[00:00:00] [music]
[00:00:04] >> So, my guest today is Joshua Nichols.
[00:00:07] He’s the director of AI and automation
[00:00:09] as Corestack. He started out in tech on
[00:00:11] an IBM contract, joined Corestack, and
[00:00:13] worked his way up to the executive
[00:00:15] force. 14 years later, he’s running AI
[00:00:17] and automation for the whole company.
[00:00:19] Joshua, welcome to Tech Unhinged. It’s
[00:00:21] really good to have you on the show.
[00:00:22] >> Thank you. It’s really nice to be here.
[00:00:24] >> All right. So, before we start, I’d love
[00:00:26] to, you know, kind of hear about a bit
[00:00:29] about yourself and what are you kind of,
[00:00:31] you know, exactly responsible for at
[00:00:32] Corestack?
[00:00:33] >> Yeah. Well, I’ll start You covered it
[00:00:35] lightly where I started, infrastructure
[00:00:37] and support. I actually IT is my second
[00:00:40] career. I had another career years and
[00:00:42] years ago and made a pivot.
[00:00:44] >> Okay.
[00:00:44] >> Back to college, got a degree in web
[00:00:47] design, which I’ve never used, but from
[00:00:50] that I’ve got into support and was doing
[00:00:53] desktop support, really enjoyed that.
[00:00:54] Got a job eventually at Corestack here,
[00:00:57] and that’s 14 years ago. I worked my way
[00:00:59] from
[00:01:00] >> Okay.
[00:01:00] >> desktop support to director of AI and
[00:01:02] automation, which I kind of really
[00:01:04] pushed for and carved out for myself.
[00:01:07] I’m I’m pretty proud of that. Okay. I
[00:01:09] think one thing I would say to people
[00:01:10] who are aspiring to learn AI is I didn’t
[00:01:13] know any more about AI than anybody else
[00:01:14] 3 years ago. I just started using it
[00:01:17] heavily, so.
[00:01:18] >> No, absolutely. I think with the advent
[00:01:20] of these LLMs and, you know, ChatGPTs,
[00:01:23] Claude 4 AI, it’s now much more
[00:01:25] accessible. I typically, you know, kind
[00:01:27] of referred it as you had your real AI
[00:01:30] with all data science and machine
[00:01:32] learning models and everything, and then
[00:01:33] you have this sort of, you know, a layer
[00:01:35] or an API that really gets done all the
[00:01:38] hard, you know, things that you have to
[00:01:39] do, and
[00:01:40] has made much easier access to, you
[00:01:44] know, all of the AI functionality that
[00:01:46] was pretty hard to build before.
[00:01:47] Building on on on that, what are kind of
[00:01:49] AI tools or systems are you kind of
[00:01:52] using in your work environment right
[00:01:53] now? How many people have access to it?
[00:01:56] Are you guys open, not open?
[00:01:58] >> We’re slightly open.
[00:01:59] >> We we have Copilot deployed to everybody
[00:02:02] who has a login and in manufacturing you
[00:02:04] do have people who don’t log into
[00:02:05] computers, but Microsoft, of course, is
[00:02:08] pretty smart to make that available, you
[00:02:10] know, the Copilot chat, the basic
[00:02:12] Copilot experience is available to
[00:02:14] anybody with whatever licensing and
[00:02:16] that’s where we’ll get lost if we start
[00:02:18] talking about Microsoft licensing, but
[00:02:20] we also do have quite a few folks with
[00:02:23] Copilot premium licensing, so that M365
[00:02:27] Copilot license. We are using sparingly
[00:02:29] using other tools like ChatGPT, Claude.
[00:02:32] We are also doing a citizen development
[00:02:35] program, so we’re really trying to
[00:02:37] democratize innovation and bring the
[00:02:40] ability to build your own tools, your
[00:02:42] own automations to individuals, so we’re
[00:02:46] doing that through Databricks and
[00:02:48] through Replit.
[00:02:48] >> Okay. Okay. So, you’re allowing people
[00:02:50] to create applications in Replit then
[00:02:53] access to that data that the
[00:02:55] applications could leverage and that’s
[00:02:56] the on them on their systems and just
[00:02:59] run with it, so you know, everybody can
[00:03:01] build those internal apps that
[00:03:02] previously used to take a lot of time.
[00:03:04] >> With some guardrails, yes. Right now,
[00:03:07] we’re not we’re not allowing
[00:03:08] integrations, so you have to keep
[00:03:10] everything within Replit and then same
[00:03:12] thing with Databricks, we’ll we’ll have
[00:03:15] plenty of integrations cuz they’re
[00:03:16] already there, so that will be a much
[00:03:18] stronger data, you know, data-based apps
[00:03:20] platform and Databricks does allow
[00:03:22] people to build apps. It’s a little
[00:03:25] bumpy in there right now, but they are
[00:03:26] coming out with a chat to build feature
[00:03:28] and I’m not trying to do a commercial
[00:03:29] for Databricks, but they do have
[00:03:30] >> Yeah, so I mean, I remember for one of
[00:03:33] our clients we had to build
[00:03:35] kind of dynamic reporting or ask your
[00:03:37] data kind of a feature where we would
[00:03:39] just simply connect to their database
[00:03:41] and the sales team could ask that, “Hey,
[00:03:43] which of my leads need, you know, are in
[00:03:45] that stage or what do I have to do
[00:03:47] today?” and something on that sort. So,
[00:03:48] and that in the past times, I mean,
[00:03:50] would have been a big big project its
[00:03:52] own self, but more recently is just, you
[00:03:54] know, just takes 4-5 weeks of work and
[00:03:56] you can put it all together. Okay,
[00:03:58] that’s good. So, what has been the most
[00:04:00] innovative app that you’ve seen that
[00:04:03] people have built in the org? That you
[00:04:04] really thought that’s a that’s an
[00:04:06] awesome use case, you know.
[00:04:07] >> There’s several. I will say I’m I’m
[00:04:09] using it for reporting. Executives want
[00:04:12] to know what we’re spending on AI,
[00:04:13] right? And they want to track that. So,
[00:04:15] I just built a Replit app that I can
[00:04:17] pull data in via spreadsheets that I was
[00:04:19] already using. And um I do intend to API
[00:04:22] that into the Microsoft environment when
[00:04:25] that’s when we decide to allow that.
[00:04:28] We’re just making sure we have our
[00:04:30] security in place to monitor those sort
[00:04:33] of API connections. So, we’re putting
[00:04:35] that in place before we do allow those
[00:04:37] connections into Replit, but once I have
[00:04:39] those, it’ll of course just be
[00:04:40] completely automated and there will be
[00:04:43] an app that reports on all of our
[00:04:45] investments in AI, not just not just a
[00:04:48] Microsoft dashboard. So, it won’t be
[00:04:50] fractured anymore. You go to one place
[00:04:52] to see it, which
[00:04:53] >> Good. Good.
[00:04:53] >> Um they seem to like.
[00:04:54] >> Yeah. Are you guys maintaining any
[00:04:56] leaderboard?
[00:04:57] >> No, we’re not. But
[00:04:58] >> That good?
[00:04:59] >> Yeah, [laughter] great great question
[00:05:01] though, cuz we actually just released
[00:05:03] the ability It’s another Replit app
[00:05:05] where if you go to it, you can see your
[00:05:07] own usage of Copilot, either chat or
[00:05:10] premium. So, not necessarily a
[00:05:11] leaderboard, but we are trying to get
[00:05:13] people a little competitive with each
[00:05:15] other.
[00:05:15] >> Based on on all of this, you know, you
[00:05:17] might have I’m sure gotten some feedback
[00:05:19] and and everything. To what extent do
[00:05:21] you trust AI to get things right?
[00:05:24] >> Yeah, this is a good question. I trust
[00:05:26] AI to get things right based on how I
[00:05:30] interact with it, right? So, if you ask
[00:05:33] who’s the home run leader MLB right now,
[00:05:35] that answer’s probably not going to be
[00:05:37] correct. So, you would have to ask one,
[00:05:39] you have to know which tool to go to cuz
[00:05:41] ChatGPT is going to be further behind
[00:05:43] than Google. So, it depends on what
[00:05:45] you’re talking about getting the answer
[00:05:47] right, and then how you prompt it. So,
[00:05:50] then ultimately, you have to verify.
[00:05:51] Cite your sources is one of my favorite
[00:05:53] endings of a prompt is tell me this and
[00:05:56] cite cite your sources, and then I can
[00:05:57] click on that, and is that a reputable
[00:05:59] source? I think we build confidence in
[00:06:01] AI getting things correct by how we
[00:06:03] interact with it, not by just taking its
[00:06:06] answers and using it as an answer
[00:06:08] machine. It’s definitely not a good
[00:06:10] idea.
[00:06:11] >> So, boss of hype, have you actually
[00:06:13] built some autonomous AI agents and
[00:06:16] moved them into production?
[00:06:17] >> Not autonomous. We do have like search
[00:06:21] agents. We have agents that do off of a
[00:06:24] trigger will do an activity, but it’s
[00:06:27] only that activity. So, it’s not, you
[00:06:28] know, when we talk about an autonomous
[00:06:30] agent, usually that means it can use
[00:06:32] whatever it it thinks its way through,
[00:06:34] and we’re being a lot more prescriptive
[00:06:36] right now, and we’re saying this is how
[00:06:39] you will do it, and here’s the workflow
[00:06:41] to follow. So, not exactly autonomous
[00:06:43] yet. I feel like that’s It’s not scary.
[00:06:46] I wouldn’t call it scary, but we’ve seen
[00:06:48] some news stories, right? Everybody’s
[00:06:49] seen the news stories. I think if you
[00:06:52] put the correct
[00:06:54] >> [laughter]
[00:06:54] >> Or yeah, yeah, it hacked its way out of
[00:06:57] the sandbox and into hugging face, yeah.
[00:06:59] [laughter] Yeah.
[00:06:59] >> So, why I mean, A, what kind of
[00:07:01] workflows have you kind of seen would be
[00:07:05] the best suited for, you know, AI agents
[00:07:07] or AI automation? Because a lot of times
[00:07:10] what what I’ve seen is that, you know,
[00:07:11] people usually tend to just say the word
[00:07:14] A, they need an AI agent, but they
[00:07:16] generally don’t because if it’s a very
[00:07:19] defined set of tasks, the good old
[00:07:21] automation workflows can work
[00:07:23] wonderfully well, right? So,
[00:07:24] >> When it’s super defined, yeah, why do
[00:07:26] you need an agent to think its way
[00:07:28] through and burn up tokens when you
[00:07:30] could just make it deterministic, which
[00:07:31] is why I’m following that that
[00:07:33] deterministic. Even though I’m using an
[00:07:34] agent, I’m still using more
[00:07:36] deterministic workflows. I have seen and
[00:07:38] we have an idea for one that we’re
[00:07:40] working on where there’s air handling
[00:07:44] and iPad system and
[00:07:47] Boomi is the iPad system, but we get a
[00:07:49] lot of errors. These are user-based
[00:07:51] errors, so the user does something wrong
[00:07:53] and then the integration goes wrong as
[00:07:54] it tries to write back to the ERP
[00:07:57] system, finance system and right now a
[00:07:59] person goes in, looks at each one of
[00:08:02] those errors and then emails the person
[00:08:05] who fat fingered whatever they fat
[00:08:07] fingered and says, “You got to fix this.
[00:08:08] This is what you need to do.” totally do
[00:08:10] that with an agent and it’s it’s hours
[00:08:12] and hours of tedious, mind-numbing work
[00:08:15] that we suddenly can free that person up
[00:08:18] to go work on and you know, these are
[00:08:20] smart people doing this. They can go
[00:08:21] work on a lot better things.
[00:08:23] >> So, that’s that’s what people and that’s
[00:08:25] what organizations are trying to do,
[00:08:27] right? So, finding those repetitive,
[00:08:29] long-hour jobs that they previously had
[00:08:32] to do, kind of automate that because you
[00:08:34] could put some level of decision making
[00:08:36] now. So, it’s a deterministic 80%, but
[00:08:39] you need some level of thought process
[00:08:40] in it and that 20%, so the agent is
[00:08:43] effectively able to do that and so
[00:08:44] paying 50% of their time and you know,
[00:08:46] that same resource or person can then
[00:08:48] focus on other things and work on those
[00:08:50] areas. So, that’s actually good.
[00:08:52] >> Yeah.
[00:08:52] >> So, I was I was reading across and and
[00:08:54] you’ve written that, you know, access is
[00:08:56] not adoption. So, what do you actually
[00:08:59] show people to prove that, hey, AI is
[00:09:02] worth their time investing into Claude,
[00:09:04] ChatGPT, Replit, you know, basically
[00:09:07] understanding a non-technical workforce,
[00:09:10] build actual confidence in AI tools
[00:09:12] because we as technology people are
[00:09:14] usually at the forefront of most of the
[00:09:16] things, right? But non-technical staff
[00:09:18] usually aren’t, so
[00:09:19] >> It’s a lot of times it’s showing them a
[00:09:22] way that, hey, this can make this
[00:09:23] better. This can do this faster or and
[00:09:26] this is more my workflow. I look to make
[00:09:29] the work richer, not necessarily faster.
[00:09:32] So, sometimes it takes me the same
[00:09:33] amount of time, but my output is
[00:09:35] significantly better than it was and
[00:09:37] that’s really going on to that second or
[00:09:39] third level of how to use AI, which is
[00:09:42] like using AI to use AI.
[00:09:44] >> [laughter]
[00:09:45] >> Um where you know, you’re having the
[00:09:47] conversation with AI, and then you’re
[00:09:49] saying, “Okay, how would I prompt this
[00:09:51] if I want to do this?” Things like that.
[00:09:53] But I do feel like I yes, I get more
[00:09:55] work done, but I also have deeper,
[00:09:57] richer work, and I try to show people
[00:10:00] that as well, depending on their role.
[00:10:02] Some people just want to you know,
[00:10:03] something that does email faster,
[00:10:05] something like that. That’s fine. If you
[00:10:06] get them comfortable with that, great.
[00:10:08] They probably start to experiment. But
[00:10:10] the one place where I started, and I
[00:10:12] feel like the people who really start to
[00:10:14] dive deeper tend to start is home use
[00:10:16] cases. Um planning a vacation, or
[00:10:19] writing a new recipe based on the
[00:10:21] ingredients in your pantry, all kinds of
[00:10:24] little things like that tend to open up
[00:10:26] ideas to how you might use it at work,
[00:10:29] too.
[00:10:29] >> Yeah. Using AI to figure out how to use
[00:10:32] AI, that’s I think that’s a real moment
[00:10:34] when people get get, you know,
[00:10:35] comfortable enough with it. Because back
[00:10:37] in the day, it was like Googling in for
[00:10:39] information. Now it’s just asking
[00:10:41] Claude, “Hey, how do I use you to, you
[00:10:43] know, to do XYZ?”
[00:10:45] >> And that’s an unlock, right? Like this
[00:10:46] is the first uh technology I’ve ever
[00:10:49] come across that you can just ask the
[00:10:51] technology how to use it, and it will
[00:10:52] explain it. That’s not a thing that’s
[00:10:54] happened before. So people are still
[00:10:56] kind of like they’re not really used to
[00:10:57] that concept. I think in in the next
[00:10:59] year, most people will be used to asking
[00:11:01] AI how do I do this when I want to use
[00:11:03] the AI.
[00:11:04] >> Somebody from my staff comes over and
[00:11:06] asks me the question, “Did you ask AI?”
[00:11:08] And No, I don’t know how to Did you ask
[00:11:10] AI on how to use it, you know, to do
[00:11:12] that?
[00:11:13] >> Yeah.
[00:11:13] >> So yeah, that’s a that becomes an
[00:11:15] eye-opener when you start really
[00:11:17] effectively using it. In all of itself,
[00:11:19] you need AI policies, right? So you
[00:11:20] mentioned harnessing, you mentioned some
[00:11:22] guardrails, everything. So let’s say if
[00:11:24] you had to And I think you earlier on
[00:11:26] you mentioned that we know we have not
[00:11:28] yet given access to data as such, or
[00:11:30] writing as such. So but if people have
[00:11:33] to really start experimenting, what is
[00:11:35] one or two policies or one or two things
[00:11:37] that you feel that we would need to give
[00:11:39] permission to people and that, you know,
[00:11:41] can be a problem from a security
[00:11:44] perspective as well.
[00:11:45] >> I think our security group is probably
[00:11:47] already a little nervous based on our
[00:11:49] policies. They did help and and actually
[00:11:51] did write the policies, but it really
[00:11:54] the policies align with the company’s
[00:11:56] risk tolerance. So, we are not really
[00:12:00] saying no to anything. Uh we can’t allow
[00:12:02] Chinese models based on being in the uh
[00:12:05] supply chain for the US government, but
[00:12:07] that’s the only thing we’re saying no.
[00:12:09] The rest we’re saying, yes, you can use
[00:12:12] it, but this is, you know, these are the
[00:12:13] guardrails. Um and if you’re using
[00:12:16] ChatGPT, which we don’t have a
[00:12:18] widespread enterprise agreement with,
[00:12:20] so, you know, you want to use your
[00:12:22] personal ChatGPT to to create a picture
[00:12:25] or something like that, fine. Just
[00:12:26] obfuscate company name.
[00:12:29] >> Any confidential
[00:12:30] >> anything confidential, just use common
[00:12:32] sense. And we’re treating our end users
[00:12:35] like adults. We’re saying, this is what
[00:12:37] you already know cuz you already know
[00:12:38] you can’t Google, you know, you
[00:12:40] shouldn’t be putting like I I use the
[00:12:42] uh the analogy Facebook, right? You
[00:12:45] wouldn’t put this on Facebook. Don’t put
[00:12:46] this
[00:12:46] >> Right.
[00:12:47] >> in your private LLM uh subscription. On
[00:12:50] the other hand, with Copilot, yeah, you
[00:12:52] can talk about most of that stuff uh and
[00:12:54] still use caution. Uh that’s our stance.
[00:12:56] I’m sure I appreciate that our security
[00:12:59] group has, you know, adopted that stance
[00:13:01] as well. It’s it’s been a nice balance.
[00:13:04] >> So, data sensitivity automatically
[00:13:06] becomes part of your AI training.
[00:13:08] >> Yeah, and it’s was there before, to be
[00:13:10] honest, cuz we have always needed data
[00:13:12] sensitivity training.
[00:13:13] >> What is the single biggest skill gap
[00:13:15] kind of holding teams back from using AI
[00:13:18] effectively? And how do you think the
[00:13:20] leadership at any organization could
[00:13:22] kind of either create new training
[00:13:25] programs or alter their existing
[00:13:27] training programs? you know, how they
[00:13:28] can go about it.
[00:13:29] >> think it’s a skill [snorts] gap that I
[00:13:31] see. Although the skill for using AI is
[00:13:34] really just to use AI more and and once
[00:13:36] you like to your point, once you unlock
[00:13:38] that ask AI how I do this. But what I
[00:13:41] see is people are not either not willing
[00:13:44] to or not thinking about how now that AI
[00:13:48] exists, how do I recreate workflows? How
[00:13:51] do I stop doing it the way a human would
[00:13:53] and start thinking about how AI can do
[00:13:55] it and not limit myself to the existing
[00:13:58] workflow. So reimagining workflows I
[00:14:00] think is one of the bigger skill gaps
[00:14:03] if if we want to use that term.
[00:14:05] It’s holding people back some. You’re
[00:14:07] still able to do great things even with
[00:14:10] current workflows, but what can you do
[00:14:12] if you reimagine the workflows?
[00:14:14] >> And that’s where the AI native
[00:14:16] organization term is also coming up,
[00:14:18] right? So where it’s not only that you
[00:14:20] have to embed AI or bring it in your
[00:14:23] process, it’s like rethinking the entire
[00:14:26] way you used to work, which is why it is
[00:14:28] much much more disruptive as the digital
[00:14:30] transformation initiative that started
[00:14:32] 20 30 years back, right? So then all the
[00:14:34] ’90s and 2000s were integrating, setting
[00:14:37] up ERPs, putting your manual, you know,
[00:14:40] getting rid of your paper-based
[00:14:41] processes and everything. And
[00:14:43] >> [laughter]
[00:14:43] >> and then organizations spent so much
[00:14:45] time like three, four decades in in
[00:14:47] setting all of that up. Now they really
[00:14:49] have to rethink that, hey, the whatever
[00:14:51] process that we defined, maybe, you
[00:14:53] know, this has to be completely
[00:14:55] rethought over.
[00:14:55] >> Yeah, I would say what thank goodness
[00:14:57] though for the digital transformation
[00:14:59] cuz without it AI couldn’t really be
[00:15:01] effective at all, right? So although we
[00:15:03] now have to rethink those workflows,
[00:15:06] that’s a cool problem to solve.
[00:15:08] >> So you mentioned that you guys have
[00:15:09] opened up, let’s say Microsoft Copilot,
[00:15:12] but are we also allowing I mean, have
[00:15:14] you integrated with the the router APIs
[00:15:16] that they offer as well that service so
[00:15:18] you can connect to Anthropic or OpenAI
[00:15:21] APIs because or how do you embed Replit
[00:15:23] in the entire process?
[00:15:24] >> Yeah, good. That’s good questions. Yes,
[00:15:27] we are allowing and have opened up other
[00:15:30] LLMs through Copilot, which is really
[00:15:33] a great unlock because when someone
[00:15:34] tells me they need Cloud, I’m like, you
[00:15:36] have Cloud. It’s in
[00:15:38] Go go go try that.
[00:15:39] >> [laughter]
[00:15:42] >> And and then yeah, how Replit fits in is
[00:15:44] it’s not integrated. So, we’re not
[00:15:47] saying you can use Replit within
[00:15:50] Copilot. That would be a cool thought
[00:15:51] though. Maybe something for me to think
[00:15:53] about trying. But right now it’s just,
[00:15:55] you know, go to Replit. We have a a
[00:15:56] short training built on a Replit. You
[00:15:58] know, it’s a Replit web app. Here’s what
[00:16:00] you can and can’t do and what you should
[00:16:02] do and how to publish all those things.
[00:16:04] It’s super fast. I built it really fast.
[00:16:06] It was really cool. But yeah, you do
[00:16:08] that. You get in. You’re a builder. You
[00:16:09] start building. Generally, we see
[00:16:11] people, you know, get in there. They
[00:16:13] build a few apps. They often will not
[00:16:15] publish those apps. Then they’ll come up
[00:16:17] with a really big idea. They’ll burn
[00:16:19] through a bunch of tokens. They’ll
[00:16:20] publish their app. And that’s kind of
[00:16:21] where it stops there a lot of times. Not
[00:16:23] a lot of people just keep building more
[00:16:25] and more apps, which I’ve seen that as a
[00:16:27] pattern now. So, it’s kind of like, I
[00:16:29] solved the problem that I wanted to
[00:16:31] solve. And now I’m busy using the app,
[00:16:34] not building another one. But we do have
[00:16:36] a few people who are real repeats. And
[00:16:39] that’s really cool to see too is people
[00:16:41] who are like, yeah, I fixed this. Now I
[00:16:42] fixed this. And it’s amazing. We have a
[00:16:46] finance app for the IT group built by
[00:16:49] one of my peers
[00:16:50] who it’s just fantastic. Made our
[00:16:52] budgeting and budget tracking and our
[00:16:56] spend super super accessible. Way easier
[00:17:00] than the spreadsheets we were using.
[00:17:01] >> So, do you guys have to make any
[00:17:03] decisions between then to buy an
[00:17:05] off-the-shelf AI tool versus building
[00:17:07] those custom applications or agents or
[00:17:10] has it largely been we’ll do everything
[00:17:11] in-house. We don’t need off-the-shelf
[00:17:13] tools anymore. The famous build versus
[00:17:15] buy discussion is back now, right? So
[00:17:18] >> Yeah, right. Right. Yeah, it is. Um I
[00:17:20] think one of the areas where we would
[00:17:22] not build is when you get into anything
[00:17:24] regulated. So, people data, tax, things
[00:17:27] like that, right? You don’t want I don’t
[00:17:29] want to We don’t want to mess with that
[00:17:30] because then you have to maintain all
[00:17:33] these different legal knowledges. And
[00:17:35] so, it’s better better to buy in that
[00:17:36] case at least from our standpoint. Big
[00:17:39] enterprise apps, we haven’t built any
[00:17:41] yet. We are thinking of that and how
[00:17:44] that it’s going to work. At that point,
[00:17:46] IT signs up to support that, right? Cuz
[00:17:48] if everybody in the company needs to use
[00:17:50] it, then IT probably needs to be owning
[00:17:52] it.
[00:17:53] >> At that stage, would you say, “Okay,
[00:17:55] this is a company-wide initiative.
[00:17:57] Almost all of the staff has to use it.
[00:17:59] So, let’s see what options are available
[00:18:01] out there whether we go the custom route
[00:18:03] or whether we go towards the buy route
[00:18:05] or the build route.”
[00:18:06] >> Yeah, but it’s also a negotiation lever
[00:18:08] to throw in. It’s like, “Well, we could
[00:18:10] build this. So, you know, if your if
[00:18:12] your quotes your bid to get our business
[00:18:16] comes in too high, we’ll probably just
[00:18:18] go the build route and who knows?”
[00:18:20] I think a lot of companies are going to
[00:18:23] move to that at least as a negotiation
[00:18:25] lever.
[00:18:25] >> Absolutely because building has become
[00:18:28] much much faster. I think the biggest
[00:18:30] element now, and that’s kind of leading
[00:18:32] into the next question as well as where
[00:18:34] is the domain knowledge or the
[00:18:37] understanding of how particular process
[00:18:39] works. That has become much It was
[00:18:41] already critical, but now it’s like
[00:18:43] super critical because
[00:18:44] the building process is much fast. This
[00:18:46] process of understanding the process
[00:18:49] itself, defining it, and and I mean, I I
[00:18:52] might just go to an extent where, you
[00:18:54] know, what I’m saying now is that the
[00:18:56] entire waterfall methodology of software
[00:18:58] development might just be coming back
[00:19:00] because
[00:19:01] >> [laughter]
[00:19:01] >> you have to give the agent a complete
[00:19:04] spec or, you know, understanding of what
[00:19:07] you want to build. And that was exactly
[00:19:08] what waterfall was, you know, figuring
[00:19:10] out the requirements. But because the
[00:19:12] other part of the agency has become
[00:19:14] rapidly fast.
[00:19:15] >> Yeah, I just to double down on that. One
[00:19:17] of my favorite ways to build is an
[00:19:20] interview with AI first. Interview me
[00:19:23] one question at a time. Uh help me
[00:19:25] unlock this idea. Build the spec sheet.
[00:19:28] Deliver the spec sheet to the agent.
[00:19:30] Then, I’m not spending a ton of time
[00:19:32] iterating back and forth with the agent.
[00:19:33] It’s like, this is a pretty pretty well
[00:19:35] thought out plan. Execute.
[00:19:37] >> Right. A lot of times, you know, this
[00:19:39] token usage and the AI cost budget has
[00:19:42] been a concern for organizations because
[00:19:45] previously you could easily predict
[00:19:46] that, “Hey, I have X seat. Each seat
[00:19:49] cost me Y dollars. And this is what I’m
[00:19:51] paying on a monthly basis.” Now, all of
[00:19:53] a sudden, I don’t know what the usage
[00:19:55] going to be. Maybe Josh spends, you
[00:19:57] know, $10,000 on tokens on
[00:19:59] [clears throat] one day.
[00:20:00] >> [laughter]
[00:20:01] >> And the bill rises up. And how do you
[00:20:03] organizations kind of are dealing with
[00:20:05] that? Uh are there limits placed or what
[00:20:07] do you think is the best way? And And
[00:20:09] that’s just one cost, but there might be
[00:20:10] some hidden technical and operational
[00:20:12] costs as well.
[00:20:13] >> Yeah, total cost of ownership. That’s
[00:20:15] one that people often overlook and have
[00:20:18] for years, [clears throat] not just in
[00:20:19] AI. So, I can tell you how we’re
[00:20:20] approaching it. And my little philosophy
[00:20:22] recently shifted because I had kind of
[00:20:25] ascribed to this like, well, you just
[00:20:27] want to give people the ability to be
[00:20:28] creative and restrictions are
[00:20:30] restrictive on creative. But, I just
[00:20:32] recently read the book and I don’t I
[00:20:34] won’t recall the author, but the book’s
[00:20:35] name is inside the box and it’s
[00:20:38] fantastic at how constraints actually in
[00:20:41] inspire creativity. And so, giving
[00:20:44] people some constraint, not hard deep
[00:20:46] constraint where, oh, you get $100 of
[00:20:49] credits and, you know, spend them
[00:20:51] wisely. They’re probably not going to
[00:20:52] use that much cuz they’re like, who
[00:20:54] wants to get addicted to that $100 worth
[00:20:57] of credits and then
[00:20:59] on the third day of the month you can’t
[00:21:01] use it anymore because you did all the
[00:21:02] good things you can do. So, you don’t
[00:21:04] want to restrict so much that the
[00:21:06] creativity and the learning and the
[00:21:08] ability to automate will go away. But
[00:21:11] you do want to set caps in a way that
[00:21:14] this is enough to get you through the
[00:21:15] month if you’re clever and you don’t
[00:21:19] overspend and you learn how to automate
[00:21:22] well. So it actually drives good
[00:21:25] learning. And so the first month they
[00:21:27] tend to blow through the you know fairly
[00:21:28] large budget and hit the cap and then
[00:21:30] they’re asking for more and we’re like
[00:21:32] yes, you can have some more. Here are
[00:21:33] some tips on how to not do this again.
[00:21:36] And we’re trying to get people to become
[00:21:38] more efficient in their use. And you
[00:21:41] know, you don’t always need to drive the
[00:21:42] Ferrari to work. You you don’t need
[00:21:44] Fable 5 every time you type in
[00:21:46] something.
[00:21:48] >> I think that the understanding of which
[00:21:49] model to use that’s also really and that
[00:21:52] could become part of the training as
[00:21:53] well, right? So if it’s something
[00:21:55] simpler or something you know, just no
[00:21:57] creativity involved, simple math, go to
[00:21:59] a lower model, reduce your cost and and
[00:22:01] then so on and so forth. So yeah, that’s
[00:22:03] a that’s an interesting way of keeping
[00:22:05] the build versus buy kind of a
[00:22:06] discussion. When a vendor or a
[00:22:08] consultant kind of pitches you AI,
[00:22:10] what’s the fastest way you know, they
[00:22:13] kind of lose their trust? What’s you
[00:22:14] know, how do you then identify whether
[00:22:17] they’re selling a hype or actually
[00:22:19] understanding your business and selling
[00:22:21] you something you know, worthwhile or
[00:22:23] >> That’s a tough one because the promises
[00:22:25] are there, often the evidence is there.
[00:22:27] I’d use the same approaches that have
[00:22:30] always traditionally used with SaaS is
[00:22:32] okay, let me talk to one of your
[00:22:33] customers. That’s a good one. Yeah, it’s
[00:22:35] a good vet. But uh I will say you know,
[00:22:38] on LinkedIn and probably end up with
[00:22:40] some connections based on this as well.
[00:22:42] Out on LinkedIn, I will get some
[00:22:43] outreach and people trying [laughter] to
[00:22:45] sell me stuff and generally it’s it’s
[00:22:47] very like, you know, it’s like, hey
[00:22:49] Josh, I liked your post, buy this. I I
[00:22:51] like and that’s just it’s a sales I find
[00:22:54] that there are people in sales who are
[00:22:55] really good at I’m going to build some
[00:22:57] rapport with this person before I pitch
[00:23:00] what I have. And honestly, don’t don’t
[00:23:02] pitch it if it doesn’t fit that person.
[00:23:04] Just move on.
[00:23:06] >> You know, 10 years back, I think
[00:23:08] uh mail tracking was was kind of
[00:23:11] launched where somebody would know that
[00:23:13] you opened their email. And we we get
[00:23:15] the same as well, but you know, I
[00:23:16] remember we somebody opened their email
[00:23:19] and you would just immediately call that
[00:23:21] person or just email and say that hey, I
[00:23:23] noticed that you opened my email and it
[00:23:25] was sound creepy.
[00:23:26] >> Stalker.
[00:23:27] >> At that yeah,
[00:23:28] >> [clears throat]
[00:23:28] >> at that early stages, nobody really knew
[00:23:30] that there were mail tracking softwares
[00:23:32] that were launched. Right? So, after a
[00:23:34] while, it it became a standard, but you
[00:23:36] know, early on, it was like, how did he
[00:23:38] how did he know?
[00:23:40] That was another, you know, kind of
[00:23:42] funny way of doing things. Couple of
[00:23:43] questions on the surrounding AI hype. I
[00:23:46] mean, everybody’s going crazy on AI and
[00:23:50] just trying to implement it or having a
[00:23:52] fear of, you know, kind of a FOMO if
[00:23:54] they don’t do it and whatnot. What do
[00:23:56] you think some of, you know, what what
[00:23:58] what do you feel is that it’s something
[00:24:00] thing that the industry is really
[00:24:01] confidently doing and that might just
[00:24:03] not age that well.
[00:24:04] >> First of all, let me say, AI is not a
[00:24:06] magical unicorn that poops money. It
[00:24:08] just it doesn’t create and if you hear
[00:24:10] that pitch, oh my god, you’re going to
[00:24:12] save so much money by using our product,
[00:24:14] that’s that’s a good vet right there.
[00:24:16] No, it’s it’s not going to just
[00:24:18] hemorrhage money into your environment.
[00:24:19] I think well, it’s already not aged
[00:24:22] well, but leaderboards, like you
[00:24:23] mentioned, that has already turned out
[00:24:26] to not be a good idea. People burning
[00:24:28] through their credits right away. I
[00:24:30] don’t have a crystal ball on what’s not
[00:24:32] going to age well that’s going on right
[00:24:34] now. Yeah, I don’t have a good answer
[00:24:36] for that other than in retrospect,
[00:24:38] obviously trying
[00:24:39] maxing was not a good idea. And I think
[00:24:42] we’re going to continue see changes that
[00:24:45] will honestly, I just had a thought.
[00:24:47] This isn’t necessarily AI, but it’s data
[00:24:50] centers. I think the AI industry is
[00:24:52] falling down on their face with the data
[00:24:55] center thing. You don’t go in, you don’t
[00:24:57] have just talk to the local mayor and
[00:25:00] city council and get this, you know,
[00:25:02] pushed in in the dark. You got to start
[00:25:04] a campaign months in advance. You got to
[00:25:06] explain to people how you’re not going
[00:25:08] to impact your their water. You’re not
[00:25:10] going to raise their rates in electric
[00:25:12] bills, things like that. You’re going to
[00:25:14] pay for the infrastructure upgrades and
[00:25:17] you’re going to create about 200 jobs
[00:25:20] and a lot of
[00:25:22] uh labor uh jobs, so skilled labor jobs.
[00:25:25] Not really a ton of like IT jobs. You’re
[00:25:27] not going to have see this small town in
[00:25:29] Kansas get filled up with a bunch of IT
[00:25:31] nerds. There’s going to be a lot of
[00:25:32] electricians, HVAC people, and uh
[00:25:35] construction workers who get jobs based
[00:25:38] on this very large build out. And yeah,
[00:25:40] it’s probably not pretty, so maybe fix
[00:25:42] that problem, too, right? Make your a
[00:25:44] data center buildings prettier. These
[00:25:46] are rounding errors in the budget for
[00:25:49] these things. Build community goodwill.
[00:25:52] Build a new high school. I mean, these
[00:25:54] are not hard [clears throat] wins.
[00:25:55] >> And that’s
[00:25:56] >> And they’re they’re trying to run these
[00:25:57] on like like they’re a shoe string
[00:25:59] budget, but the return on their
[00:26:00] investment is massive. They need to get
[00:26:03] out in front of this cuz they’re losing.
[00:26:05] Like states are writing laws that are
[00:26:07] pretty much prohibitive. The need for
[00:26:09] data centers isn’t going to go away, so
[00:26:11] we’re really creating I think the push
[00:26:13] back is right. Uh don’t try to do this
[00:26:15] in the dark. Don’t try to sneak it in on
[00:26:17] communities. That’s not the way to do
[00:26:18] this.
[00:26:19] >> No, I agree and that’s actually a good
[00:26:21] point because that’s a debate that has
[00:26:23] now started in terms of what those data
[00:26:25] centers are and it’s largely a cooling
[00:26:27] problem, so that can be optimized over
[00:26:29] time, absolutely, but large swaths of
[00:26:31] lands are being purchased and and and
[00:26:33] those problems are are building up. It’s
[00:26:35] a good answer that.
[00:26:35] >> Yeah, solve those problems, too, right?
[00:26:37] >> Solve those problems, absolutely.
[00:26:39] >> hemorrhage water and if you are using
[00:26:41] that system, you’re using the wrong
[00:26:42] system cuz there’s a closed system where
[00:26:43] you don’t hemorrhage water.
[00:26:45] >> Final question. What’s the hardest
[00:26:47] lesson you’ve learned while you’ve
[00:26:49] progressed your career, worked in AI?
[00:26:51] How do you think that can help the next
[00:26:53] generation that’s coming in?
[00:26:54] >> The hardest lesson in the AI space that
[00:26:57] I have learned so far is it takes a lot
[00:27:00] of practice and I would consider myself
[00:27:03] pretty good at being creative with AI,
[00:27:05] pretty good at prompting, and good at
[00:27:07] using AI as a thought partner. I really
[00:27:09] had to dig in and I did it one way. I
[00:27:12] think other people are growing slower. I
[00:27:14] did a very fast growth track, but that
[00:27:16] took a lot of time on my own just
[00:27:18] playing with AI, practicing, trying to
[00:27:20] build out reports or trying to build out
[00:27:23] workflows, things like that. It was a
[00:27:25] lot of thought partner work. The reason
[00:27:27] that was a hard lesson was because it
[00:27:30] took a lot of my personal investment. I
[00:27:31] was working, you know, doing nights,
[00:27:33] weekends, evenings, but I came out of
[00:27:36] that not needing to do quite as much of
[00:27:39] that. And then the other hard lesson is
[00:27:41] when to let it go. As someone who’s, I
[00:27:43] would consider myself an AI power user,
[00:27:45] there’s always one more little prompt to
[00:27:47] send and it’s like, “I’ve got this call
[00:27:49] in 2 minutes, but I can send this prompt
[00:27:51] and then AI can be working on that in
[00:27:53] the background well while I was
[00:27:54] >> That is so true. Yeah.
[00:27:56] >> It’s
[00:27:56] >> [laughter]
[00:27:57] >> It’s like let it go. Like you
[00:27:59] It’s okay to let that go once in a while
[00:28:01] and and maybe have a AI-free day every
[00:28:04] week where, you know, Saturday or
[00:28:05] Sunday.
[00:28:06] >> Yeah.
[00:28:07] >> Don’t don’t touch AI at all.
[00:28:08] >> No, you’re absolutely right on that. You
[00:28:11] know, I’ll just send this prompt out and
[00:28:12] then I’ll do something else because
[00:28:14] it’ll keep working on the background.
[00:28:16] >> But you’re always trying to get that one
[00:28:17] more thing done. Yeah.
[00:28:18] >> I’m I’m I’m guilty of that, but in the
[00:28:20] last month or so I I figured that I
[00:28:22] don’t need I I I will not open up Claude
[00:28:24] for 2 days a week, you know, maybe
[00:28:25] Saturday, Sunday, or at least give a
[00:28:27] day’s gap in between. It just drives you
[00:28:29] nuts at times
[00:28:30] and yeah, that’s that’s a good lesson to
[00:28:32] leave it.
[00:28:33] >> Problem-solving machines, right? So
[00:28:35] we’re
[00:28:36] And so when you have such a powerful
[00:28:38] tool to solve problems, you can kind of
[00:28:39] get addicted to that.
[00:28:41] >> Yeah, that’s true. All right, Joshua,
[00:28:42] that was really good. I think I enjoyed
[00:28:44] the conversation. I hope you did as
[00:28:46] well. Thank you. Bye. Bye-bye.
[00:28:51] >> [music]
