPUs are going there too, to build that context: "This is the data point; this is EBITDA; this is the number; and here's the context and information, and it's accurate."
Lucas Hendrich: I'm going to change tack a bit, because there's a theme behind the podcast's name, CTO to CTO. It started with my kids asking what I do every day, and my finding it impossible to explain simply. And as we discussed before recording, there are names and titles for roles, but when you're working within a portfolio to achieve business results, they all compress into the same thing.
Now you're moving from a CTO role to a CIO role. Is there a distinction? I've discussed with guests who are more product-focused — maybe because they build digital products — that the CPTO role is moving closer to the product side, because the technology side is increasingly a question of the right prompting. What are your thoughts? Will the role itself become an artifact, and ultimately just be about business enablement?
Garret Fitzgerald: It's funny. I've been a chief digital officer, a chief technology officer, a chief information officer, and a general manager, and it almost feels like the job has always been the same. Where it sits in the business and how it works influences the title. I should say I'm not a computer scientist; I didn't study computer science. I studied business management and finance, then spent a career in technology roles.
To the more prescient part of your question — the future of the position, whatever the title — first, I still think there's an ongoing role, and always will be, for the deep technologist who understands technology architecture and the decisions around which databases or front-end technologies to use, and who can architect a system. AI is a huge enabler there and will keep improving. But that systems thinking and overall solution-building will remain, both in product and in IT. How do you wire up the company? That's a very important skill set and will continue to be.
What changes is that the era of the ivory tower for IT leaders, whatever the title, is over.
Lucas Hendrich: Mm-hmm.
Garret Fitzgerald: You have to be both a great leader and a builder. To some extent that's true for junior roles too: they have to elevate faster into systems thinkers and leaders, because the era of "I do this one task well" is also vanishing. So there's some convergence, but great leaders will be deeply involved in building, whatever their title.
Garret Fitzgerald: The analogy I keep using for this moment: imagine you're a CFO who doesn't know what a VLOOKUP is, because you've never used Excel. That's the parallel today for a CIO or CTO who doesn't get into the granular details of using these new AI tools. If you're a CFO who's never done a VLOOKUP, how can you lead a finance team? Now it's, how can you lead a technology organization if you haven't done the prompting and experimenting yourself? It's a strange time. You really have to get into the building, if only to build your own mental model and lead your team to do it well.
Lucas Hendrich: I agree. You answered a question I hadn't even asked this way. I think this is where things are converging: many more people will become general managers, in the sense that you'll be able to orchestrate not only how your business runs technology, but also how it runs finance or HR.
And I think the transformation happening now has precedent. If you were around for the DevOps revolution, which felt like one at the time, think of the quintessential admin sitting at the keyboard, the gatekeeper saying, "None shall pass." The whole DevOps movement asked, "Why is a human making that decision? Create rules for it and automate everything."
You could call it the revolution of the lazy, but it's not laziness. It's: Why have a human do what a machine can? Let's be smart about orchestrating and delegating tasks and distributing our cognitive workload.
Now, think about your subject matter expertise — what's readily recallable in your biological hardware and storage. You can't really claim that as value anymore either, because there's a corpus with a thousand perspectives you don't have. So what kernel is left? I think it's a combination of the hardest one — intuition gained through experience, from repeatedly making mistakes over time, in business and in life — which will be a differentiator, and obviously a challenge for people entering the workforce. Plus agency and critical thinking.
I think I saw a post on X from Marc Andreessen saying something like: computer science is becoming like the laws of physics. You can go study it, but it's about defining the boundaries. It's no longer a question of craft at the code level. It's systems thinking and how you plug things together, given those first principles. I don't know if I was asking a question or just agreeing.
Garret Fitzgerald: I do think systems thinking, general critical thinking, and the ability to learn quickly have always been incredibly important, and they're both enabled and more important now. Learning quickly isn't something you're born with. You have to learn how to learn. And part of learning how to learn now is learning to use AI to learn, and to get back non-generic responses. There are many ways to build that skill.
For people entering the workforce, as you mentioned, the bear case is that they're entering a workforce where learned experience matters more. They have less access to human experts than before, and fewer of the jobs that let them build experience over a longer period. So it is a challenge for them.
The bull case: early in my career, I took an IT role without a technology background and had to learn networking. So I'd buy our site's head of networking lunch once a week and pick his brain, plus do Google searches.
Lucas Hendrich: Mm-hmm.
Garret Fitzgerald: Now, what took me six months to learn, I could learn in three days with AI, if I needed to. Many of those tasks are one-off anyway; it's about the result you want.
Lucas Hendrich: For sure. You could learn it, execute the tasks, let it slip out of your brain, and go back to it if you ever need it again.
Garret Fitzgerald: There's a lot of that. But back to systems thinking: really learning means not just the throwaway skill of programming a firewall, but using AI to learn how the system should work and the best ways to implement a network. What are the first principles of networking? What are the trade-offs? Security — how does it affect user experience? Answering those questions and building those skills is now highly accessible.
Lucas Hendrich: One last question that isn't about digital technology, but definitely about labor at portfolio companies that do manufacturing. Is AI adoption keeping pace with robotics, ahead of it, or being considered together with it?
Garret Fitzgerald: Good question. Part of the answer depends on company scale.
Lucas Hendrich: Mm-hmm.
Garret Fitzgerald: At smaller scale, I think they're still very distinct in the conversation. One reason is that much of the robotics in manufacturing is highly specialized and built by robotics companies. You may be able to get data out of it more easily and report on your process, but that's still distinct from a small business building its own custom applications with AI or implementing AI agents in its customer service flow.
For larger companies, there's a much better opportunity to inject AI into existing large-scale robotic operations and manufacturing. With scale, you probably have the leverage to work more closely with the robotics equipment manufacturer and get deeper under the hood.
But even simple things, like pointing a camera at your production line to identify defects — I remember doing that ten years ago as a multimillion-dollar project with highly specialized cameras. Now a middle-market manufacturer can probably get that running in a week with a $200 or $300 camera.
Lucas Hendrich: Right. The building blocks are smaller, more configurable, and easier to produce. I wonder whether there'll be a robotics trend in that direction, so that for whatever specialized widget I'm making, there are component building blocks to start from if I'm doing something custom — instead of essentially needing a separate company to do it.
Garret Fitzgerald: Take the very basic example I described. You see a defect, and there's a way to use AI to get into the PLC and stop the line. So at a smaller or middle-market manufacturer that used to produce a thousand defects before someone noticed and stopped the line, a one-week project with a camera on the line and an automatic stop function means you stop it after ten defects. That's a huge win.
Lucas Hendrich: Yes.
Garret Fitzgerald: And AI's ability to code that app, code the interface to the PLC, and do video analytics is pretty incredible.
Lucas Hendrich: I'm curious about your new role. I know it's new, so there may not be much you can share. You'll be CIO of a data center company. Tell me a bit about the company, and how you see your mental map shifting for this role.
Garret Fitzgerald: I'm very excited to be joining Salute. They provide data center services across the data center lifecycle, including commissioning work for new builds…
Lucas Hendrich: Mm-hmm.
Garret Fitzgerald: …and servicing existing data centers — things like on-site power generation and liquid cooling systems. It's very interesting. There's a huge field services workforce, national and international. So there's an opportunity to equip field services teams doing the real work with better information about the customers and systems they're working with.
I also think we'll be able to do a lot of great things on the sales side, enabling those teams in an obviously fast-growing industry.
Lucas Hendrich: Mm-hmm.
Garret Fitzgerald: There are a lot of new data centers, and we can help the teams get out and sell better. So I'm really excited about those use cases at scale, getting thousands of people the education they need. It's also a growing business with a number of acquisitions, private-equity-backed. I think a lot of great foundations are already in place.
Lucas Hendrich: Mm-hmm.
Garret Fitzgerald: We'll keep building out enterprise processes while maintaining nimble, startup speed.
Lucas Hendrich: In a growing business right now, when you say nimble, I think of digitally native companies that have worked in two-week intervals for the past two decades. The timescale of that, the way of working, is changing again. I don't think we talk about it enough, and I don't think people have settled on what orchestrating work looks like. It's an interesting time to be expanding a business, given how ways of working are shifting. You're probably not looking at 18-month project plans in a big Gantt chart anymore. You're getting out and getting things done.
The other really interesting thing, which I don't think people talk about enough, is that enabling AI creates a ton of jobs. You mentioned field service technicians. Do you see a lot of hiring and real headcount growth? Those are jobs that can't easily be automated.
Garret Fitzgerald: Without a doubt. As we build more and more data centers globally, a lot of jobs will be created in construction, and all the way through to the hands and feet needed to keep them running.
Lucas Hendrich: Electricians. Nuclear physicists.
Garret Fitzgerald: Yes. The other interesting thing for us in IT, as part of a growing business serving more internal customers — in accounting, field services, or sales — is being able to do so much more with each person. Not just more in quantity, but in quality: people can do things they didn't previously have the skills for, pick up the mantle, and move faster. That's a really exciting proposition.
There are metrics you can put around it in certain roles. But back to dashboards: how many of our users have been asking for a new dashboard for two years, and it's never made it above the line of things we can work on?
Lucas Hendrich: "I'm going back to my spreadsheet."
Garret Fitzgerald: Exactly. Or they go hire their own consultant. Just getting to more as an IT team is highly impactful for the business and the decisions business teams can make.
Lucas Hendrich: Anything that was previously IT's purview is becoming democratized and decentralized. So when we talk about field technicians, I'm sure we're not talking about the same technician as 10 or 15 years ago. We're talking about someone who's also AI-enabled. Is that what you're seeing?
Garret Fitzgerald: Yes. Getting people in the field good spec information about what they're working on, for one. But even before they reach the field, there's training: bringing in new people and teaching them the skills they need. AI is a huge part of that.
Lucas Hendrich: Before we started, we talked about OpenClaw at scale, and it made me think of two paradigms for that kind of agentic development and implementation.
One paradigm is OpenClaw-ing my own functions. You see the memes — "OpenClaw started an LLC for me, then voted me off the board and went off and did its own thing." But that started with me: my identity, my WhatsApp, my authentication.
The other paradigm is OpenClaw as an extension of people. I have Kimi Claw running and doing things for me right now, but it doesn't require access to my bank account, or even a real identity. It's just a set of actions, running.
I'm curious what you think. It almost goes in a philosophical direction. At what point do we get used to personas in the world that are Claws, embodied in a sense — with social media profiles? We still live in a world where we deal with individual agents, whether biological or otherwise.
Garret Fitzgerald: You laid out the distinction. There will continue to be many advances in the coworker-style agent: acting on a person's behalf, making their life easier, pulling information together for them. But the real magic of the OpenClaw moment, and where I see this year heading…
Lucas Hendrich: Mm-hmm.
Garret Fitzgerald: …is the agent as its own entity, specifically a general-purpose agent. I think the people doing things really well with OpenClaw are creating a distinct identity for each agent, with its own accounts: its own Notion account, its own Gmail account…
Lucas Hendrich: Mm-hmm.
Garret Fitzgerald: …and its own Slack account, so it truly operates as an independent entity.
Garret Fitzgerald: It will be important to distinguish between the two. I don't think it's a good idea to pass an agent off as a person, and I think people will increasingly want to make sure it's known that they're legitimate people.
So I think we need a new or modified identity schema. Even the basics — password managers, API key management systems — aren't built for agents today; they're built for people. I think we'll probably see a burgeoning product category explode in the next six months, because identity, password, and API key management are among the biggest security issues with agents right now.
If we get that right — and you see companies like NVIDIA releasing an enterprise wrapper for OpenClaw, and Claude continuing to push its Cowork product in that direction — I think we'll very soon have good, enterprise-grade, general-purpose agents.
Lucas Hendrich: I agree. A couple of things could happen along the way. One is that new governance needs to be put in place. I hope it's not formal regulation, because that tends to reflect the lowest common denominator of understanding of the technology, and it could hamstring something that would otherwise drive a lot of progress.
By the same token, I've received email from personas: there's a LinkedIn profile and a picture, the story's a bit vague, and you start questioning it. Then I ask jailbreak questions, and sometimes I get a model response. I find that fun but also annoying.
Garret Fitzgerald: Certainly.
Lucas Hendrich: Maybe part of it is that I don't trust it; I treat it as a sophisticated marketing ploy, spam, or an attempt to hook me. Maybe it'll take agents doing something genuinely meaningful that creates value for me before I'll choose to interact. Either way, I think it's a real challenge if, to your point, there's currently no foolproof way to tell them apart.
Garret Fitzgerald: I'd like to believe companies and platforms like LinkedIn will find ways to solve it themselves. I get far less value from LinkedIn than I did two years ago, and a huge part of that is the volume of messages and slop. At some point that has to hurt the business model, and they'll find a way to adjust.
Lucas Hendrich: Twitter has a legacy of that too, which may be where the bad taste comes from when you wonder, "Are you human or not?" It'll be interesting to see how that paradigm changes. Maybe it'll come from things like Moltbook — some emergent value there. I don't know that it will be actual AGI-style emergence, rather than just some kind of product-market fit. And I do think there's utility in sometimes getting a direct response instead of chattiness.
Garret Fitzgerald: I think within a few years, it will be very normal to be in a group chat where 60% of the participants are agents. I think that will be normalized fairly quickly, and that's probably a good place to be. I just want to know which ones are agents and what their roles are — just as in a five-person group chat, you want to know who the five people are.
It's the same concept: three of them are agents, and here's the role each is serving. In the workplace, knowing that matters. Beyond that, whatever gets to the solution faster. The same goes for customer service agents. As a consumer, if I get my answer better and faster and my issue is resolved, that's what matters.
Lucas Hendrich: Absolutely. I guess it's all about intent, and whether the intent is clear. Maybe that will be the distinguishing factor: "Hi, I'm an agent, and this is my intent."
Garret Fitzgerald: I think Jensen's quote that every company needs an OpenClaw strategy really rings true.
Lucas Hendrich: I think it addresses exactly this issue.
Garret Fitzgerald: That's not to say you should roll out 10,000 unsecured OpenClaws across your company tomorrow. That would be a bad idea. But you need to experiment, learn, and develop your strategy for how and where to inject it. I have no doubt that a year from now there will be very secure ways to implement OpenClaw-like agents, if not OpenClaw itself.
Lucas Hendrich: I also think there's a discipline there, going back to everyone becoming a general manager. There's a heuristic, a mental map, even a skill, to deciding which tasks to give agents and how to divide up your own productivity: "I'll set this task when I go to bed, so it's done when I wake up." There's an art to it that isn't readily apparent. It comes from understanding your role, what you're doing, and how you can bring the most value — not just going through the motions. I look forward to agents running all the meetings, making the lower-level decisions, and presenting the higher-level ones that really matter. It'll change how we perceive time.
Garret Fitzgerald: An interesting thought on time: I think the hours of the workday will keep blurring, especially for agents, for two reasons. First, when is power cheaper? Second, when is usage lower, so model providers offer tokens more cheaply? You already see providers like Claude offering 2x your subscription usage during off-peak hours.
Lucas Hendrich: As a promotion, yes.
Garret Fitzgerald: I think that will keep tracking when energy is cheaper and usage is lower, since providers have a high base load of energy and capex to level out. That will affect when you want agents to run, and what the workday and workweek look like for agents.
Lucas Hendrich: And on the economics, tokens are already becoming a form of currency. People are being given token allowances, but I think we're still far from being good general managers of token use. There are also challenges around token economics across foundation models.
Garret Fitzgerald: Assuming you're a skilled, active AI user, the API token budget your company gives you will be an important number, because it directly affects your total productivity as an employee. You're already starting to see it come up in high-tech companies, and I think it will spread to other roles too.
Lucas Hendrich: I agree. I'd just love to see more maturity, clarity, or connectedness in benchmarking token use and using it effectively. I'm doing a lot and burning a lot of tokens, and I don't have a strong sense of where I got the best return on that burn. Maybe I learned something, or avoided a mistake, because I burned a lot of tokens to reach a conclusion. But it's still not clear to me.
Another layer comes from my experience in the PE world. For ten years, we pushed everyone to the cloud because it made the most economic sense. Then suddenly it didn't, and you were captive to it.
So I'm concerned about the transparency of token economics, and about becoming captive as Claude Cowork gets embedded in companies' businesses, especially in the enterprise. How will that change things, and how will mid-market companies think about it? Maybe the best option is to roll our own — even without the same performance — going with what I'd call open source, though it's really open weights, rather than Claude for Enterprise.
Garret Fitzgerald: On token ROI and cost, I encourage separating things into two distinct categories. The first is when you've built an agent-based process automation. There, you want to focus — maybe not during the initial build, but as you deploy and run it — on token spend versus ROI. That's something you expect to run for a long time, potentially at high transaction volume. If you're getting a $50,000 ROI, you want to make sure you're not spending $75,000 on tokens, compared with a manual process or a process automation that doesn't use an LLM. That's the ROI conversation.
The second is personal productivity and general-purpose agents. At least at this moment, if you focus too much on ROI use case by use case, or outcome by outcome, you're missing the forest for the trees. You have to take a longer time horizon, because people need to learn and build new mental models right now. You have to invest in that, just as we've always had to invest in training.
Lucas Hendrich: You're in the dip of the hockey stick.
Garret Fitzgerald: Even after that, I think you have to look year over year at the overall impact on the business. Have I been able to hold flat at ten people but double my business? That won't be a month-to-month or quarter-to-quarter metric; it'll be year-to-year. So there's a bit of a leap of faith here. But you'll get anecdotal and qualitative indicators along the way too.
Lucas Hendrich: And I think this is being figured out at the micro and macro levels at the same time. One fear I read about just today is that tokens could become what's called a market for lemons: you can't really price them because you don't know the result — what you're actually getting from your token use. I don't think it's that extreme, but it is hard to know what good looks like: the right ratio of token burn to productivity. I think we're still pretty far from that.
Garret Fitzgerald: It's also really interesting to decide whether to use the latest, higher-cost model or a slightly older one. In the first category, LLM-based process automation, that's a bit more clear-cut. When I develop a custom app with Claude Code, I know Opus tokens cost more, but I feel it gets things done faster and better the first time. So am I actually spending less than I would on rework with other models? People have to go by feel. It's still a feel, not a science, at least for the average practitioner. I'm sure someone at Anthropic is studying it.
Lucas Hendrich: I think you're absolutely right. And the deeper you go with Claude, the less capable you become of evaluating other models. As an individual, you have to choose to go deep somewhere. Maybe you have more flexibility, but as an organization, you can't experiment across the board for too long. You have to make a choice.
Garret Fitzgerald: In the first category, where you build something that references LLMs, OpenRouter is a fantastic platform.
Lucas Hendrich: Mm-hmm.
Garret Fitzgerald: One API key gets you access to whatever model you want. Plus, for many open-weights and other models, there are companies dedicated to running them very inexpensively, and those are accessible through it too. It lets you swap models in and out seamlessly, since all you're doing is changing the model reference.
Lucas Hendrich: Mm-hmm.
Garret Fitzgerald: You can do a lot of A/B testing very quickly.
Lucas Hendrich: I think what you're getting at is: How much intelligence does a process actually need? I've seen a lot of prototypes where you don't need a chain of math geniuses to automate your CI/CD process. You do need something with a high degree of intelligence checking your pull requests, but not necessarily for the mechanics at every point in the process. In the worst cases, you end up going back to something rules-based — good old-fashioned AI.
Garret Fitzgerald: It's ironic you say that, because I think many people are moving from rules-based to LLM-based approaches in cases where they shouldn't. What makes it ironic is that you can now use AI to build much better rules-based controls much faster.
So instead of asking how to use an LLM to run the process, ask how to use Claude Code, Codex, or whatever to design and build a comprehensive rules-based approach. Historically you couldn't do that in one shot. Now, when there's an error in your rules, you say, "Go update it," and boom.
Lucas Hendrich: Yes.
Garret Fitzgerald: And you can use the LLM to monitor results against what the rules expect and continually improve them, especially with uncertain data sets. So I think many people should first ask, "Can I do this with rules?" and then, "How do I use AI to help me build those rules better and more reliably?" — rather than putting an LLM in the loop and spending tokens every time.
Lucas Hendrich: I'm sure you've been on or led teams that had to build a rules engine at some point; it's a common concept. It's interesting to think of the LLM's role there as just reviewing the rules and checking for compliance changes. That's the Claw going out, doing all that work, coming back, updating your rules engine when appropriate, and letting you know.
Garret Fitzgerald: We all have a mental model from past rules engines: they're very hard, they take a lot of work, refinement, and person-hours to encode human cognition…
Lucas Hendrich: A lot of human cognition.
Garret Fitzgerald: That's our mental model, which is why people say, "I'll just put an LLM in, and that's better than a rules engine." But you need to break that old mental model. Building a deterministic rules engine isn't as hard as it used to be.
Lucas Hendrich: Right. And you could build it in Bash.
Garret Fitzgerald: Exactly. Have fun reading that code.
Lucas Hendrich: Elon was saying they're just going to go straight to binary.
Garret Fitzgerald: Right — don't even bother.
Lucas Hendrich: I like to save a couple of questions for the end. I'm not sure how much time we have left. I think you've already answered this one in a different way, but what do you think is most overrated about AI right now, and what's most underrated or underutilized, broadly speaking?
Garret Fitzgerald: It's almost the same answer to both. I think the most overrated part is really an "us" issue — a perception issue. There's an unrealistic expectation that AI is a godlike black box that will produce a perfect answer from a generic input. Then you get something back, it's hallucinated, and people say, "See? It doesn't work." I think that's still pretty pervasive.
Garret Fitzgerald: That goes back to what I said earlier: putting in place a mental model for building AI-based systems on the same models you used to design human process flows and teams is critical. And a very underrated part of AI is doing the work to understand the process, dig into the outcome, and really get into the detail — using AI to do that, and then using AI in the solution. There's a lot of "I one-shotted this app" on X, and I think that misses the point. What's underrated is the impact you can have on process.
Lucas Hendrich: And again, it goes back to the general manager concept. You have to be very judicious about what you assign to your agents or ask your LLM.
Garret Fitzgerald: The other one, maybe a mix of both, is that many people still underrate how fast things are improving. I like to think I've been close to the forefront — the operational-use forefront, not the PhD model-building one. In the last three months, the capability has been a hundred times what it was, and it improved over the two years before that too. It's a step change.
And then there's usability for the finance analyst: getting integrated into Excel and PowerPoint, on the desktop, with these agents. I think that will continue. There's a line of thinking that says, "I'll wait and use it when it's better." By then, you could be so far behind, because you have to build the new mental models. Everyone just has to push.
Lucas Hendrich: I think people are caught in that cycle right now: Do I wait to adopt or not? Where does it become fixed as part of my processes?
Lucas Hendrich: The last question: what's your favorite representation of AI in movies, books, or TV?
Garret Fitzgerald: Oh man. Most films are pretty dystopian about it. Well, not all. Jarvis is pretty nice — until he turned into Ultron, I guess. I grew up in the era of the Terminator movies, which were pretty good, and he turned out to be a nice guy eventually. The Matrix was a great set of movies too; I grew up on those.
Lucas Hendrich: How about the most realistic representation of AI?
Garret Fitzgerald: Oh gosh. I hope those don't become realistic representations of AI — I don't think they will. But setting aside the machines taking over the world, The Matrix has a lot of aspects — building virtual worlds and simulations and understanding — that will advance everything about our civilization in the coming years. As long as we're not stuck in one. If you think of that as a use case for advancing human technology, it's going to be big.
Lucas Hendrich: That's the next step, right? World-building. That's where reinforcement learning comes in, and where you overcome the limits of LLMs. On overrated versus underrated, I think LLMs are slightly overrated in terms of general intelligence, because so much of what makes us intelligent is our constant feedback loop with the physical world and its boundaries.
I don't even have a good answer to my own question; there are so many. But I always liked the voice of HAL in 2001, just for the way he declines a request.
Garret Fitzgerald: There you go. Maybe the Star Trek assistant, too.
Lucas Hendrich: That's a good one. One guest picked Her.
Garret Fitzgerald: Oh, yes.
Lucas Hendrich: That's probably the most realistic, in terms of people finding companions in ChatGPT.
Garret Fitzgerald: It's funny: there are the big existential questions — will AI be harmful or helpful, and so on. The one thing I know for certain is that being better at using it, and being on its cutting edge, can't be worse than hiding under the covers right now.
Lucas Hendrich: Totally agree.
Garret Fitzgerald: I don't control the future destiny of mankind and AI. So let's set that aside and do what we can control for our companies and ourselves.
Lucas Hendrich: I agree. And as a final note, my kids know AI's exact limits because they're constantly testing it, so they have no illusions about Skynet.
Garret Fitzgerald: My oldest is seven, and I'm already trying to get him using AI more to ask questions and learn. I'll give him my phone. I think it'll be an inherent part of their lives, and they can learn a lot faster and better by using it.
Lucas Hendrich: Garret, thank you for joining us.
Garret Fitzgerald: Thank you for having me. It was awesome.
Lucas Hendrich: A pleasure.