AI Hype Detox for Manufacturers and Distributors with Heather Hershey

Aaron Sheehan: Foreign. Welcome back after a very long absence to the B2B Uncut podcast. I am still your host, Aaron Sheehan, uh, with Oro Commerce, who helpfully sponsors and distributes this podcast. We have to make up for the long, long absence of not having any content. I'm so sorry. Uh, and there are reasons for that, and people have been beaten, um, to, uh, to atone for that, that error in judgment. But to make up for that, I am joined by, uh, Heather Hershey, who should need no introduction, but we're going to give her a chance to do it anyway, uh, because that's always helpful. But, um, Heather and I have known each other for some years. I, uh, would say she is ubiquitous on LinkedIn, um, like so many, like so many people are. And I would describe her as one of the sharper tools in the shed. But I will let her introduce herself to you. Heather, who are you and why should people care what you have to say?

Heather Hershey: Oh, great question, Aaron. Thank you for inviting me on the pod. Um, I'm Heather Hershey. I'm an analyst from a tier one analyst firm. If you want to figure out which one, please go to, uh, my LinkedIn profile. Heather Hershey, one on LinkedIn. Um, but just so you know, uh, what I'm about to say on this podcast is a reflection of my own opinions. So that comes with all the biases that that would imply, but is not a reflection of my employer. All right, caveat over. My background is really complex. Um, but like a lot of analysts before me, this path led to an interesting place, uh, where I'm talking to line of business people about it things and vice versa. So in my life, uh, I am a former commerce professional, former, uh, AI researcher. I went to UGA for AI for a master's. Um, my actual emphasis was in natural language programming and sentiment analysis, um, which has moved very, very far beyond what it was back in 2013 when I was in my master's program. So I'm dating myself as a very old millennial, and I write about commerce and agentic commerce, B2B commerce, um, how B2C commerce platforms are not always the greatest fit for B2B commerce, um, and topics of that nature. So thank you for giving me an opportunity to be on my soapbox for a moment there.

Aaron Sheehan: Oh, I love that. And it's a theme that of course resonates very strongly with us over at Oro. And so we were talking, you know, about what do we, what do we want to discuss? And it felt like to kick off 2026 the heels of NRF, National Retail Federation. It's a big tech and commerce event that takes place in New York, uh, every January in the cold, cold, uh, wind. That show is often a kind of a kickoff for what's the theme of the year going to be? And I think it's A.I. um, I think that may be the theme. And so I thought, well, who better to sort of introduce us, uh, and have a spicy conversation about what is true, what is not true, what is AI really? And helping make sense of that for folks in the manufacturing and distribution space who are absolutely bombarded with, um, ferocious claims right now from lots of ends. What is going on? So I think maybe let's start with some definition of terms. I have three terms I want you to define for our audience. We'll start with an easy1. What's AI?

Heather Hershey: Oh, gosh, you would think this would be an easy one. Um, but it's not really, because the problem is that AI has been around for a very long time. So of course it stands for artificial intelligence. And what that means can either be incredibly figurative or incredibly literal, depending on that interpretation. So I think some of this is kind of generational bias in a way. But it's also just a matter of how pedantic do we want to be? Because, I mean, if this concept has been around for. For, uh, almost an entire century at this point, it kind of does speak to the fact that everyone's gonna have some stake in it. We've seen many, many waves of it. In general, it's anytime a machine can approximate something that a human would otherwise be doing manually. And as these processes become more digitized, more philosophical questions about what that entails start to creep into the mix. So. So, um, it becomes more of an issue of what is intelligence, right? So there's a lot of philosophy. It's multidisciplinary. If you're a researcher, like I was, it could involve social science, it can involve psychology, a lot of computer science, uh, hardcore logic. It's kind of an amalgam of all of these things, because the goal is to try to approximate human intelligence and human intelligence. Again, it's a very subjective sort of thing. And is human, like, intelligence even the goal anymore? Like, if you talk to someone who's really advocating for something like asi, for example, obviously you've introduced another acronym.

Aaron Sheehan: You weren't supposed to do that, not till the end.

Heather Hershey: Okay, sorry, sorry, sorry, sorry. But, um, yeah, like, literally, uh, if what you want is a superior intelligence, right, Then what you're basically Assuming by default is that intelligence is not bound to something that looks like something a human can do. So AI in general is all these things and more. In the old school definition, it could have been a workflow engine, right? Like it could be something pretty basic with some if then triggers, but by some people's standards, that bar is significantly higher. And if it's not a machine that's literally doing all the thinking and all the execution, then it, it's not really AI.

Aaron Sheehan: Right, got it. So starting with the calculator and kind of, uh, the trusted TIA and moving, moving its way kind of up the food chain a little bit. So is AI the same thing as an LLM then?

Heather Hershey: Well, an LLM is a flavor of AI, so you've got AI, which again can be incredibly basic. It's this big nebulous umbrella that could mean essentially everything and nothing all at once. Which is why it's so horrible to use as a marketing term because it means again, nothing. But you drill down and you notice that there are different flavors of it. For example, machine learning. Right. Machine learning ML is basically when, uh, you get a machine to train itself. Right. Um, and one of the ways it does this is through neural nets. If you use a lot of neural nets to do more complex calculations of this nature, then you get what's called deep learning. An LLM is built upon this deep learning technology. So what an LLM is doing is all of this and more on a much, much larger scale. It's taking in literally billions of parameters of data m mainly in the form of words or code. Uh, it's good to remember that an LLM by design is meant to do something that communicates. The whole point of the exercise is that it communicates. So frankly, I think just a little side note that we should consider like using LLM when we're talking to it, or people who actually have to work within these models or really directly with them, but use nlp, natural Language Processing, when we are talking to marketers and people who are only interfacing with the chat, because the whole point of LLMs is that they enable NLP.

Aaron Sheehan: Got it. Okay. Well, that makes a certain amount of sense since to me, uh, of course we've had NLP around for quite some time, so that's not particularly new and sexy.

Heather Hershey: No. Um, but it's a very sophisticated way of doing it.

Aaron Sheehan: That's fair. We strive to be accurate on this podcast, uh, to inform and to, and to entertain at the same time. So what, uh, I have heard. How I have heard LLMs described is they are basically predictive text guessers in a certain way. They are their models. This deep learning that you speak of is trying to sort of almost type ahead and guess like, what's the next? At a very, very sort of like, very simplified level. What's the next character that makes mathematical sense probabilistically in completing or answering this sentence? I have also heard the term RAG to sort of describe the process of tuning or interacting with the LLM. Walk us through what RAG is and then we're going to stop at the acronyms and get into the like, unfounded opinions. But like, let's go.

Heather Hershey: Yeah. Okay, so RAG is Retrieval Augmented Generation. So again, what LLMs are doing is that they're generating content code language, right? So you can either use that, uh, public model, you know, ChatGPT, Gemini, whatever's available, you can develop your own models using like hugging face or something like that, or you could buy software or some other thing that enables you to take your own data sources and securely use that within the context of the model. So that way when the model is trying to figure out what the next best word or phrase is, it's literally matching it to, uh, information that you have in your own existing systems and have control over and is a little bit more proprietary, um, and specific.

Aaron Sheehan: Got it. That makes sense. So, you know, a lot of our, a lot of our customers here, a lot of our listeners, I expect, are manufacturers. And I think about like an error rate on a production line, like a 1% error rate on a production line would be actually pretty awful for a manufacturer. Like, that's a lot of waste, that's a lot of rework, that's a lot of wasted material and time and, you know, and, and really like, capacity. But if an LLM is fundamentally probabilistic, how do you reconcile sort of a technology that's kind of based around guessing with an industry like ours that requires a lot more certainty? I suppose.

Heather Hershey: Uh, I don't know if agentic AI is a comfortable fit for a lot of things that need to be done within the context of B2B commerce. Because, I mean, uh, I think about things like edi, for example, Right? Like, and I've made this argument many, many times. Um, it's, you know, EDI hasn't gone away and we've all heard of mechanisms that have tried to kill it. Um, you know, I've often heard the debate about, you know, why edi and not APIs, et cetera. Right. When they do fundamentally slightly different things. Right. Uh, I think that the reason why EDI and technologies of that nature persist is because they're secure. Right. Because you know, there isn't a lot of unknown going on there. It's very easily auditable and traceable and it accommodates really high volume transactions which we don't know a lot of that stuff about agentic AI when it comes to B2B payments at this point. Right. So that's just one example of how this touches commerce. But you know, when we're Talking about these LLMs, they are the biggest of possible black boxes. And um, you basically have to take it on faith when they tell you what it is that they're doing. And that's just going to be an uncomfortable fit any way you swing it.

Aaron Sheehan: Yeah, that's, that's, that, that makes a certain amount of sense. And certainly we, we at ORO use LLM integrations into the, into the platform to accomplish very specific outcomes, but they're very bounded and very trained, um, and focused on very specific problem sets and workflows that need to be automated or need to be evaluated. I think there's a large difference between AI and well, I'll say LLM because it's more precise to your earlier point. An LLM that is completely making decisions on its own. This is sort of the agentic kind of like concept and an agent that is helping someone, a person accomplish something complicated on their own. But there is a person continually engaged. And I think a little bit about this is kind of the, like the, the size matter question a little bit. It's the large language model. But as you pointed out, we've had nlp, we've had machine learning, um, we've had algorithms. Another fun word that we didn't, did, didn't define. Um, for a really long time, um, like m. A really, really long time. Obviously, I mean algorithms have been around for, depending on how you want to define it, centuries. Is an LLM overkill then compared to sort of like boring machine learning, which is sort of like a very rules based system. Uh, and what's, are you seeing a lot of repackaging of machine learning into generative AI and LLMs? And should we actually be a little relieved by that? Because what, what that means is that means the technology is actually boring and trustworthy. It's just being gussied up by the marketing people. Is that a fair way to think of it?

Heather Hershey: Yes, uh, to a certain extent. There's a profound amount of agent washing that I'm witnessing. Right. And the thing is I want to, want to do two things. One, I want to circle back to something that you mentioned because I forgot to answer it directly was the bit about the you know, smaller models. Um, I'm a huge fan of distillation. Um, it's less resource intensive and it's, it works very quickly. You can stand up those models very quickly compared to the parent, whatever that may be. Um, so I think that yes, there's definitely a place for smaller models that may be specifically purpose built for industries, um, and particular use cases that are not quite so broad and generalized. Because that's kind of the problem with the LLMs. They're incredibly generalized by default because all these companies are in the pursuit of AGI, artificial general intelligence. And there's this assumption that if they just you know, build bigger models uh, with, with more, you know, data and more um, parameters that they will be able to achieve that goal someday. Whether or not that's true is an entirely different thing. But here's the thing. A lot of that stuff um, is very cool and very buzzy and gets a lot of attention with the investor class. But I feel like professionals in commerce, particularly folks who are, you know, the rubber meets the road, they're in fulfillment or something like that. You know, hardcore ops where you know, not every part of it is digitized or, or can be digitized. You know, they're sitting there going what, what's in this for me? What does any of this have to do with my day to day? And I think that boring AI is very useful. Workflows are great for most parts of B2B commerce that at least that I'm aware of. You don't have to go back through and cleanse every single bit of the data that your organization has ever produced in its entire history in order to leverage it. Right. Like uh, and you have more control. And I think that that's really more to the point when you are considering adding anything agentic or that has this LLM technology to what you're doing in commerce, remember what it's there for, right? Like it's there to help you communicate. So in the case of having the chat interface, on the operational side it's to help you interface with operations or like different parts of the commerce platform or whatever software you're working in. But on the customer side, right, it's to help them, you know, whether that's with it self service or it's with you know, finding the right product, configuring products, maybe even someday things like that, right. Like those are all things that could potentially be done through chat. And then you have to ask yourself the next logical step. Would it make sense for someone to engage with a chatbot instead of just the ui? And if you can't answer that question definitively, maybe an LLM is overkill.

Aaron Sheehan: Yeah, but they're so cheap. They're so.

Heather Hershey: Sure, sure, yeah,

Aaron Sheehan: that's me being facetious, but I think. Yeah. So the technical reality that we see on our end is that for a lot of our customers and folks in our orbit in manufacturing, distribution, supply, wholesale, anyone kind of uh, in that into the supply chain, there's all of this sort of like cutting edge AI buzz about LLMs. But where the sort of center of percussion is for a lot of these businesses is an erp, often an old ERP with a green screen or you uh, know, a mainframe of some kind. Or just maybe it's, maybe it is like, you know it's a, it's, it's S4 or you know, it's, it's an old version of like Great Plains or something like that on the, on the accounting side. And then after years and years of merger and acquisition, what you actually have is now like lots of sources of truth about very specific pieces of data. And what we've seen is a lot of folks because B2B is often run by the finance and ops folks. And so reporting is kind of like the change agent I guess, enforcing like a data lake or data warehouse to be built that says, okay, we're going to take all of the little contained silos that we have acquired and still operate across all of these different lines of business. We're going to pour their contents into some kind of normalized data lake that's going to be our uh, we're going to use that as an insights and reporting engine. We're going to put a reporting engine on it and use that to generate insights compliance and all that kinds of stuff. What's the technical reality of trying to put an AI agent that does communication on top of a constellation of disconnected systems like that? Because that's been the goal for the last few years is best of breed. Right? We have very specific point solutions that do very specific things. Maybe they're really modern, but they're the best in breed and there's some kind of glue that holds them all together. But don't worry about that, they're really good separately and, or you've just inherited that organically over time because you bought, I, I bought uh, a, you buy a distributor here, you buy a distributor there next Thing you know, you got a bunch of distributors, you got a bunch, and you get a bunch of erps. What happens when you try to take an LLM and stick it in that construct and, and more, I think practically, what do you have to do if you're in business and you're being. You need to like, figure out where to put the LLM. Where do you start with something like that because your data is so fragmented.

Heather Hershey: Yeah, that's, that's really rough, honestly. Um, and there's actually a pretty huge market of startups, for example, that are trying to create these intermediary layers for exactly this sort of thing. Um, and, and I've seen this, you know, in various flavors. Whether, um, it's looking at this from a systems integration sort of perspective and, and how to harness all that data. But I mean that, that's even putting the cart before the horse a little bit. That's assuming that that data can even be used by that agent in some way. Um, and that it isn't redundant, et cetera. Um, you know, they, It's a complex thing. I think this is a lot of what I view is going on with the agentic AI push because, you know, as you notice when I try to reframe LLM as nlp, it becomes a little less sexy. Right? Like that is kind of what I've been doing as Captain Wet Blanket on some of these Biggie.

Aaron Sheehan: You have T shirts.

Heather Hershey: Uh, yeah, but it's like all, uh. There's a lot of buzz meant to generate hype for various reasons from various companies. That's cool. There's a lot of profound utility in doing what you've just described, Aaron. But the way to get there is unfortunately usually the same route that you've been hearing anytime a big push towards digital transformation hits the scene. And this is, this is one of those things, right? Just like the Metaverse, just like Covid, right? Like if, if you are a holdout, you have to, you know, get, get a cons, you know, a firm, an si, uh, somebody enabled to help you kind of get, um, all of your data, right? And that way you can create the data layer required to leverage this technology properly. And what you plan to do with it dictates how big the scope of that actually is. If you're trying to do this across multiple erps simultaneously, that's a big deal. And this, and the thing is it gets more expensive every year. So like, if you have been holding off on that, that's going to continue to be a problem. I would suggest that one thing though, that I've pretty much been a advocate for since the beginning is if you can find software that can help you take chunks of that at a time and rationalize it to leverage AI in very specific ways, like for customer service or something like that, that's really helpful. Right. And you can use that in tandem with the Strangler pattern to decouple your ERP if you want. And you don't want to do a whole rip and replace of that system because it's painful. But uh, you know, you don't have to go that far. You can also buy one of these apps that acts as an intermediary layer to create kind of that source of truth that the AI can then leverage for its processes.

Aaron Sheehan: Yeah, and, and, and for those listeners who are like, what on earth is a Strangler pattern? Um, that is, I'm gonna, I'm m gonna butcher this a little bit. But it is effectively a software design pattern where when you have a big system that does a lot of things, you start snipping off little pieces of it and giving that little piece to a better system to go do. So that over time slowly your giant behemoth of an IT stack becomes more decentralized and uh, you know, you are less dependent basically on one central database and one central application. It's named after the Strangler fig, I believe.

Heather Hershey: Yes.

Aaron Sheehan: Um, and we can talk about that plenty. I think one of this is. As an aside, I would like to get your reaction to this statement. Um, and rest assured, if you give the right answer, we will probably use it in some vertical video on LinkedIn after this publishes M. Yeah, yeah. So no pressure. What we see is that in B2B we'll call it E Commerce Digital Commerce. That's not necessarily the right word for it, but let's, let's say an E commerce platform often ends up as that orchestration layer of bringing context to a lot of disconnected data. They connect orders, transactions, financial records, inventory product data, customer data, customer master records, user tables, like a single sign on, quote history, customer service requests, returns, warranties, all of this stuff. That is often there's bits of that, the bits of that a customer experience sort of just like all over the floor like you had uh, just a bunch of dogs shedding everywhere. You've got a living room, you got shag carpet, there is fur everywhere. And kind of like that's, that's kind of where the data is. And E Commerce platforms in B2B are often the new sexy thing, but because they Sort of sit on top of all of those back office systems and are then used by buyers, customers and internal teams. They become the sort of like aggregator of all of that information and then become the natural place to put an LLM interface or a better uh, NLP for uh, natural language interface or boring old machine learning to generate predictive insights. Have you seen the same thing? Would you agree or disagree with that?

Heather Hershey: I 100% agree with that statement.

Aaron Sheehan: That's amazing.

Heather Hershey: LinkedIn is that, that's why I mentioned the strangler pattern, because you know, I'm, I'm like a broken record. I mention it all the time, but I, I think that it's kind of easy to interpret from how bearish I have been in previous sections of this conversation regarding the hype of AI. That doesn't mean that I don't think it has utility. And in commerce there are plenty of places to put LLM technology that are of high value. Right. Like again, if you want to interface with the system to pull up analytic data on the fly, that's a great way to do it. Right? If you're comfortable doing that with a chat interface on the front end. Again, I implore people to think about combining their IT help desk with the chat bot for customer success, customer service, whatever you're building on the front of your website because you could use that repository of data to seamlessly go from. So you have to have knowledge bases for both. Right, Is what I'm saying. You could use those repositories of information seamlessly within the same interface. Kind of like how Google combined their search bars back in the day. So you could just use one to rule them all. Right. That way you are not developing a bunch of chatbots on all over the place, all using different sources of truth, because you have to manage all those and update them. Right. It gives you the opportunity to kind of consolidate some of the ways that you interface with your customers on the front end through an actual novel interface that the chatbot. So there are many, many ways that you could, you can configure that, but that's just one example of how this can be leveraged in a novel way.

Aaron Sheehan: Exactly. It's a great use case for a customer portal where the customers can see their order history, which means the portal has to then know the full order history, then can raise a ticket or have an interaction about items in those orders and then there's full visibility around that also. Bonus points, Heather, for using the word seamlessly twice. In that explanation, you are ready for a career in product marketing, should you so choose, um, at least in SaaS, uh, or sewing. Okay, so let's say you're the CTO of a modern distributor and you have to spend a hundred thousand dollars on A.I. uh, to satisfy the board, but you actually want to see a return on that. Where would you put your money?

Heather Hershey: Two things. One, I would probably focus on security first. And that's not super sexy, right? Because it's not splashy. But I think it's necessary to see

Aaron Sheehan: a return, Heather, not flush it down the toilet.

Heather Hershey: No, the thing is this is, this is so important because as agents get more sophisticated, we can't assume that only good faith actors are going to have access to them. I definitely see an uptick of this and there is more concern about it. And the thing is that it goes hand in hand with enabling more of it for your own use. Right, because you would want to put more security in place and guardrails for what you're doing with your own data. But you want to make sure that whatever you're doing isn't also going to expose it to more bad faith actors in the process. So things like detecting when, uh, a bot, like an agentic procurement bot for example, is trying to uh, interface with your website or your agents as opposed to a human. That's going to be good information for the future. But I don't think investing in that right now is as important as just laying the groundwork for the security apparatus that would enable that kind of pattern recognition later down the road. And then I do think that sales enablement would probably be a really great place with this, you know, like guided sales, something of that nature. Um, something that really kind of streamlined some of those processes so customers don't have to sit there and wait for two days before, you know, sales rep gets back to them via email. That'd be great.

Aaron Sheehan: Well, and that's the, that's always been the promise of E commerce. But a lot of times in B2B you still have a sort of a complex RFQ. You know, somebody, somebody emails an RFP because they have a project they're bidding on and they're like, here's a spreadsheet full of stuff. Get back to me. You know, speed kills deals. If you take too long to respond to that RFP or that request, they may have moved on to, to a competitor at that point. So I totally hear you. Yo Gardner, uh, has been talking about digital sales rooms for, for some time. Guided, uh, selling of course is a big, a big topic there and it's something that you see a lot in manufacturing, especially that we see, which is a lot of guided selling for the sales reps to use in a CPQ tool. But they're rarely customer facing. And so there's still this sort of like asynchronous way workflow involved where a person still has to spend the time to deconstruct that spreadsheet and turn it into a bill of materials with a quote on it. And that is a great use case for AI. I agree. I think we will probably see some interesting stories on the security side of this, um, probably in 2026 where bad actors are using agents to go and try to gamify commerce. I think actually as you were talking I was reminded of a conversation I have had I don't know how many times over the last 15 years around SEO, which is part of the company, wants to keep the catalog under lockdown. You ah, got to have a login and a password to see the products and all. We can't let Google see our product. I don't care about SEO. We're, you know, our customers know who we are. They'll find us. They don't need our. They don't need to. They don't need to. The web doesn't need to know what we sell. Right. And then they're like, oh, but the web does need to know. People won't find us unless they're looking for a specific product. They'll start with the product and then they'll find their way to us and then they'll request a quote and then they'll become a customer. That the sort of. The two, the two orbits of that really agentic commerce, any kind of agent or crawler is following a very similar path that a googlebot is following for SEO purposes. So are, uh, maybe we're going to reignite that debate inside the boardrooms of let's say large distributors. Do we lock everything down to prevent. We're preventing E procurement bots, but we're also preventing Google and we're preventing people. Um, but we're safer. Or uh, maybe we'll open it up and then we'll have to invest. We'll have to take that $50,000. Let's say that um, Heather is spending on security to invest in uh, countermeasures against, against bad actors. Well, so I guess to that end then we should probably talk about the prisoner's dilemma.

Heather Hershey: Oh yeah. This is something that you'll see most often in D2C B2C commerce. Um, and it's mainly related to what's going on with the large language platforms themselves. So your OpenAI, you know they're chatgpt so Claude, maybe not so much, but you'll see it in perplexity. ChatGPT, Gemini, they are basically, and Amazon to a certain extent too are very much trying to brute force enforce this motion where they disintermediate all of D2C E commerce and basically just become one big marketplace aggregator to rule the mall, um, where you shop through chatbot forever and ever and ever.

Aaron Sheehan: I'd say Shopify is doing this pretty explicitly now too.

Heather Hershey: Yeah, um, I have some issues with this. Namely the example I like to use is thus why this value prop has a pretty glaring hole. I'm going to hang out at night and just have Netflix on in the background and I'm scrolling mindlessly adding crap to my cart, mindlessly. I might buy it. If they send me an email with a deal, I'll probably pull that trigger. But I'm um, mindlessly scrolling with very little direct intent. But I'm probably going to buy something. You have to have pretty strong intent to be able to communicate with a chat bot what it is that you actually want it to go buy on your behalf, which is the entire agentic commerce premise, right? That it's going to do your purchasing for you and it can't do that without you instructing it on what you're looking for, at what price point, etc. Etc. And you, you'll notice if you go into like Google's Gemini toolkit for, for agentic shopping, it's basically a workflow, uh, engine that, that you know, you're using to basically trigger it to purchase things on your behalf based on certain criteria that you're setting out in advance. It's very deterministic. Which again makes me question how agentic is all of that. Right? Because agents do have deterministic elements, but they are by and large prob. So uh, but anyway, what you're leaving on the cutting room floor are all of these mindless scroll sessions from somebody who didn't actually want to read an essay. They just wanted to look at pretty pictures and add stuff to their cart. So you know, those impulse buys matter, right? Um, and part of the problem with what's going on in um, the prisoner's dilemma concept is that for a lot of merchants this sounds like a terrible idea, right? They may want to put chat interfaces on their website that that version of uh, agentic commerce doesn't really offend them because they control it but you know, it is kind of out in left field for them to be completely cool with this dinner disintermediation play and to give up most of uh, top of funnel discovery which was basically hijacked whether they like it or not by these answer engines. And then now mid funnel and Commerce are up for grabs apparently. And that's, that's got some of them scratching their heads and going why do we need to do this? Because inherently, even if they can't put their, their finger on why, it's because it is a prisoner's dilemma. What a prisoner's dilemma is in game, uh, theory is essentially you got two prisoners and they're being interrogated separately by separate cops and they are told separately. Here's the deal. If you rat out your partner, you go free. The thing is, is that they would both go free if they both kept their mouths shut.

Aaron Sheehan: So you're talking about basically two, two rational individuals having conflicting self interest is in their interest to cooperate.

Heather Hershey: Right.

Aaron Sheehan: At some level. But also not.

Heather Hershey: No, it's. The thing is like the, the overlying logic is not in their, either their favor to talk.

Aaron Sheehan: If we take anything away from this, it's that LLMs are going to be sending people to jail. And I think I, I fall back to the, the wisdom that my Uncle Roscoe told me a long time ago, which is uh, just never talk to the cops. And then, you know, you make the rational uh, decision. The prisoner's dilemma. I'll leave it to your imagination listeners to decide, uh, to fill in the

Heather Hershey: mental picture entire Law and Order episode on top of this premise.

Aaron Sheehan: Huh? Yes. Excellent. But that'll be in the, the sequel, the, the follow up to this, this episode. Heather, you, you don't know, but we're, we're going to do a little true crime. Yes, exactly right. Uh, from the headlines.

Heather Hershey: But that's it. The merchants know this isn't to their benefit. Right. Because there might be a first adopter advantage. But that's evanescent. Right? Because eventually if enough people join, all it's going to really do is commoditize everything on the front end of commerce and turned most merchants into drop shippers for these LLM platforms.

Aaron Sheehan: I mean, I mean, uh. Laughs in Bezos like, I mean this is a little bit. I think we've, we've seen this particular play before and you know, bringing it Back to, to B2B. Obviously Amazon Business has grown a lot and is doing quite well. I think one of the reasons for that is they have fulfillment centers and OpenAI does not. And at the end of the day there's a value proposition, there's a disintermediating going on, but there's also a value being being delivered and, and thus far simply doing it with software, uh, it does not do that. So you've mentioned agentic commerce a lot. And uh, I haven't actually, you know, really dug into what that means and I don't even want to because it just makes my head hurt. As to what the actual definition of agentic commerce is, I think it's fair to say that, you know, for it to be a useful term, it needs to be different. Agentic AI needs to be different from other forms of AI. And the implication of agentic is that it does stuff that for you, it is acting as an agent on your behalf. It understands what your desires are and it goes out and tries to make those desires come true. That may be m, you know, negotiating the best possible price for a particular product. It may be, you know, returning a bunch of, buying a bunch of stuff and then returning a bunch of stuff, you know, things that are rational for it to do but not in the best interest obviously of the distributor, let's say, who's having to like bear all that cost. So what is your working model or definition of agentic commerce and in B2B specifically? Because I've seen a lot of predictions around, well, agentic commerce is going to disrupt eprocurement, Agent E commerce is going to disrupt portals, it's going to disrupt edi. Sort of one of your earlier points. What is your prediction for how we'll be engaging with this term in the B2B world in the next two years? What is going to get disrupted significantly by agentic commerce?

Heather Hershey: See, I don't, I don't really know how much of this is going to penetrate B2B because I, I definitely believe that the market, based on what I've seen and especially validated after NRF and seeing, you know, the scant handful of B2B vendors, uh, shying away from using agents in their go to market strategy at the event, I'm definitely of the opinion that my definition would still apply in a B2B context because I made it intentionally very, very broad and functional. Um, so just to cut chase, my definition is essentially based on what is making the decisions when it comes to the purchase, the actual payment and then the fulfillment. Right. Um, now not to say that every element of things like fulfillment need to be completely operated or that these things should be operating without guardrails or Some kind of human oversight, but that the majority of the decision making process is done by the AI itself via agentic means. And what that means on a technical level is that an agent is m more than just a reasoning engine. It basically is leveraging the LLMs, usually multiple, simultaneously while it is leveraging pieces of software as tools. Right. It's using those as tools. It's using all sorts of different things combined, uh, in this sort of multimodal way to run loops in pursuit of solving a problem, making a decision, et cetera. Right. Like whatever it's, it's been set to do, it's just going to keep looping and looping and looping until it determines probabilistically that it is come across the best answer and then uses the LLM, um, technology to generate that answer. So I think that it's not necessarily the most intuitive fit for B2B. Right. Because would you want an agent making the decision about the purchase? Right. Like that is, you know, the promise of a lot of what's going on in agentic procurement. Right. Like I would argue that where that would be most beneficial are for repetitive purchases of known products. And in that case maybe the purchasing decision isn't being done identically as much as the modeling about the predictive aspects of what should be ordered. Right. Like, so you kind of have to look at what is, what is actually being decided upon and what role the agent is playing in here to kind of make sense of it. Because if you are just creating a workflow like what Google has come up with for its agentic B2C purchasing, but you just create something like that for procurement that's still humans providing the logic, humans making the decision. If it's, if it's a situation where I have a checkout in a B2C context that's not agentic because I'm still manually going through a regular E commerce cart. And likewise if I have to submit a purchase order, um, you know, to a sales rep, I don't necessarily feel like there's much of anything agentic about that.

Aaron Sheehan: Uh, so yeah, that's fair. I mean for our product we have a workflow, you know, that handles that exact use case where a purchase ah, order gets emailed to a sales rep and then automatically basically ingested through inbox integration and then using uh, OCR and an LLM reads that purchase order in and translates it into a digital clone basically of that purchase order validates, and this is the important part, validates the guesses against what is on record for that Customer. So is that shipped to one that I've seen before on this particular account? Is this per line in price for that SKU accurate? Is that the right payment method? None of that is agentic in the sense of it is sort of like, you know, robot take the wheel. Like uh, it is making decisions about my business. For me it is an automated workflow where AI is being very specifically applied I think, but there's still what the result of that workflow is a dashboard and a queue for a person to come in and review what's there and then, you know, make a correction. It is a quality of life and automation tool. It is not running your business for you, which is I think what, where some of the talk seems to be today. Uh, and I think a little bit about what you said around edi. And now we. EDI is a standard.

Heather Hershey: Yeah.

Aaron Sheehan: Fundamentally it's just a standard of standard of documentation. Now we have more standards for AI with Google's UCP and OpenAI's. Uh, was it ACP? Right, yeah, ACP. I see. You get, you get, you get lost pretty quickly trying to keep track of them and everybody's got lots of different standards. But fundamentally a standard is just an agreed upon way of exchanging information. It's not replacing what a person is doing to some extent. Last sort of serious question then. I'd love to get some validation on this point. I have heard that no one's Googling anymore. The kid, the youth aren't Googling. A couple years ago the youth weren't Googling because they were just doing all their search in TikTok. Okay. Now the youth aren't Googling because they are doing all their searching in chatgpt. Um, isn't it kind of true that everyone is using LLMs now instead of searching for stuff?

Heather Hershey: I don't think we can trust any of the data that is coming out around that subject at this time. And I think the reason why I'm so skeptical about that is because they are adding it to every enterprise app under the sun and it is being added to the heart of all of our mobile devices and it's being added to, uh, everything, search everything. So I mean, I mean we've known for a long time in marketing that, you know, people don't like to scroll beyond the first third of the page, you know, below the fold. And so that. That is the golden real estate. Well, all that's been shifted over multiple tabs. And the golden real estate is the answer engine that the AI is generating. Which means that now the game is trying to figure out how to get yourself into that. So when people talk about how everyone is using LLMs, that's because it's the default. So if, if it's, you know, the customers are being asked to use this and they choose to use it, that's one thing. But if it's the convenient option because it's the default shoved in their faces, we all know, we have always known. And it's always centered around search. Aaron. Like the, the, the whole debacle with Internet Explorer back in the day. Right, right. Like the browser wars. Right. It was because that was the default. So people just stuck with it. We've known this for a really long time and so use that same skepticism and knowledge of, of tech history when thinking critically about these AI topics and the stats associated. Another thing I really want to put on people's radar is be really picky about the distinction between discovery and commerce. Because while discovery is certainly a part of commerce, particularly on the B2C side, obviously commerce is a lot m. More. Right. Because discovery usually happens near the top of the funnel and commerce, even B to C sense, is way at the

Aaron Sheehan: bottom, definitionally is the bottom of the funnel.

Heather Hershey: Yes, absolutely. And so I, uh, mean, you can't fix the problems at the top of the funnel by completely rerouting what's going on at the bottom. Um, and that's kind of the leap in logic that's going on with a lot of this conversation that conflates discovery and commerce. And I'm not going to name names, but if you look at my LinkedIn feed, you will see that I, I've been pretty critical of reports coming from certain organizations that are in the consulting world that do this prolifically on some of their most often recirculated reports. And it is maddening.

Aaron Sheehan: Well, are we sure that the lom M Didn't write the report? I think sure.

Heather Hershey: Wrote a huge. Honestly, it's not this, it's that, right?

Aaron Sheehan: M. Dash. M. Dash. M. Dash. Well, Heather, we're going to find out, I suppose, pretty quickly if the skepticism, uh, is warranted, if your predictions or anti predictions come true or don't come true, depending on what they are. Um, and so we should probably revisit this conversation in say 12 months and see like what is the state of agentic AI and commerce? Have the robots, um, come for us all and turn to nutrient paste? Uh, or are people still around and, and, and flourishing with um, bespoke AI tools? Or uh, are we, are we thriving? Who knows? There's only one possible way to find out, and that that will be to. To listen to this podcast. Well, I have one last question for you that I ask all of the guests. And if you've, um, if you've listened to this podcast before, you'll know what's coming. And if you haven't, this will be a glorious surprise. But we always ask, I always ask for one piece of media that you have consumed in the last few months that you would recommend to people. And sometimes this is a book, could be a novel, could be a nonfiction book, could be a podcast, could be a TV show, could be a movie, uh, a play, an opera, interpretive dance, whatever you want. A piece of art, a piece of media that you like and would recommend to other people. This is not graded, so you don't have to give a very sort of like, propeller head answer, that is extra credit. But you can certainly, if you want to. What do you have for the guests, Heather?

Heather Hershey: It's not very nerdy. I've been getting into cooking more of my own food lately, and so I think the, the most recent purchase and, and media consumption on my part has been from America's Test Kitchen. I got a cookbook about Mediterranean food and I'm so stoked. I need to make lambusaka and all these other things. Um, I am not necessarily the world's best, uh, knife expert, but I'm sure that I will find plenty of excuses to try to refine those skills as I experiment.

Aaron Sheehan: Well, I'm not especially because you, like, you did this motion. And so I don't know, like, I agree with you like, that you are not the foremost. Like, I would be nervous around you with a knife just based on the hand gesture that you made. Ah, but no, uh, America's Test Kitchen, Mediterranean cooking. I think that sounds amazing. Next, uh, time we meet, if you could bring hummus, that would be fantastic. We love, we love a good hummus at Oro. Well, really appreciate you coming, uh, on, on the podcast to sort of like reignite 2026 for us. And reignite is not our code, uh, word, uh, for the year. It's just a word I chose. But I think this was a good level setting for, uh, all of our audience. Uh, uh, for the B2B Uncut podcast, where can people. You've, you've mentioned it, but we're going to ask again because a lot of people skip to the end or they listen it to it to X speed. Where can people find you if they want to learn more?

Heather Hershey: I'm on LinkedIn Uh, so Heather Hershey won super easy. And it's spelled just like the candy bar. And yes, I am related, but I didn't inherit anything from that.

Aaron Sheehan: Not even. Not even a bar of chocolate. That's unfortunate.

Heather Hershey: If you go to Hershey's Chocolate World, though, they give everyone a chocolate bar at the entrance.

Aaron Sheehan: Everyone can be special at Hershey's Chocolate World and at Heather Hershey's LinkedIn feed. Uh, thank you so much, Heather. Really appreciate it. This was a blast, as always. We, uh, will talk later, and as I said, we will revisit this in a year to see whether or not humanity still exists. Thank you so much, and thanks all of you for listening. Goodbye.