AI Hype Detox for Manufacturers and Distributors with Heather Hershey

Heather Hershey: Thank you for inviting me on the pod. I’m Heather Hershey. I’m an analyst at a Tier 1 analyst firm. If you want to figure out which one, please go to my LinkedIn profile—HeatherHershey1 on LinkedIn.

But just so you know, what I’m about to say on this podcast is a reflection of my own opinion. So that comes with all the biases that would imply, and it is not a reflection of my employer. All right—caveat over.

My background is complicated, but like a lot of analysts, the path led to an interesting place where I’m talking to line-of-business people about IT things, and vice versa. I’m a former commerce professional and a former AI researcher. I went to UGA for AI for a master’s. My emphasis was in natural language programming and sentiment analysis, which has moved very far beyond what it was back in 2013, so I’m dating myself as a very old millennial.

I write about commerce—agentic commerce, B2B commerce, how B2C commerce platforms aren’t always the greatest fit for B2B commerce, and topics like that. So thank you for giving me an opportunity to be on my soapbox for a moment.

Aaron Sheehan: I love that, and it’s a theme that resonates strongly with us at OroCommerce. We were talking about what we wanted to discuss, and it felt like, to kick off 2026 on the heels of NRF—the National Retail Federation show in New York every January in the cold, cold wind—that show is often a kickoff for what the theme of the year is going to be.

I think it’s AI. I think that may be the theme. So who better to introduce us—and have a spicy conversation—about what’s true, what’s not true, what AI really is, and help make sense of it for folks in manufacturing and distribution who are absolutely bombarded with ferocious claims right now from lots of ends?

Let’s start with definitions. I have three terms I want you to define for our audience. We’ll start with an easy one. What’s AI?

Heather Hershey: You would think this would be an easy one, but it’s not. AI has been around for a very long time. Of course it stands for artificial intelligence, but what that means can be either incredibly figurative or incredibly literal, depending on how you interpret it.

Some of this is generational bias, but it’s also: how pedantic do we want to be? If this concept has been around for almost an entire century, everyone’s going to have a stake in it. We’ve seen many waves of it.

In general, it’s any time a machine can approximate something that a human would otherwise be doing manually. As processes become more digitized, more philosophical questions creep in. It becomes: what is intelligence? It’s multidisciplinary. It can involve social science, psychology, computer science, hardcore logic—an amalgam of all of these things—because the goal is to approximate human intelligence.

And human intelligence is subjective. Also: is human-like intelligence even the goal anymore? If you talk to someone advocating for ASI—

Aaron Sheehan: You introduced another acronym. You weren’t supposed to do that. Not till the end.

Heather Hershey: Sorry, sorry. But literally, if what you want is a superior intelligence, you’re assuming intelligence isn’t bound to something that looks like what a human can do. So AI is all these things and more.

In the old-school definition, it could be a workflow engine. It could be something basic with if-then triggers. But by some people’s standards, the bar is much higher. If it’s not a machine doing all the thinking and execution, then it’s “not really AI.”

Aaron Sheehan: Got it. So starting with the calculator and the trusty TI-80 and moving its way up the food chain a little bit. Is AI the same thing as an LLM, then?

Heather Hershey: An LLM is a flavor of AI. You’ve got AI, which can be incredibly basic—it’s a big, nebulous umbrella that can mean essentially everything and nothing all at once. That’s why it’s horrible as a marketing term, because it means… nothing.

But you drill down and there are different flavors. Machine learning—ML—is when you get a machine to train itself. One way it does this is through neural nets. If you use a lot of neural nets to do more complex calculations, you get deep learning. An LLM is built on deep learning.

So what an LLM is doing is all of this and more on a much larger scale. It’s taking in billions of parameters of data, mainly words or code. It’s good to remember: an LLM, by design, is meant to do something that communicates. The whole point is communication.

Side note: I think we should use “LLM” when we’re talking to IT or people who work directly with these models, but use “NLP,” natural language processing, when we’re 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. That makes sense. We’ve had NLP around for quite some time, so that’s not particularly new and sexy.

Heather Hershey: But it’s a very sophisticated way of doing it.

Aaron Sheehan: Fair. We strive to be accurate on this podcast—and entertain at the same time.

The way I’ve heard LLMs described is: they’re basically predictive text guessers. The deep learning is trying to “type ahead” and guess what the next character should be, probabilistically, to complete or answer a sentence. I’ve also heard the term RAG to 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 unfounded opinions.

Heather Hershey: Yeah. RAG is retrieval-augmented generation. LLMs generate content—code, language.

You can use a public model—ChatGPT, Gemini, whatever—or you can develop your own models using something like Hugging Face. Or you can buy software that lets you take your own data sources and securely use them within the context of the model.

That way, when the model is trying to figure out the next best word or phrase, it’s matching it to information you already have in your systems—data you control, that’s proprietary and specific.

Aaron Sheehan: Got it. Makes sense.

Aaron Sheehan: A lot of our customers—and a lot of our listeners—are manufacturers. And I think about something like an error rate on a production line. A 1% error rate would be pretty awful. That’s a lot of waste, rework, wasted material and time, and lost capacity.

But if an LLM is fundamentally probabilistic, how do you reconcile technology that’s based around guessing with an industry like ours that requires a lot more certainty?

Heather Hershey: I don’t know if agentic AI is a comfortable fit for a lot of what needs to happen in B2B commerce.

I think about things like EDI. I’ve made this argument many times: EDI hasn’t gone away, and we’ve all heard about mechanisms that tried to kill it. People debate “why EDI and not APIs,” even though they do slightly different things.

The reason EDI and technologies like it persist is that they’re secure. There isn’t a lot of unknown going on. It’s auditable and traceable, and it handles really high-volume transactions.

We don’t know a lot of that stuff about agentic AI in a B2B context yet. And when we’re talking about LLMs, they’re basically the biggest possible black boxes. You have to take it on faith when they tell you what they’re doing. That’s going to be uncomfortable no matter how you frame it.

Aaron Sheehan: That makes sense. And certainly, at OroCommerce, we use LLM integrations to accomplish very specific outcomes—but they’re bounded and focused on specific workflows that need to be automated or evaluated.

There’s a big difference between an LLM making decisions on its own—the agentic concept—and an agent that helps a person accomplish something complicated, with a human continually engaged.

That leads to another question. We’ve had NLP, machine learning, and algorithms for a long time. Is an LLM overkill compared to “boring” machine learning—more rules-based systems?

And are you seeing a lot of repackaging of machine learning into generative AI and LLMs? Should we be relieved by that—because it’s actually boring and trustworthy, just dressed up by marketing? Is that a fair way to think about it?

Heather Hershey: Yes, to a certain extent. There’s a profound amount of agent washing happening.

Let me circle back to something you mentioned earlier, too—the idea of smaller models. I’m a huge fan of distillation. It’s less resource intensive, it works quickly, and you can stand up those models much faster than the parent model.

There’s definitely a place for smaller, purpose-built models for specific industries and use cases that aren’t so broad and generalized. That’s kind of the problem with LLMs: they’re generalized by default, because a lot of companies are chasing AGI—artificial general intelligence. The assumption is: build bigger models with more data and more parameters, and you’ll get there someday. Whether that’s true is a different conversation.

A lot of that is cool and buzzy and gets investor attention. But commerce professionals—the people where the rubber meets the road, in fulfillment, hardcore ops, where not everything can be digitized—are sitting there going: What’s in this for me? What does this have to do with my day-to-day?

Boring AI is very useful. Workflows are great for most parts of B2B commerce. You don’t have to cleanse every single piece of data your organization has ever produced just to leverage it. And you have more control. That’s really the point.

When you’re considering adding anything agentic—or anything that uses LLM technology—remember what it’s there for. It’s there to help you communicate.

On the operational side, a chat interface can help you interact with operations or different parts of your commerce platform, pull up analytics on the fly, things like that. On the customer side, it can help with self-service, finding the right product, configuring products—maybe someday.

But you have to ask the next logical step: Would it actually make sense for someone to engage with a chatbot instead of the UI? If you can’t answer that definitively, an LLM might be overkill.

Aaron Sheehan: Yeah, but they’re so cheap.

Heather Hershey: Sure, sure!

Aaron Sheehan: That was me being facetious. But the technical reality we see is: for a lot of manufacturers and distributors, the center of gravity is an ERP—often an old ERP with a green screen or a mainframe, or something like that.

And then after years of M&A, what you actually have is lots of sources of truth for specific pieces of data. A lot of these businesses are run by finance and ops, so reporting becomes the change agent. They build a data lake or a warehouse and say, “We’re going to take all these silos we acquired, pour them into a normalized data layer, and use it for insights, reporting, compliance.”

So what’s the reality of trying to put an AI agent—something built for communication—on top of a constellation of disconnected systems like that?

Because for the last few years the pitch has been “best of breed”: point solutions that do specific things, plus some glue that holds them together. Or you inherited it organically. You buy a distributor here, you buy a distributor there, and next thing you know, you’ve got a bunch of distributors—and a bunch of ERPs.

What happens when you try to stick an LLM into that construct? And practically: if you’re running the business and trying to figure out where to put it, where do you even start when your data is so fragmented?

Heather Hershey: Yeah, that’s really rough, honestly. And there’s a pretty huge market of startups trying to create intermediary layers for exactly this problem.

I’ve seen it in various flavors—systems integration approaches, data harnessing approaches—but it can still be cart-before-the-horse. That assumes the data can even be used by the agent in a meaningful way, that it isn’t redundant, and that you can rationalize it enough to be useful.

A lot of what’s happening with the agentic AI push is hype. When I try to reframe LLM as NLP, it gets a little less sexy—so I’ve basically been Captain Wet Blanket on some of this.

Aaron Sheehan: Do you have t-shirts?

Heather Hershey: Yeah—but there’s a lot of buzz meant to generate hype for various reasons, from various companies. That’s fine. There’s also real utility in what you described.

But the way to get there is usually the same route you’ve heard any time a big push toward digital transformation shows up—metaverse, COVID, you name it. If you’ve been holding out, you typically need help getting your data in order so you can create the data layer required to leverage this properly. And what you plan to do with it dictates how big that scope is.

If you’re trying to do this across multiple ERPs simultaneously, that’s a big deal. And it gets more expensive every year you wait.

One thing I’ve advocated for from the beginning: if you can find software that helps you take this in chunks and rationalize it in specific ways—customer service is a common example—that helps. You can use that in tandem with the strangler pattern to decouple your ERP, without doing a painful rip-and-replace.

You can also buy one of these apps that acts as an intermediary layer—creating a source of truth the AI can leverage.

Aaron Sheehan: For listeners who are thinking, “What on earth is a strangler pattern?”—that’s a software design pattern where you have a big system that does a lot, and you start snipping off little pieces and handing them to a better system over time. Slowly, your giant IT behemoth becomes more decentralized, and you’re less dependent on one central database and one central application. It’s named after the strangler fig.

Heather Hershey: Yes.

Aaron Sheehan: As an aside, I’d like your reaction to a statement. And if you give the right answer, we’ll probably use it in some vertical video on LinkedIn after this publishes—so no pressure.

What we see in B2B digital commerce is that an eCommerce platform often ends up as the orchestration layer that brings context to disconnected data. It connects orders, transactions, financial records, inventory, product data, customer data, customer master records, user tables, single sign-on, quote history, customer service requests, returns, warranties—stuff that’s scattered everywhere.

Because it sits on top of those back-office systems and gets used by buyers, customers, and internal teams, it becomes an aggregator of all that information—and then the natural place to put a natural-language interface, whether that’s an LLM chat layer or “boring” machine learning for predictive insights. Have you seen the same thing? Would you agree or disagree?

Heather Hershey: I 100% agree with that statement.

Aaron Sheehan: Amazing. LinkedIn is so happy.

Heather Hershey: That’s why I mentioned the strangler pattern. I know I’ve been bearish in parts of this conversation about the hype, but that doesn’t mean I don’t think AI has utility.

In commerce, there are plenty of high-value places to put LLM technology. If you want to interface with the system to pull up analytic data on the fly, that’s a great use case—if you’re comfortable doing that through a chat interface.

On the front end, I’d also push people to think about combining their IT help desk with a chatbot for customer success or customer service. You already have to maintain knowledge bases for both. If you structure it well, you can use those repositories within the same interface—so you’re not building a bunch of chatbots across the organization, each with different sources of truth you have to manage and update.

It’s an opportunity to consolidate how you interface with customers through a genuinely new interface: the chatbot. There are a lot of ways to configure it, but that’s one example of how this can be leveraged in a novel way.

Aaron Sheehan: Exactly. A great example is a customer portal where customers can see order history—which means the portal has to know the full order history—then raise a ticket or ask questions about items in those orders, with full visibility around it.

Also, bonus points for using the word “seamlessly” twice. You’re ready for a career in product marketing.

Okay. Let’s say you’re the CTO of a modern distributor and you have to spend $100,000 on AI to satisfy the board—but you actually want to see a return. Where would you put your money?

Heather Hershey: Two things. First, I’d focus on security.

It’s not splashy, but it’s necessary. As agents get more sophisticated, we can’t assume only good-faith actors will have access to them. There’s already an uptick in related activity—there’s data I can’t quote directly, but it’s real.

Security goes hand-in-hand with enabling more AI for your own use. You need guardrails around your data and your implementations, and you need to make sure you’re not exposing yourself to more bad actors in the process.

For example: detecting when a bot—like an agentic procurement bot—is trying to interface with your website or your systems, versus a human. That’s going to matter. But I’m not sure investing heavily in that detection right now is as important as laying the groundwork: security controls, governance, and an architecture that can support better pattern recognition later.

Second, I think sales enablement is a great place to invest. Guided selling, streamlining buying processes—so customers aren’t waiting two days for a sales rep to respond to an email. That’s meaningful.

Aaron Sheehan: That’s always been the promise of eCommerce, but in B2B you still have complex RFQs. Someone emails an RFP because they’re bidding a project, and it’s a spreadsheet full of items. If you take too long to respond, they move on. Speed kills deals.

Gartner’s talked about digital sales rooms for a while. Guided selling is big. In manufacturing, we see a lot of guided selling inside CPQ tools, but it’s rarely customer-facing. So there’s still an asynchronous workflow where someone has to deconstruct that spreadsheet and turn it into a bill of materials and a quote. That’s a strong use case for AI.

On the security side, I think we’re going to hear some interesting stories in 2026—bad actors using agents to game commerce.

It also reminds me of a conversation I’ve had a hundred times around SEO. Part of the company wants the catalog locked down—login required, keep Google out. “Our customers know who we are.” Then someone else argues the opposite: people start with the product search, find you, request a quote, and become a customer.

Agentic commerce—any crawler or agent—is following a similar path to a Googlebot. So maybe we reignite that debate in distributor boardrooms: lock everything down to block procurement bots, but also block discovery and people… or open it up and invest in countermeasures.

We should probably talk about the prisoner’s dilemma.

Aaron Sheehan: We should probably talk about the prisoner’s dilemma.

Heather Hershey: You see this most often in D2C and B2C. It’s mainly tied to what’s happening with the big language platforms themselves—ChatGPT, Gemini, Perplexity—and to some extent Amazon.

A lot of them are trying to brute-force a move where they disintermediate D2C commerce and become the marketplace aggregator where you shop.

Aaron Sheehan: Through chatbot, forever and ever. Shopify is doing this pretty explicitly now, too.

Heather Hershey: I have issues with it. Here’s the example I use for why the value prop has a glaring hole.

I’ll be up at night with Netflix on in the background, scrolling and adding random stuff to my cart. I might buy it. If they email me a deal, I’ll probably pull the trigger. It’s low-intent, but it converts.

To buy through a chatbot, you need strong intent. You have to be able to describe what you want, at what price point, with what constraints. That’s the whole agentic commerce premise—that it purchases on your behalf.

And if you look at Google’s Gemini toolkit for agentic shopping, it’s basically a workflow engine. You’re setting criteria in advance and triggering purchases. It’s deterministic.

Which makes me question how “agentic” any of that really is. Agents can have deterministic elements, but they’re still largely probability engines.

But either way, what gets left on the cutting room floor is all that mindless browsing. Some people don’t want to talk to a bot—they want to look at pictures and add things to a cart. Those impulse buys matter.

For a lot of merchants, the disintermediation play sounds like a terrible idea. A chat interface on their own site doesn’t bother them as much because they control it.

But giving up top-of-funnel discovery—because answer engines hijack it—and now mid-funnel and commerce are apparently up for grabs too? That has merchants asking, “Why would we do this?”

And even if they can’t fully articulate it, it’s because it functions like a prisoner’s dilemma.

A prisoner’s dilemma is a game theory scenario: two prisoners are interrogated separately. Each is told, “If you rat out your partner, you go free.” But if they both keep quiet, they both do better.

Aaron Sheehan: Right.

Heather Hershey: It’s two rational people with conflicting incentives. It’s in their interest to cooperate—at some level—but also not, because they’re being asked to take it on faith that the “deal” will be honored.

If enough merchants participate, any early advantage disappears. Eventually, it commoditizes the front end of commerce and turns most merchants into dropshippers for these LLM platforms.

Aaron Sheehan: I mean… laughs in Bezos. We’ve seen this play before.

Bringing it back to B2B: Amazon Business has grown a lot and is doing quite well. One reason is they have fulfillment centers. OpenAI doesn’t.

There’s disintermediation, sure, but there’s also real value being delivered. So far, pure software doesn’t replace that.

You’ve mentioned agentic commerce a lot, and I haven’t dug into what it actually means—partly because it makes my head hurt. For it to be a useful term, agentic AI has to mean something distinct: it does things for you. It acts on your behalf. It negotiates, purchases, returns—whatever it’s instructed to do—sometimes in ways that are rational for the agent but not exactly in the distributor’s best interest.

So what’s your working definition of agentic commerce in B2B? And what do you think it disrupts over the next couple of years—eProcurement, portals, EDI, all the things people keep predicting?

Heather Hershey: I don’t really know how much of this is going to penetrate B2B. Based on what I’ve seen—and especially after NRF, where even the small handful of B2B vendors there seemed to shy away from using “agents” in their go-to-market—I’m skeptical about how far this goes in the next couple of years.

My definition still applies in a B2B context because I try to keep it broad and functional. For me, it comes down to: what’s making the decisions when it comes to the purchase, the payment, and the fulfillment.

That’s not to say every element of fulfillment needs to be fully automated, or that any of this should run without guardrails or human oversight. But the question is whether the majority of the decision-making is being done by AI.

Technically, an agent isn’t only a reasoning engine. It usually leverages one or more LLMs while using software as tools. It’s combining different inputs in a multimodal way and running loops in pursuit of solving a problem—making a decision, completing a task. It keeps looping until it determines, probabilistically, that it has the best answer, then uses LLM technology to generate the output.

And I don’t think that’s an intuitive fit for B2B. Would you want an agent making a purchase decision? That’s the promise in a lot of the agentic procurement narrative.

Where it could be useful is repetitive purchases of known products. But in that case, it may not be “agentic purchasing” as much as predictive modeling—what should be ordered, when, in what quantities. So you have to look closely at what’s actually being decided and what role the agent is really playing.

If what you’re building is basically a workflow—like what Google has put together for agentic B2C purchasing—that’s still humans providing the logic and humans making the decisions.

If you’re in a normal B2C cart and checkout flow, that’s not agentic. You’re still doing it manually. And if you have to submit a purchase order to a sales rep, there isn’t much agentic about that either.

Aaron Sheehan: That’s fair. In our product, we have a workflow that handles that exact use case: a purchase order gets emailed to a sales rep and then automatically ingested through inbox integration. Using OCR and an LLM, it reads the PO and translates it into a digital clone of that purchase order.

But the important part is that it validates the guesses against what’s already on record for that customer. Is this ship-to address one I’ve seen before on this account? Is the per-line-item price for that SKU accurate? Is that the right payment method?

None of that is agentic in the “robot take the wheel” sense. It’s not making decisions about my business for me. It’s an automated workflow where AI is applied in a specific way.

And the output is a dashboard and a queue for a person to review, correct if needed, and approve. It’s a quality-of-life and automation tool. It’s not running your business, which is where some of the conversation seems to be going.

It also makes me think about what you said around EDI. EDI is a standard—fundamentally, it’s just an agreed-upon way of exchanging information. Now we have standards for AI too, like Google’s UCP and OpenAI’s… is it ACP?

Heather Hershey: Yes.

Aaron Sheehan: You get lost pretty quickly trying to keep track. But the point is: a standard is just an agreed-upon way of exchanging information. It’s not replacing what a person is doing.

Last serious question. I’d love some validation on this point.

I keep hearing that no one Googles anymore. A couple years ago, the story was that younger people weren’t Googling because they searched on TikTok. Now the story is they aren’t Googling because they search in ChatGPT. Is it kind of true that everyone is using LLMs now instead of searching for stuff?

Heather Hershey: I don’t think we can trust the data coming out about that right now. The reason I’m skeptical is that LLMs are being added to every enterprise app under the sun. They’re being built into our phones. They’re being injected into search—into everything.

In marketing, we’ve known forever that people don’t scroll very far. That first slice of the page is prime real estate. Now that’s been shifted, and the new prime real estate is the AI-generated answer at the top.

So when people say “everyone is using LLMs,” part of that is because it’s the default. If customers choose it when it’s optional, that’s one thing. But when it’s the convenient option shoved in your face, adoption stats can get misleading.

We’ve seen this movie before. It’s always been about defaults—think about the browser wars and Internet Explorer. People stuck with what was shipped to them.

Use that same skepticism and tech-history memory when you’re evaluating AI stats.

Also, be picky about the distinction between discovery and commerce. Discovery is part of commerce—especially in B2C—but commerce is more than discovery. Discovery happens near the top of funnel. Commerce is at the bottom.

Aaron Sheehan: Definitionally the bottom of the funnel, yes.

Heather Hershey: Exactly. You can’t fix top-of-funnel problems by rerouting what happens at the bottom. That leap in logic shows up in a lot of AI conversation, where discovery and commerce get conflated. I’m not going to name names, but if you look at my LinkedIn feed, you’ll see I’ve been pretty critical of certain consulting-style reports that do this constantly. It’s maddening.

Aaron Sheehan: Are we sure an LLM didn’t write the report?

Heather Hershey: Honestly…

Aaron Sheehan: Okay. Well, we’ll find out soon enough whether the skepticism is warranted. We should revisit this in 12 months and see where agentic AI in commerce actually is—did the robots turn us into nutrient paste, or are we thriving with more bespoke tools?

There’s only one way to find out: listen to this podcast.

I have one last question that I ask all guests. Recommend one piece of media you’ve consumed in the last few months—a book, a podcast, a TV show, a movie, a play, an opera, interpretive dance—whatever. What should people check out?

Heather Hershey: It’s not very nerdy. I’ve been cooking more lately, and the most recent thing I’ve bought—and used a lot—is from America’s Test Kitchen. I got a Mediterranean cookbook.

I’m stoked. I want to make lamb moussaka and a bunch of other things. I’m not necessarily the world’s best knife expert, but I’m sure I’ll find plenty of excuses to refine those skills as I experiment.

Aaron Sheehan: I’m not sure I’m comforted by that, especially based on the hand gesture you just made. I’d be nervous around you with a knife.

But America’s Test Kitchen, Mediterranean cooking—that sounds amazing. Next time we meet, bring hummus. We love a good hummus at OroCommerce.

Really appreciate you coming on to kick off 2026 for us. Where can people find you if they want to learn more?

Heather Hershey: I’m on LinkedIn—HeatherHershey1. Super easy. 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 a bar of chocolate. That’s unfortunate.

Heather Hershey: If you go to Hershey’s Chocolate World, 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.

Thank you so much, Heather. This was a blast, as always. We’ll talk later, and we’ll revisit this in a year to see whether or not humanity still exists.

Thanks to all of you for listening. Goodbye!