So What? Theoretical AI Usage

So What? Marketing Analytics and Insights Live

airs every Thursday at 1 pm EST.

You can watch on YouTube Live. Be sure to subscribe and follow so you never miss an episode!

In this episode, Katie, Chris, and John dissect theoretical AI usage and contrast academic research with practical implementation.

You will uncover hidden capabilities within modern technology platforms to transform daily operations. A clear view of theoretical AI usage reveals organizational obstacles that hinder successful technology adoption. This shift allows decision makers to isolate high-value tasks and eliminate workflow friction. By applying theoretical AI usage to strategic planning, teams unlock massive productivity gains.

Watch the video here:

So What? Theoretical AI Usage

Watch this video on YouTube

Can’t see anything? Watch it on YouTube here.

In this episode you’ll learn:

Where commonly recited theoretical AI usage stats come from
What percentage of work can truly be automated
What the difference between automated and augmented is

Transcript:

What follows is an AI-generated transcript. The transcript may contain errors and is not a substitute for listening to the episode.

Katie Robbert – 00:31

Well, hey everyone. Happy Thursday. Welcome to So What?, the Marketing Analytics and Insights live show. I’m Katie, joined by Chris and John. Howdy, fellas up top today. This week we are talking about theoretical AI usage. So in theory, Chris, what are we talking about?

Christopher Penn – 00:51

We’re specifically talking about the new paper that we published on AI, What Jobs Can AI Do?, which, if you are at all in our sphere, you have gotten multiple emails about. This paper comes in two flavors: a helpful, friendly summary and a “here’s a whole bunch of data” academic paper.

Fundamentally, what this paper is about—back in, was it February or March?—Anthropic released a paper about labor, about the future of work and labor. That paper was very, very widely cited. I could pull it up: Anthropic Labor Market Impacts of AI. Katie, we have shown this—

Christopher Penn – 01:42

I’ve lost track of how many times we’ve shown this on various podcasts, blogs, and stuff like that, but this was their paper. The way they put this together was they used a paper from 2023 that said, using the O*NET SOC database from the US government, here are 18,000 tasks. A panel of human judges, plus at the time OpenAI GPT-4—which is now a dinosaur—were asked, “Here’s this task. Could a machine do it in half the time as a human?”

That’s what creates this blue line on this chart. Then Anthropic mined millions of Claude conversations with their Claude thing and said, “Do we see evidence in these different professions, based on the context of the chat, that people are using AI in this discipline?” The red line is what they came up with.

Christopher Penn – 02:37

And so, Katie, you and I have been doing a ton of work recently using the TRIPS framework by Trust Insights, which is a 5-part framework, to say: Is this a task well-suited for AI based on time, repetitiveness, importance, pain, and sufficient data? Combined with the Artificial Analysis benchmarks for whether AI has the capabilities to do these things—can it see, can it read, can it reason things out?—that was the genesis of our version of the paper, which asked what AI could plausibly do and where we land on that. I’ll bring up our version of the chart, which requires a lot of scrolling.

Katie Robbert – 03:27

So while you’re bringing that up, before we get into our version, I first want to talk about—can we go back to the Anthropic version for a second? I want to talk about the methodology a little bit. I’ve mentioned a handful of times—or in every call that we have—that I have done true clinical trials and academic research, and I am a super stickler for methodology.

So when I hear “theoretical”—and, again, to their credit, they use the word theoretical—this is highly problematic because people look at this and go, “Oh my God, this is black and white. This is what this means.” They said, “What could AI do that humans currently do?” So that’s blue. Then they mined chat conversations, probably without our consent.

Katie Robbert – 04:23

That’s neither here nor there; that’s for a different episode of how people were actually using it. We know from our own talking with people that the limitation of people using these tools is that they don’t know what is possible.

So I feel like that red section of how people are using it is the wrong way to go about it. Most people—not in the marketing industry or data science, but the general population—are still using a large language model as a search engine, or “write this email for me,” or “take these notes,” very basic functionality, because they don’t know what’s possible. First and foremost, just acknowledge that this specific analysis is riddled with methodology problems if you’re looking at it from a truly evidence-based perspective.

Katie Robbert – 05:23

And I would say it’s not.

Christopher Penn – 05:26

Yeah. One of the things about Claude in particular is that Claude does not have a lot of the modalities that other tools do. It cannot generate images, for example. It cannot generate video. It is a very capable text-based tool, and it does have a mobile app that has the ability to interpret images, which is pretty good. But there are other tools—ChatGPT, for example, Gemini, etc.—that have a lot more multimodality. In some of these fields, to your point, Katie, there are capabilities of AI that it can do that people don’t know it can do.

Christopher Penn – 06:05

Or, as was pointed out many times on LinkedIn by folks like Ethan Mollick over at Wharton, people in the general population are using the free versions of tools that are so reduced in capabilities compared to the paid versions that they never truly experience what a Fable 5 model can do because it’s behind a paywall.

A classic example is installation and repair. I use AI all the time for this. One of the things I do in one of my keynotes is show a video of me taking my phone, walking up to my furnace system, turning on the camera, and saying, “Here’s the model of my furnace. Here’s what it is.” I point to the valve and say, “Why is this leaking?” It says, “That’s a pressure release valve. It’s fine. Just hand-tighten it until it’s tight, and it will stop leaking.”

Christopher Penn – 06:53

“It just got loose.” I’m like, okay, that saved me $175 from calling out the plumber for what is clearly just loose. I did Google it just to make sure that was actually what that part was, but that is a use of AI in this case that, A, goes beyond chat because it’s multimedia, and B, is in installation and repair, where Anthropic reports no usage at all.

Katie Robbert – 07:19

Well, I feel like we could debate the finer points of whether that is really just a Google search or actual true use of AI, but your point is well taken. This is something that we point out in our paper: in a lot of these industries where there is a notion of “Can AI do it?”, there is actual opportunity. It’s not that the AI is going to do this for me—AI is not going to hold the torque wrench in a mechanic shop and fix the engine, but it is going to guide you through perhaps a different model of engine you’ve never seen before.

Katie Robbert – 08:03

So it’s thinking about it in terms of those assistant categories and educational categories, versus “it’s going to pick up the tools and fix the engine for me.” Maybe in some factories they have those robotic assembly lines—that’s one thing, but it doesn’t extend to your everyday mechanic. I think that, again, is some of the blind spots in this analysis.

Christopher Penn – 08:37

Totally. We are already seeing advanced use of AI in agriculture. For example, there are autonomous robots that have been deployed to fields that have computer vision and essentially look at a tomato plant and assess, based on leaf coloration and things, “Do we need to make an adjustment in the fertilizer mix?” It does it fully autonomously.

There’s no human chatting with Claude about whether this tomato plant needs maintenance or not. It is just this gigantic machine that moves through the field.

Christopher Penn – 09:08

So there are, as you said, Kate, a lot of gaps in this because you’re not using an LLM to do that work, but you 100% are using AI. You’re using computer vision combined with a vision language model to make that assessment and go, “Yeah, the soil here needs more calcium. The leaves are turning this color because there’s a deficiency of calcium.” Then it looks in the eight tanks of chemicals it has and says, “Okay, I’m going to spray some fertilizer on this.”

Katie Robbert – 09:32

What’s also not acknowledged here—because you just brought up a really good point—is that while you have these robotic systems, it doesn’t speak to the availability of these resources. You mentioned things like Fable 5 being behind a paywall. If Claude was just scraping free chat conversations, then I would fully expect those conversations to be fairly basic.

Whereas with what you’re describing, someone has to have access to an autonomous agentic robot that is pre-programmed to go out into a field to be able to assess soil, temperature, and other things. That’s not captured in this analysis.

Katie Robbert – 10:17

So when someone is looking at this graph—whatever you want to call this—it doesn’t tell that story, but it does scare the bejesus out of people.

Christopher Penn – 10:32

Yes, that is true. Our version of this uses job descriptions instead. In a lot of cases, there are things in job descriptions where, to your point, Katie, someone may not think, “Oh, I could use an AI tool for that,” but then they try it out, see the results, and go, “Oh, I didn’t know it could do that.”

It’s one of my least favorite terms, but it applies here: it is the art of the possible, saying what AI can do. So what we did was we fed 90,000 tasks through two different models: Claude Haiku and DeepSeek.

Christopher Penn – 11:14

No, it’s three different: DeepSeek V4 and MiniMax M3. We ran them through the TRIPS framework plus our AI Likelihood Index to say, of this job description, what are the tangible outputs of this job? Then, of those tangible outputs, what is the likelihood that AI would be capable of doing those tasks?

What we end up with—which is our red line here—is that in a lot of cases, there are big chunks of jobs, even in things like computers and management, where talking to your team, coaching people, and collaboration are things machines are not going to do, or you probably don’t want them doing. In other professions like production, building and grounds, and healthcare support, there are tasks that machines can do.

Christopher Penn – 12:03

People may not be having conversations about them with Claude right now, but they absolutely are in that wheelhouse. That’s where we landed with our version of this chart.

Katie Robbert – 12:14

Again, it comes down to availability. Our good buddy Brian—hey, Brian—said, “When we start seeing AI being more integrated into robots, these blue-collar gaps are going to fill quickly.” I agree with you to an extent. I think theoretically those blue lines are going to fill in more quickly, but the red—the actual “can we do it, do we have access”—I think that’s going to be tough because it’s going to cost money. It’s an expense that people may or may not be willing to take on.

As we know, AI or a robot or a machine, whatever it is, is not a set-it-and-forget-it. It needs maintenance. You need someone on your team who actually knows—let’s say the robot breaks down in the middle of the tomato field.

Katie Robbert – 13:07

Is anybody going to fix it? Does anybody know how to fix it? Is there a reboot button? Very quick side note: it reminds me of one of my favorite cult movies, Return to Oz. There’s a character called Tik-Tok—prior to the actual app—and he breaks down, and there are instructions written on him about how to get him restarted.

Without those instructions, this smaller version of Dorothy is just standing around going, “What the heck do I do? I have this robot, this mechanical man standing in front of me, and I have no idea how to boot him back up.” I feel like that’s going to be the challenge as you bring these tools into less conventional industries.

Katie Robbert – 13:51

That’s not to say people won’t be able to figure it out, but it’s going to be a steeper learning curve because it’s not part of their everyday use. This is something that I have the privilege of hearing about from my husband, who works at a company where the overlords believe very much in “AI everything.” But the everyday staff members—the team members who are actually working with their hands and dealing with people—just keep having AI thrust upon them. Nobody has time for it, it’s a whole different workflow, and there’s nobody around to fix it when it breaks.

So I just want to flag those things: yes, AI is going to become more capable, but no, humans are not going to want to use it more.

Christopher Penn – 14:41

There is actually a video from UBTECH that shows the reverse of that Tin Man thing. This is a total distraction, but what the heck, we’ll go with it: this robot pulls its own battery out of its back, goes to the charging station, grabs a new battery for itself, and recharges itself on the floor. I think that’s a lot of fun—sort of the Tin Man being able to repair itself.

Katie Robbert – 15:13

John, what do you got for me?

Christopher Penn – 15:15

Help me.

John Wall – 15:15

Well, it’s got to have two batteries, right? That’s the problem with that one.

Katie Robbert – 15:22

All right, so back to the point.

Christopher Penn – 15:25

Yes, back to the point. There is a lot more that AI can do in those low-represented fields, and there is less that AI maybe should do in the highly represented fields. One of the fields that was most highly represented was business and finance. According to the Orlando study from 2023, 94% of the tasks in that field machines could do twice as fast as humans. I don’t know about that.

There’s actually one other interesting blind spot here too that you mentioned, Katie, that just occurred to me. Claude is a wealthy AI, meaning you have to have money to use it. If you’re using Claude Max 20, for example, that’s a $200-a-month subscription.

Christopher Penn – 16:10

When you look at the universe of AI and what people are actually using out in the field, it is Chinese models, like DeepSeek V4 Flash. The new version of V4 Flash is an incredibly powerful model. It is very smart, and it is one-fortieth the cost of Claude—literally a penny on the dollar for Claude.

If we look at OpenRouter’s most recent leaderboard, DeepSeek V4 Flash is six trillion tokens. It is the largest model in use in the OpenRouter development system, followed by High3 by Tencent Mimo, the new version of Flash GPT 5.6 Luna, and the smallest version of OpenAI. Claude doesn’t even make it into the top 10 because of its cost.

Christopher Penn – 17:05

So a paper that is based on Claude usage inherently is not even in the top 10 of the most used models in some of these things. How can we know that’s even representative of what the AI universe is?

Katie Robbert – 17:20

I feel like one of the distinctions you just made—maybe I’m mishearing, but you’re saying “in the development space.” So if I’m in healthcare and I’m a nurse being told to use AI, my first thought isn’t, “Huh, I wonder if there’s a cheaper Chinese model that I can use.” It’s “What is right in front of me?”

So I hear you when you say that these alternative models to the big three are being used more, but what this paper doesn’t say is who is using these things. What are their professions? What is their level of expertise? Because if it’s pulling—and I’m thinking of the Claude research specifically, not ours—it’s just pulling chats from literally everyone.

Katie Robbert – 18:10

So it could be your 83-year-old grandmother who’s looking for a blueberry buckle recipe because she just got a bunch of blueberries from the farm stand that’s represented here, but she doesn’t have a profession that’s represented. It’s definitely skewed and misrepresented data.

I also want to acknowledge that when the general population sees data like this, we tend to panic, because the “AI taking my job” panic is really real and these papers don’t do a great job of really explaining what it means. Hopefully, that’s what we’re able to do today in addition to providing our own version of the research.

Christopher Penn – 18:57

Exactly. So that’s the short version of what we found: there is much more opportunity in the underrepresented fields. We were talking about this on the podcast this week. If you are a sniper and you’re doing counter-sniper operations, you can use vision language models to assist you. If you are a lifeguard, you can use a vision language model to assist you.

That’s not going to be represented because most people either don’t know the capability exists, don’t have the knowledge for how to ask the machine for help with something like that, or don’t know how to build a system that is agentic in nature that could handle a task like that on an ongoing basis.

Christopher Penn – 19:47

The other thing that was interesting is that the Anthropic numbers, when you look at them in terms of how well they represent the goodness of fit to all the different categories—Anthropic’s Claude usage numbers only have a goodness of fit of about 33%. What that means is, compared to what’s possible, the way people are using Claude is not a good fit. It inherently says the way people use Claude is not a good fit to the way AI could be used.

Our score is a much closer fit to that, showing that if you’re going to try to figure out how we should be using AI, the methodology using the TRIPS framework is going to get you better results.

Katie Robbert – 20:43

Which I think is more helpful for someone trying to understand where AI fits into their business. That’s really what we’re talking about—anyone from any walk of life is welcome to read this paper, but our ideal customer for this paper is someone in a business who is trying to figure out that next step in their AI enablement.

For us to ground our analysis in that makes a lot more sense than just taking a look at what people are talking about or what questions people are asking these chatbots. That’s not helpful.

Christopher Penn – 21:25

Yep. It’s interesting because when you start digging into the individual roles themselves—out of the 90,000 different tasks, this one is food and protein production at Cargill—you can start to see what tasks are either lower or higher value. Recording production and inventory data in computer systems—yeah, that is clearly an AI task. Whereas down here, manual meat and protein processing, I think your husband would agree, I probably don’t want AI doing that.

Katie Robbert – 22:00

No. For those who don’t know, my husband’s a butcher. But that’s exactly it—this passes what you call the sniff test of logically making sense. If you really dig in, there are probably ways to incorporate AI, but in broad strokes, that all tracks. That makes sense.

John Wall – 22:21

Yeah.

Christopher Penn – 22:21

To your point, that is how we validated that the roll-up results were good: by looking at some of these 90,000 tasks through the 3,600 jobs that we dug through to see if the tasks flagged as good for AI were actually good for AI, or if something went wrong. If it said chopping meat is a good task for AI, that would be an indicator that something went terribly wrong.

So, Katie, based on the papers, the analysis, and all these things, if someone’s saying, “Okay, I understand theoretical usage of AI. Now how do I make more use of it? Our company has spent a gazillion dollars, we budgeted 11% of our operating expenses next year, and we don’t know where to start.” How should somebody get started with this sort of AI enablement?

Katie Robbert – 23:20

I am so glad you asked. We can help at Trust Insights with AI enablement—truly, we really can help you. But in all seriousness, I think that what we’ve seen a lot of our clients doing successfully comes down to structure. It’s not that people are uneducated about AI or don’t know where they want to take it; it’s that oftentimes they have limited resources, or the shape of their organization is the limiting factor.

In the Inbox Insights newsletter this week, which you can get at TrustInsights.ai/newsletter, I talk about the four basic structures of an organization and the limitations that they put in place if you’re trying to scale AI.

Katie Robbert – 24:10

Most enterprise-sized companies—and I’m saying “most” in broad strokes—are hierarchical organizations where if John is the CMO, Chris is the VP, and I’m the analyst, I can’t talk to John; I have to talk to Chris, who can then talk to John. Those layers moving up and down provide a huge block for AI to be able to scale.

Then you have functional organizations where you are clustered by discipline—so you have marketing, HR, etc. You can scale AI really well within that specific discipline, but then it doesn’t translate if finance is doing similar things to marketing and the two organizations have a virtual firewall. That’s another limitation. You also have a flat organization where people have a lot of autonomy, but nobody’s really in charge and nobody has authority.

Katie Robbert – 25:08

Nobody can say, “This is how things should be moving forward,” because nobody is used to having to answer to anyone. Then there is a matrix organization, which on paper is the most ideal for scaling AI, but is the least ideal in a business operational sense because it brings resource and billing headaches.

In a matrix organization, let me say I’m the manager and Chris and John are my two direct reports, but then I also have another team under me. They can all share resources, but then Chris is overbooked while John doesn’t have enough, and who’s footing the bill and making that decision on resources? There’s more collaboration for AI, but more of a headache for humans. These are the limitations that AI runs into, but it doesn’t know that’s the limitation.

Katie Robbert – 26:05

This is where a lot of organizations are today. First and foremost, I recommend people start by getting a sense of what those blocks are, starting with looking at your organizational structure. Once you understand what the blocks are, you can start to pick apart and prioritize how to move them.

If you’re in a hierarchical organization where only certain levels can talk to certain levels, it may be an opportunity to revisit your AI Champions program or your Center of Excellence to ask, “Is it doing what we need it to do? Is it pulling people from across these different levels so that they can collaborate together?” And so on and so forth. Then, I would take a look at your use cases.

Katie Robbert – 26:51

This is something a lot of organizations have—they say, “Here’s all the things I want to be able to do with AI, but I’m not currently doing them.” You can run it through what Chris mentioned: the TRIPS framework. We have a free version of the TRIPS framework on our website.

You can go to TrustInsights.ai/trips and do a very simple exercise where you score each of these tasks for time, repetitiveness, importance, pain, and sufficient data. That will start to give you a sense of priority on where to move next. If you have a larger-scale project, that’s where we come in. We can do a lot of this for you because we have automated it with our own proprietary software.

Katie Robbert – 27:41

I feel like I’m ranting a lot today.

Christopher Penn – 27:43

That’s all right, that’s all right. The other thing I would recommend, from a technological perspective, is checking in once a quarter to see what the capabilities of the tools themselves are and to understand what is out there.

One of the sites that I spend probably the most time on is called Artificial Analysis. It’s a company that does independent benchmarking of all the different AI models. You can get a sense of who is neighboring whom. Claude Opus 5 and Claude Fable are neighbors—they’re very close in terms of intelligence. When you look at their agentic capabilities, they’re somewhat neighborly.

Christopher Penn – 28:23

There’s a spectrum of all the different companies—594 major model families out there. On the leaderboards are other things like text-to-image (what models are good at that), image editing, text-to-speech, and transcription/speech-to-text.

As you’re starting to think through this, especially when we’re in budgeting season, it’s a good idea to take a step back and ask: What are our use cases? What are our needs? As Katie said, do the TRIPS framework and then see if there is an ecosystem mismatch. For example, if you are a Microsoft shop using Microsoft Copilot, Copilot’s primary model is Phi-4 Multimodal Instruct.

Christopher Penn – 29:13

When you look at Phi-4 on these charts, you go, “Oh, well…” In fact, I think I can even pull up Microsoft Phi-4 Multimodal. This is the model that powers Copilot. Look where it stands down here on the intelligence index.

Katie Robbert – 29:37

Our friend Erica brings up that, as we were talking about organizational structure, matrix organizations can also end up with every marketing team—digital, Corp Comms, content—using different tools for different functions. That’s exactly right.

This is why I recommend starting with what kind of organizational structure you have. You could have on paper a matrix organization, but then within that, you have a functional organization, and that’s where you start to run into those issues again. Once you know what that looks like, the next step, to Erica’s point, is to do a tech stack audit: What are you using, why are you using it, and how are you using it?

Katie Robbert – 30:18

Then this leads you into your use cases of what you actually want to do. It’s a lot of work, but if you are an organization looking to really scale AI in 2027, this is the work you have to do.

Christopher Penn – 30:34

Exactly. It is occasionally useful—probably more than occasionally useful—to bring in an outside perspective. In any organization, even one as small as Trust Insights, you kind of get used to doing things a certain way. You have a certain toolset you’re working with.

If you’re not leaving your desk, getting out in the world, and seeing what’s out there, you may not see some of those use cases and realize, “Oh, I didn’t know it could do that.” That is very different. It’s one of the reasons why I spend a lot of time reading academic papers published from major AI conferences to see how they’re doing things in drug discovery.

Christopher Penn – 31:20

I didn’t even think about using a language model in that capacity, and now I can mentally add that to my toolbelt of things to try.

Katie Robbert – 31:30

The way that I describe that in this week’s newsletter is like those blind spots. If you think about the layout of your house, everyone has issues with the layout of their house, regardless of where they live. We just get so used to working around it because it’s easier than taking the time and resources to change the layout.

If your laundry is in the basement, but it would be better on the first floor, you might complain about it, but given the time, energy, and budget that it takes to fix it, you’re just going to work around it. You’re going to figure out a different system to make it work. John, you love/loathe home repair.

Katie Robbert – 32:09

I’m sure you run into this a lot: “If it was just moved this far over,” or “If we had just thought to swap these rooms…” But you learn to live with it.

John Wall – 32:20

Yeah, carrying a washing machine up cellar stairs—it’s cheaper to just leave the trash in your basement and buy a new one. Switching costs are real, and design is critical. If the shutoff to your oven is behind it and you can’t move the oven, you’re just going to burn the house down. You need to think about this stuff and take it into account.

Christopher Penn – 32:45

The last thing I’ll add on theoretical AI usage—and this is something from two weeks ago in the newsletter you wrote, Katie: if you have something and you’re not sure if it’s a use case, ask the machine, “Hey, how could you do this? If I give you this task, how could you tackle it, or what parts of this task could you do?”

Clearly, the machine is not going to be handling cutting meat for you. However, it might be able to teach you a thing or two about recognizing different patterns, things to look for, or “Hey, that’s got Cyclospora all over it.”

Katie Robbert – 33:26

I don’t even know where to go from that! Basically, the use case that you’re describing—a lot of it is educational. I would say that with the caveat that one of the things we know about these large language models is that they are very overeager, programmed to be incredibly helpful, and say yes to everything.

If you give a large language model a use case without any real guardrails, restrictions, or knowledge of what you and your company can’t do, you may be led down a false path thinking, “Well, AI said it could do it, so let’s just have it do it.” Then you find out along the way that you have all these different bottlenecks or don’t have the capability.

Katie Robbert – 34:09

So I would say ask AI, “Can you do it?” But also say, “Here are the limitations, and here are the things that we as an organization can’t do because of budget, resources, education, or whatever it is. Can you work within these limitations: yes or no?” That’s going to give you a better-quality answer.

Christopher Penn – 34:29

Yep. Another thing—and you do this a lot, Katie, with Claude Cowork—is to add a connector to whatever your company’s collaboration environment is and say, “Hey, pull out the last seven days’ worth of tasks that I’ve talked about in our system and make a catalog of it.” Then you put that through the TRIPS framework to ask, “Okay, what did I do this week? Oh, I did all this stuff, and I probably don’t need to do half of it because a machine should be doing it—like, look, yet another status update.”

Yeah, you should have handed that off to Claude a while ago. A lot of the time we have task blindness—we get so used to doing something that we don’t even physically remember it anymore.

Christopher Penn – 35:09

Like, “Oh, yeah, did I do the monthly report? I don’t remember if I did the monthly report, and I don’t remember what’s in it.” That was before AI. Now, if you use these tools to dig into the knowledge bases you have, audit the systems, and audit yourself—have it look at your outbox, have it look at your direct messages, and ask, “What did I do this week?”

Katie Robbert – 35:32

I do that a lot because I can’t for the life of me remember anymore, for a variety of reasons. I’ve found it to be an incredibly useful tool because then I can push it into our project management system, so I am not solely responsible for remembering everything that I committed to.

My next evolution—it sounds really basic—is to start using AI a little bit more in my inbox, because right now I kind of keep everything in there. I’m a bit of a hoarder in my inbox, and it gets overwhelming. There’s stuff in there that I don’t need to keep. I can move it into more useful places like a task list, or I can archive things that I’m just CC’d on where I have no action to take.

Katie Robbert – 36:16

There are better ways to do it. It’s very much as you said: it’s that task blindness of “Well, this is how I’ve always managed my inbox. Why would I do it any differently?” But it’s getting overwhelming.

Christopher Penn – 36:29

To Brian’s point earlier, as you start getting the systems—and this will be happening over the next year or two—integrating into the physical world, you’re going to see many, many more opportunities. For example, if you’ve ever seen a Raspberry Pi, tiny, super-cheap computers, you can get those with tiny little cameras. If you wanted to—I’ve done this myself—you can plug that little computer into your laptop, and Claude or the tool of your choice (ChatGPT, Copilot, whatever) can write the code that those little boxes speak, because they’re really just Linux boxes, and you can start building your own real-world things.

Christopher Penn – 37:12

For example, you can take a Raspberry Pi and a little camera module, and have a tool build a bird-watching app that you put next to your bird feeder. When a bird lands, it turns the camera on and starts sending you screenshots of what bird landed on your feeder, or you can build your own flock camera or whatever.

If you have things like Arduinos, which are robotic systems that you can plug into, you can start to build in the real world using AI to engineer the software for it. In terms of theoretical use of AI, we’re getting out of Black Mirror and getting into real life. There are a number of open-source Wi-Fi systems, for example.

Christopher Penn – 37:56

If you want to construct your own Wi-Fi network at home, it’s easy to do. We do this a ton with servers that we operate, like our email marketing automation server. I have Claude Code on that, just helping me maintain the sucker.

If you go on Amazon or the device vendor of your choice, there are weather stations you can buy for your backyard that have open software so you can modify them to your heart’s content to say, “Notify me when the backyard temperature gets above this,” or “when it’s pouring rain,” or whatever, as I realize some of these use cases—

Katie Robbert – 38:29

Can’t you just look out the window?

Christopher Penn – 38:33

You can look out the window, but you don’t necessarily remember to record the data. Right now, for example, we are entering what is called a super El Niño, where the weather patterns in the Pacific actually switch directions. You’re going to see very large climate changes across the planet because effectively the Pacific Gulf Stream is reversed.

If you have that data plus an AI tool plus your own backyard data, you can start to project and measure its impact and say, “Okay, if I’m growing crops—as we saw in the food production and farming section of theoretical usage—now you have systems that can help you anticipate and get ahead of problems before they cost you your entire season.”

Katie Robbert – 39:19

You had me at “bird-watching app.” Don’t get me anything for Christmas: build me a bird-watching app!

John, what do you think? You’re on the front lines of hearing how people are trying to scale their AI and what’s next for their AI enablement. What are your thoughts on where people should go?

John Wall – 39:47

A key point to this is that theoretical is great to discuss, but the reality is there are human factors that have to be part of it. Theoretically, there’s no reason we shouldn’t already have electric cars—it’s technically superior in every way, really. So you’ve got to align the human forces. You’ve got to get incentives square, because we know orgs right now where cultural and political issues will actively fight against this. Even though it’s profit-motivated, it can still be sabotaged.

So you need to get ahead of it, get people aligned and trained so they can understand the opportunities. We’re, of course, glad to help with this kind of stuff.

John Wall – 40:36

We’ve worked with other orgs that have struggled to do this stuff. It’s not enough to just check boxes and say, “Okay, we’re using it.” There’s more to do than create bad pictures for your slide deck; there are other places to go.

Take full advantage. There’s never been a greater opportunity in business in the last 20 years. Since the dawn of the internet, this is the next opportunity to get ahead of everybody else and find a new way of doing business. Take your shot.

Christopher Penn – 41:08

I’ll give you one last example. I was doing this for a self-storage company that has gate hardware and software that basically controls the front door, the lockers, and stuff like that. This one company was saying, “We’re ending our desktop software package. You now all have to pay 100x more for the SaaS software version of this that is internet-controlled.”

Out of curiosity, I asked, “Can we find the hardware specs for this system?” Yes, we could—it’s all made of very common components.

Christopher Penn – 41:40

So on a lark, I had an AI system build net-new software for it, because it was not a complicated system to begin with. It produced it and passed all of its test uses. I don’t own a self-storage facility, so I had no way of actually testing this in production, but if you have hardware today that has an open standard, there is no limit to what you can do with this.

Theoretical AI usage goes way beyond “summarize this email” into “Hey, I just saw this company that produces open-source hardware for wearables like watches and things. I’m going to make my own fitness software.”

Christopher Penn – 42:19

What the heck, why not? Folks should figure out: What could you do if you have access to the system and you can get a machine to talk to it? What could you do? And I guess we’re making a bird-watching app.

Katie Robbert – 42:37

Yeah.

Christopher Penn – 42:41

All right, folks, we have—

Katie Robbert – 42:43

Well, I was going to say, we mentioned two papers: our academic paper and our more user-friendly paper. We have links to those papers in our free Slack community, Analytics for Marketers, which you can join at TrustInsights.ai/analyticsformarketers.

If you want to learn more about how we can help, you can go to TrustInsights.ai/contact. Talk to a real human—he’s right here, his name is John! Or you can learn more specifically about how we can help you with your AI enablement at TrustInsights.ai/aienablement, and then you don’t have to talk to John. But if you would like to, you still can.

Christopher Penn – 43:18

It’s fine talking to John, not Ricard.

John Wall – 43:20

You don’t have to. Only if you want to.

Katie Robbert – 43:22

I love talking to John. I don’t know why you wouldn’t want to.

Christopher Penn – 43:26

All right, folks, that’s going to do it for this week’s show. Thanks for tuning in, and we will talk to you all on the next one. Thanks for watching today. Be sure to subscribe to our show wherever you’re watching it.

For more resources and to learn more, check out the Trust Insights podcast at TrustInsights.ai/tipodcast and our weekly email newsletter at TrustInsights.ai. Got questions about what you saw in today’s episode? Join our free Analytics for Marketers Slack group at TrustInsights.ai/analyticsformarketers. See you next time!

Need help with your marketing AI and analytics?

So What? Theoretical AI Usage

So What? Marketing Analytics and Insights Live

airs every Thursday at 1 pm EST.

In this episode you’ll learn:

Transcript:

Leave a Reply Cancel reply

Pin It on Pinterest