Latest / Machine Dreams / How to Make Your AI Work in B2B Commerce

Transcript

How to Make Your AI Work in B2B Commerce · Machine Dreams

• speaker-0: Hi everyone and welcome back to Machine Dreams, the show where we get real about what's working in AI today, what's really not, and how it could all impact us. I am Colabole Samuel, a deviant technology analyst covering AI, cybersecurity, and enterprise technology. And as always, ⁓ as usual, my course would be here, but unfortunately she's not here. Leah is observing our Passover holidays. So it's just me, but we're going to have a great time nonetheless. Today I'm talking with Aaron Sheehan. He is the VP of Strategy and Partnerships at Oro Inc. The company behind Oro Commerce, a B2B e-commerce platform built specifically for manufacturers, distributors and wholesalers. Before this, Aaron ran competitive strategy at B Commerce. He also hosts his own podcast. called Beats will be Uncaught. Aaron, welcome to Machine Dreams. speaker-1: Thank you very much, Sam. It's great to be here. speaker-0: Yeah, great to have you. And I was just telling you the other time about how much of a great setup you have and you know, it's really fascinating the artwork and the design. speaker-1: Thank you. You are glad that the clutter the clutter comes across attractively on screen. speaker-0: I would call it creative chaos. speaker-1: There we go, I like that, I may put that on a t-shirt. speaker-0: Just make sure you pay me some commission for it. course. Great. So, work with manufacturers, distributors, wholesalers every day. And so really curious about what problem most of them are actually trying to solve when they come to you. Not the one that they think they should have. I mean, the real one, the real problem. speaker-1: Of course, always. Yeah, it's interesting. would say when it comes to businesses who sell B2B, ⁓ they all look a little bit different from one another because they tend to have organically grown up over time. And the problem that they actually have usually is pretty simple. it's all of these businesses have a series of back office systems. They have an ERP. that manages maybe pricing and timesheets and some inventory and financials. They may have an order management system. They may have a customer relationship management system. They may have a help desk system. What they really have most of them is like 10,000 Excel spreadsheets that are on people's desktops or laptops or on SharePoint or something like that. And their customers don't know any of that, right? Their customers want to buy from them. They want to get the right product. They want to get the right quantity. They want to know where's my order, what happened to it? Can I get more? Can I split between these two locations? ⁓ Things like that. And today, what has been the case really for decades is that the user interface between all of those separate systems and spreadsheets and the actual customers is a call center and a ton of people. And it tends to make finding information very, I would say, manual and labor intensive. So we have a lot of conversations where people, sort of introduce us correctly as a B2B e-commerce platform. It's absolutely correct. A lot of our customers don't think that what they need is an e-commerce platform. E-commerce is like a marginal channel for them if it exists at all. What they really want is a portal. The word portal makes complete sense. Everybody knows what that is. And what they want is they want a window into their back office systems that their customers can interact with without having to pick up a phone, without having to send an email and learn things and do things and interact and buy more stuff and self-service their account. But they also want to be able to do that and scale it. And they want to make more money on those accounts. So they want to be able to upsell and cross sell. Hey, we have a new product. Push an alert out to those people so they can buy it, right? Today, a lot of that is very labor intensive. If you talk to people about basic things like, how long does it take to pay an invoice? How many people have to touch that piece of paper before an invoice can be paid? Sometimes it's a staggering amount. What if we could digitize that? So I would say digitization for the goal of better customer experience and automation of processes, that is the problem they are trying to solve. And we are one solution to that. speaker-0: Great. ⁓ So where does AI actually fit into that? Because I imagine that a lot of these companies are being pitched AI solutions that sound great in demos, but don't actually address the core problem, which is actually a problem across ⁓ several other industries. ⁓ It's something that I've heard. I don't know if it's the same in your industry, but a lot of companies are getting pitched flashy demos that don't actually fix the problem that they are trying to solve for. So I'm curious to know where actually AI fits into all of these great things that you've just talked about. speaker-1: That's a pretty good point. think our general point of view on this is that ⁓ AI should be making the experience frictionless and faster and better. should speed and accuracy have to be maintained. A team of people who are trained to work in an ERP, they're listening, they've got their headset on and they're tinkering away. They're very fast. Mostly reliable. Your sales team may be a different story, but like if you've got a trained call center, they're pretty good at things like entering orders in off of a fax or off of a call or off of an email that comes in. Where we see AI fitting into for manufacturers and distributors is take the workflows that you have already and digitize them and scale them and make them work faster. ⁓ It means basically meeting your buyers in the channel that they are used to transacting in. So you're, you're just trying to make the existing touch points go faster. We think of it like this and you said AI, I'm going to talk about LLMs specifically because AI has a lot of different meanings. But an LLM is effectively, right. It's effectively, it's a natural language interface to do stuff. I talk to it. I literally talk to it or I type into it. Like I'm having a conversation and it figures out what it is I'm actually trying to do based on the context of who I am needs to know who I am. It needs to know what kinds of things I'm trying to do, kinds of things I'm allowed to do. And then it is taking actions behind the scenes based on what it thinks I'm trying to do. That has to be pretty accurate. That has to have guardrails around it. Obviously, you're often dealing with very large transactions. And so ⁓ what we're trying to do is take LLMs, take AI. and put it into the stuff that we're already doing as a platform because those are the problems that people are hiring us to solve, not totally different shiny flashy problems that we are being hired to and licensed to solve a very specific need in the business. speaker-0: Yeah. On the back of that, maybe you can talk now more specifically about some of the things you've built at OroConverse. I know that you've built a smart order, which extracts data from, you know, email PDFs and a lot more. And I know that you've also built smart agents, which lets buyers order through conversation instead of navigating or going through a complex ⁓ portal. So could you walk me through the practical impact of this? products and what was broken before, what was the challenge and ⁓ how does this solve this challenge? speaker-1: Absolutely. the just in general, what was broken before is just often a lot of time wasted. So it takes what we've seen on average about 30 minutes for a person to open up Outlook, open up an email from a customer, a buyer of theirs. They've got a PDF of a purchase order. Everybody's purchase orders look differently. So you've got a sort of person has to scan it and figure out where's the where the items where's it shipping to? How are they paying? Who is it for all of that stuff? That's 30 minutes. That's not wasted time in the abstract. Those are a series of steps, basically, that someone is engaging in. And the question, the failure point or the problem, the thing that's broken is, is the highest and best use of my sales rep's time manually typing text from one document into a computer or from one computer into another computer? The answer is generally no. And if that could be done more reliably without having to spend 30 minutes per order doing it, we would like technology to do that. That's now possible. And it's a free feature of ours. If you're an oral commerce customer, you get access to that. So what you're describing is sort of the smart order capability that PDF becomes a digital order instantly and works across very large, you know, hundreds of line items, know, dozens and dozens of pages even on the, the PDF or the spreadsheet. ⁓ And the other side of that is what happens if there's a mistake. So If I'm typing, if somebody has sent me like a 30 page purchase order, there are probably hundreds of line items on there. ⁓ If I type the wrong SKU, there's a substitution that I need to make that I don't or the quantity, I type the quantity wrong. Like because these are B2B transactions, the impact of that, you're now impacting someone's business. It's not like I ordered the wrong color sock. ⁓ well, right. If I ordered the wrong size of coupling, a project may end up being delayed for weeks and now I've got a big business problem in my hand and who am I going to be mad at? The person who types the order in wrong for me, right? And so I think that you're taking what is some level of cognitive work that a human is doing, reading a document your customer has sent you. And if you can programmatically do a reliable job of extracting that information for them, they can then come over the top, approve, make edits if they want. But like 80 % of that work has now been done for them. And now they're focusing on the details, right? And so the Pareto principle is always at work, which is that 20 % of the, causes 80 % of your problems. So that's very much true in B2B. speaker-0: You're neck deep in this world and you have great experience doing what you do. My question is how do you decide what's worth building versus what's just AI for the sake of having AI? speaker-1: Yeah, it's a really great question. so we, this is, will say, you know, we have constant founder and executive conversations about this topic here at, at, or this is a company mostly of technologists. So this is a, and this is really what you just asked. It's absolutely applicable to AI, but frankly, it's applicable to any software that any software company builds, which is, you building features for feature sake? Are you adding value for your, for your customers? AI power is sort of like a feature of everything that's being sold right now. There's a lot of hype, there's a lot of demos, AI power literally everything. ⁓ you know, the honest question to ask is what decision is AI making that a human used to make and what happens when it's wrong, right? And that's, there's a lot of black box thinking in AI, which is like, I'm going to take this magic thing called AI and I'm going to plop it down in the middle of my business. And it's a magic eight ball. It's just going to give me answers. How did it come up with the answers? I don't know. Like, ⁓ Will it come up with this? Can it reliably give the same answer to the same question every single time? No, LLMs aren't good at that. Consistency is not something that they're particularly good at. So you have to build guardrails and rules around it. so the question that we're asking is, where is AI fitting into the sphere of things that we are doing as a company? And how are we layering the right auditing and logging and decision making? And honestly, is a large language model, the right tool when maybe a very small language model or a deterministic piece of a deterministic algorithm is actually the better, cheaper to run, more reliable solution. think it's not putting LLMs absolutely everywhere and having them sort of take the place of deterministic programming, which is kind of a thing that I think is it. There's a little bit of risk in doing it that way because LLMs are not inherently consistent. So I think you have to take AI in the B2B world, you have to use it in narrow and specific ways where there's a well-defined input, a well-defined output, and you can measure and track the outcome. speaker-0: Great. You know, on the back of that, and this is also from something that I read on your company blog, which I think is really interesting. There was an article called, titled, The Honest Guide to AI Claims and B2B Commerce. I think it's a really great article. If anyone is watching, our viewers, I mean, you should definitely go read that article. My question is, you know, if I'm a wholesaler and I'm not so technical, I don't really understand so much about the AI world, but I've I've been hearing a lot about it and I'm curious and obviously there's a lot of hype as you've mentioned before now. What should I be looking for, you know, when I'm making a decision to use a platform like yours? ⁓ What are companies being sold right now? I guess that's my big question. What are companies being sold right now that does not actually work? speaker-1: A lot of time, so there's been a lot of survey data on this and we've got some as well, but generally what we were finding that buyers are looking for, I'm meaning manufacturers and distributors buying software are looking for is they're looking for AI tools that fit inside their existing stack of tools. One of the challenges that you have with ⁓ plugging in outside AI vendors into your existing business is a data problem. Typically, the data that the AI, I'm talking about the AI like it's the Wizard of Oz, it is this magic all-knowing sort of like Oracle, but we'll just, we'll keep talking about it that way, because that's certainly how it's being sold. ⁓ The data and business rules that you want the AI to reliably be using are just spread out across all of those spreadsheets and all those separate applications. So when you introduce another piece of software, Now you have to figure out how to normalize and ingest all of not just the data, but the decision tree that you want the AI to follow the rules you want it to sort of like live inside and the guardrails, etc. What we find is that people are looking for their existing vendors to please solve the AI problem for them. How are you? The question that we get is sort of like what you asked, which is As our service provider, what are you, how are you guys investing in AI? How are you guys surfacing AI tools to me in a way that doesn't require me to re platform something or, you know, implement middleware or create a whole data governance project that I don't want to do right now. Can you make our AI work on you? That's, that's, that's what they're looking for. And e-commerce platforms are in a really unique position in the tech stack because e-commerce platforms are an orchestration layer in and of themselves. If you remember what I talked about, you know, like a few moments ago, it is that all of there's all of the silos and fragmented data and processes that are kind of behind the scenes for these companies. Well, e-commerce forces you to get your inventory and your pricing and your customer data and your images and your customer service and your sales. personas and rules and functions altogether in one place. Right? Typically, those are all spread out across different places. So because e-commerce platform brings everything your business knows about itself together with everything your buyers know about themselves, it's the only place that typically lives in a way that's accessible to both sides. That means that it is the natural place for a lot of LLMs and AI to come in and add value because the commerce platform has the context, has the context for how you run your business. It has the context for how your buyers interact with your business. So therefore, without having to sort of change anything, we can add AI power into that to speed up the existing processes that are already there. speaker-0: You know, you spent years at BigCommerce before, you know, AuraCommerce. ⁓ I'm curious to know if there were things you learned there that sort of changed how you think about building for B2B. Are there things that are different about ⁓ how manufacturers ⁓ and distributors operate that consumer facing companies don't really understand? speaker-1: ⁓ there's a, so there's a ton. I would say I've been in the business of building e-commerce platforms for B2B companies for well over a decade actually, even prior to coming to big commerce. What I've learned from, I would say a very long time in technology working in lots of different B2B arenas, healthcare insurance, as well as manufacturing and distribution is that ⁓ B2B companies are tend to be, you have to remember that everybody who's buying from you is buying from you because it's their job. That's the big difference. So what they want out of an experience is different than what a consumer wants when they're buying socks. And so one of the things that gets said a lot in this space is B2B buyers want a B2C like experience, meaning that if I am buying ⁓ forklift repair parts for my system, my network of say 80 garages that service industrial equipment and I need to buy and restock basically consumables and MRO equipment for them. I want my buying experience to be just like when I log into Amazon or when I go to a Shopify site and buy socks or a t-shirt or something like that. And the actual truth is like, no, that's actually not right. And I think trying to that's because when people are buying for their job, they want more friction. actually, want someone making sure that that's the right product. They want somebody making sure that it's the right quantity that is actually going to arrive on time. Maybe they need to get an approval because maybe they can buy up to $10,000, but they need their supervisor to come in to buy $20,000 worth of parts. Those aren't things that we deal with and we're purchasing for ourselves unless we have a very, very managerial spouse. Maybe that approval workflow works that way. Generally speaking, what I've learned is that what the buyer wants, they will actually tell you what they want. And what they want is reliability and trust. If you can nail reliability and trust in the digital interactions, you will have that buyer forever. But you have to establish those two things first. And that's much less true in consumer e-commerce, where I click on a Facebook or Meta ad. that I like and I just give somebody my money that I've never talked to before because why is 20 bucks? You know, who cares? Not, it's not who cares when I'm ordering the forklift parts. speaker-0: I think that's interesting. when it comes to AI specifically, what works in B2C that just doesn't translate to B2B? speaker-1: And that's a really interesting question. ⁓ think a lot of the there's two answers to that because one of them is what doesn't work because the work hasn't been done yet. And I'll start with that. A lot of what works very well in B2C that's AI powered is product recommendations, search and merchandising. Meaning that like I go in and I search for a product and I'll get like people who bought this also bought this other thing or related products and things like that. Of course, our platform supports all of that. But what is typically a large gap is knowing that this part over here is often bought by people who buy this other part. The relationship of how product data is cross sold and merchandise, frankly, a lot of times, especially in distribution, your product data is often a schematic and two lines of text and a price. That's it. So there's nothing really to go off of. No one's gone through and cleaned up the data. to say that if you buy from this category, you probably want to buy from this category or this product relates to this product. Some distributors and manufacturers have this down and have done a really good job, but a startling number of them have not. so therefore it's hard for a consumer grade search and merchandising tool to come over because it's expecting lots of rich data about every product in my SKU. Well, if I'm selling 300,000 parts and I have very little information about them written down anywhere, what is the AI going to do? Right. That's less true in consumer world where a lot of this stuff is much more richly described. And so therefore an LLM or an algorithm has more to go on. But if you don't have a lot to go on, there's not much you can do. The other side of like what is B2C that doesn't work in B2B really is, I think, to some extent, the sort of the conversational stuff that we're hearing so much about through Google, through OpenAI and Shopify to some extent. I go to chat, GPT, and I buy something. I go to Gemini and I buy something. B2B land, I don't think it's going to be there for very long time because those protocols for buying inside of an LLM are not aware of who I am as a person and what my negotiated pricing is, what my ship to addresses are, right? There's a lot of e-procurement work that needs to be done to glue those things together. And because B2B buying is serious business with so many approvals and friction and things, the idea that I'm going to sort of impulsively log into it, a chat window and just sort of like order $10,000 worth of products, I think is unlikely. There a of people who disagree on this, who disagree with me on this, but I do think we're a long ways off from that coming true. speaker-0: So back of that and maybe my final question today. ⁓ There's a lot of hype right now that I've seen about what is called agentic commerce, which is fitting into the old ⁓ trend about agentic AI. Would you say that is real for B2B or are we still years out? ⁓ speaker-1: Well, that's a bit of what I was just describing is I think we're a ways out. I think that the timeline is being dramatically compressed in the marketing is what I would say. Like a lot of tech hype, right? I think the vision is compelling and especially in kind of an e-procurement system where a lot of your buyers are buying from you in their own software. And so they're not even coming to a portal, they're not coming to a website. Right? Well, why couldn't that software be? It could, but you're talking about disrupting an entrenched group of existing, very workflow and permission driven systems out there like, ⁓ Koopa and Ariba and others that solve the specific problem. they're, they're, like I said, there's an agent at commerce has an auditability problem. Anytime the LLM is just like spitting stuff out, it can't reliably do the same thing twice. And it also can't tell you how it arrived at that. And the auditability and governance of how products get purchased in the whole transaction chain. It's incredibly important to these businesses to be able to show their work when needed. And I personally don't think the agentic commerce and I will say there are no for the other thing I would say Sam is no one can agree on what the heck agentic commerce even means. Does it mean just product discovery where I am but I'm transacting offsite? This is open AI, of course, is kind of like reluctantly forced into this position. Gemini, Google Gemini, on the other hand, is trying really hard to get you to stay squarely inside the Google box at all times. But this is a sort of a classic attention capture kind of challenge that we have seen through social and conversational commerce for decades now. And I don't think it's going away, but I also don't think agentic commerce is necessarily coming anytime soon to the B2B world. speaker-0: I know I said that was my last question, this is probably something our viewers would be interested in. Where do you think B2B companies should be putting their AI budgets? I'm sort of putting you on the spot, but where do you think they should be putting their AI budgets? speaker-1: That's a really great question. I think you've got to get your data right first. And this is a conversation that is, you know, I'm seeing and hearing play out across a lot of touch points in the industry is garbage in garbage out and computer programming, right? And so you have to have good solid data and a system for governing that data because you don't just get your data right once and then you walk away from that. And then next six months later, you're like, ⁓ my God, my data is terrible again. What happened? Well, because it's always changing. You're always adding orders, adding products, removing product lines, a lot of merger and acquisition activity happening. There's so much stuff that goes on that you need to have a program for governing data before you can reliably put AI into your business. Otherwise, AI is going to be hallucinating to fill in gaps in your data. How comfortable are you with your customer facing tools hallucinating things? Not very usually. so fix the data is the first place to investment and have the sort of governance and oversight. I would say second, would say look for ways to automate your sort of high volume, low judgment work. So if something is a scale problem for you, you need to know your own business. Like a McKinsey consultant would, you need to have timed and measured and understand where the bottlenecks like if I, if I can automate my, ⁓ my distribution centers to where the picking and packing takes place by robots and it's super awesome. But the process of getting trucks loaded up to the docs. is the same as it's always been. I'm going to maybe have boxes spilling out into the hallway, but my loading dock is still a bottleneck for me and I haven't fixed that. What have I done? I haven't actually shipped anything. I have not made any more money from all of the automation that I have employed. So I think looking at your business holistically and look at things to automate high volume, low judgment work. So order processing, catalog matching, basic customer inquiry stuff like It's not glamorous, but you're going to get ROI from that. And third, I would say take your AI and put it inside the workflow and not like adjacent to it. So again, this goes back to the, it's challenging to buy an external third party AI magic, magic thing, and sort of like drop it on top of your existing operation. The investment that pays off is AI that's embedded in the system that your buyers are using and your reps are using. So if you're making somebody log into something new to go use AI, they won't. speaker-0: Great. Thank you so much, Aaron Sheehan, VP of Strategy and Partnerships at Auro Inc. Appreciate you for joining Machine Dreams and for the reality check on what AI in B2B actually looks like. It was great having you. speaker-1: Thanks for the time, I enjoyed it. speaker-0: And thanks to everyone listening. If this was valuable, share it and subscribe. We'll see you next time on Machinedrons. you