重塑万亿级采购市场：AI Agent 如何颠覆企业采购与端到端协同

Vlad: If you want to build an airplane, you need to attract thousands of suppliers. Someone sends a confirmation: "Hey, sorry, this part will arrive 2 weeks late." And if they miss this letter, it's hundreds of millions in losses. Procurement has historically been more siloed, and now it involves lawyers, finance, a bunch of different software systems, and people. The opportunity for an AI startup is to say, "We're going to take control of this entire process from start to finish." No company or enterprise starts with fully autonomous negotiation agents from day one. Why? Because they don't trust us and they don't trust the technology from day one. Thanks to the human-in-the-loop approach, we infuse our agent with feedback and knowledge, and over time they start to trust us in 10,000, 20,000, 100,000 negotiations. When you think about what a sustainable vertical AI company looks like, what qualities do you look for? It is very, very difficult to predict your protective barrier in the future. If you look at all the best companies, in the early stages they were...

Elina: Welcome back to the a16z podcast. I'm Elina Berglund, and today I'm joined by Sima Embl, a partner at a16z, and Vlad Kyle, co-founder and CEO of Leo, a company that builds AI agents for corporate procurement. Simo, you recently wrote an article called "The Incumbents Are Coming," where you posed a question that almost every AI company faces: if a well-known software vendor already has customers and data, and the model works in its tools, what is the advantage of a startup? We have Vlad, and he can really help us understand what the startup advantage is. So, I think it's worth starting with the arguments "for" and "against" the market leaders. So, if a company can connect a powerful AI agent to its software, where is the place for another application and what is its advantage?

Seema: So, let me step back a little and outline the situation. Historically, we thought there was a market leader and there was a startup, and they were fighting—it was a fight between distribution and innovation, to quote my partner Alex Rempell. Um, but now there's this third element that you're pointing out, which is that the leading company can add one of the models on top, and then a much more serious competitor appears on the market. So why do you need an AI startup when you have CloudForce, which takes the cloud and Salesforce, combines them, and, you know, you already have all the data, and your employees are already used to using this product. So why another product? Um, I'm still absolutely convinced that there's obviously a case for an AI startup, and they're really focused on the fact that a legacy company is limited by their accounting system and the data they have, and they're not doing the job from start to finish. Let me explain this with a more specific example. Um, let's say you're a customer and you call and say you were charged after you canceled the service. This, um, resolving the cancellation history— it's not just, you know, a customer comes into the chat and says, "Hey, I was overcharged." And the answer to this: you need to contact billing, view the entire chat history, and review the contract. This is not one accounting system, it is knowledge about the customer and everything they interact with, and this is something that one accounting system will not cover. But the opportunity for an AI startup is to say, "We're going to take over this whole end-to-end process." This could be a legal area, for example, covering the entire process from preparation of materials to trial. Well, I can talk more about procurement, but it's really the concept of owning the entire workflow from start to finish.

Elina: Yes. Vlad, do you want to tell me where exactly this manifests itself in procurement? Where an existing leader plus a model is simply not enough, and what have you seen in the companies you work with?

Vlad: Well, when we think about procurement, I guess most people think about prices, right? So, what is the final price we agreed on? And, strangely enough, the recording always looks very, very simple and easy. It's just 8 thousand. This is the result, for example. And I mean, it's all the same in sales, right? So, even if I go to Sima and tell her, “Hey, we’re now collaborating with another company. Look at this signature on the contract." It looks very simple, but Seema doesn't see all the work behind it, does she? So, there were maybe 30 stakeholder meetings, 500 emails, 20 Excel spreadsheets, and the same goes for procurement from the counterparty, right? So you see 8 thousand for aluminum in your ERP system, but you don't see that perhaps the supplier has made a counterclaim and asked for 10 thousand. You don't see that the estimating engineer spent three weeks working with Excel spreadsheets and 3D modeling to figure out part prices and everything else. This is what we see. So most of the work in procurement actually happens outside of this ERP or any accounting system.

Elina: Yes. Yes. And I'm sure that in the workflow you described, an agent can do a huge number of things, and different types of agents can also perform many tasks. And this is actually a good bridge to the next question: a year ago, I think, many market players were releasing chatbots, and that was pretty much all you could see. But, Simo, in this article you wrote, you highlight four different types of agents. The search agents we are used to are process agents, policy agents, and principal agents. So can you tell us about each of them and explain what exactly changes in the judgments required for each of them to work?

Seema: Yes. A year ago, I created this meme about the "build a chatbot" strategy, which was that all large companies essentially had a chatbot on top of their accounting system that you could talk to to get information or do some light analytics. And this really fit into that first category —search agents. And maybe let me give an example of each of them— search agent, process, policy, and principal—using the example of customer service, just because it's easier to understand. So, imagine you're a customer, there's a service outage, and you call to say, "Hey, I want to get compensated for this outage." The information retrieval assistant that an incumbent supplier might have will be able to find, "Yeah, you know, this is what was said in the contract terms, and yeah, there was a failure." And just confirm this information. It pulls customer data from a database and simply shares it, perhaps summarizing it. The second agent, or the second step in the agent sequence, is the process agent. So, this process agent can provide credit approval and say, "Okay, according to our policy manual, the service was unavailable on these dates." "That's why we say you're entitled to compensation." And he can just apply it to the account within the due process. No judgments are needed here. Then, going to the policy agent, it doesn't just apply the process, it says, "Okay, in this situation, the failure lasted 20 minutes." "That's enough to consider it a significant failure." And they apply this judgment because there is actually no clear definition of the term. And in the final stage, where the lead agent works, you're actually weighing up, "Okay, should we offer more compensation, because this was a pretty terrible failure and we want to preserve the relationship, so it's worth doing more than the processes or the policy require." The importance of these four stages lies in the fact that most existing companies, if they started at all, started in the first category. Now they at least say they are moving towards processes and policies, which means they will be able to apply more judgment. And if you look at what they've launched, these are workflow agents that will help you sign a document or enter information from a transcription, or something like that.

Seema: They are still very limited to information retrieval and partly processes. They haven't yet reached more complex judgments, and I can explain why. There are many incentives that prevent them from doing so. But these companies are trying, and I think what they can't do on their own is making some of them think, "Okay, let me partner with OpenAI, Anthropic, or one of the labs to try to leverage their modeling capabilities and complement what they have and make it significantly stronger."

Elina: Yes. Or charge, I would say. Yes. Well, do you want to tell me why some incumbents are holding back?

Seema: Yes. Well, I would say that they are not so much restraining themselves as being restrained. Of course. Um, yes, I'm sure they want to go full force, but there are two sides to this. First, they have a distribution advantage, right? They have the trust of customers, which allows them to sell more products to those customers. Here, let's take Salesforce. When Agent Force launched, it was very easy for customers to say, "Yes, I'll sign up for a Salesforce agent." "Especially if it was offered at almost no additional cost." Um, and because they have that trust and distribution, getting the product connected is usually pretty easy. On the other hand, there are problems with intrinsic motivation, right? That is, if you move to more complex agents, there is an internal conflict between the existing product offering the workflow and you, who directly performs the work. These are two different products: for example, when you solve a customer support problem from start to finish, versus providing a workflow for a support agent, a human. These are different buyers. How do you, well, you know, these two teams conflict, and besides, from a sales perspective, what exactly are you offering the client? And this is often a classic problem of incumbent market leaders, right? This is when there are two vice presidents, right? They are from different departments, sell different products, and they can never agree on what incentives are right and who to sell to. But, be that as it may, there are all these classic problems that leading companies face.

Elina: Logically. Vlad, where on this spectrum— search agent, process agent, policy agent, executive agent— is Leo?

Vlad: Yes, we cover, I would say, all of those categories, and it really depends on the complexity and the risk that our agents are taking, right? So, sometimes Leo and I already have cases where we work completely autonomously. Sometimes a person remains in the control loop. It really depends on budget approval, complexity, and risk level. But perhaps, going back to what Sima said about trust in leading companies, you were talking about internal trust. I think there's also the issue of external trust, right? So how do you convince someone to go from "it's just an agent that pulls information and maybe runs processes" to "doing something completely autonomously"? Obviously, there is a technological component, but the main thing is the human factor, you need to trust them. Um, and we, as a startup or a company at the scaling stage, have to earn this trust. Market leaders already have trust, but that also means they can lose it if they release something too early and the product doesn't work, is of poor quality, or makes the wrong decisions. So, this is what we can do. And what's interesting, or rather what really surprised us, is that these process agents have become a kind of side task for us when we talk about the invoice processing process or invoice agents as a process. So, we get some information, compare it with other documents, and then transfer it back to SAP or Oracle. This is a very clear, understandable process. And then we realized: okay, this is 100% of the software market, right? This is accounting software. That's how it's created today, but in reality, it's only 20% of the total work that needs to be done, or of the problem itself. Because 80% of the problem is, for example, what to do if the account is fraudulent? What to do if there is a disagreement? What if we deviate from the standard scenario? And so we very quickly moved on to the next stage— agents that handle such exceptions. We convinced clients because we were probably lucky enough to always be a little ahead of the curve to offer them next- generation agents. For example, when we started 3 years ago, it was just pulling out a document. Today it's not impressive at all, but 3 years ago it was incredible. So we offered this to customers. We figured out, "Okay, this is a real problem." This is a real use case that they are willing to pay for. And then we were able to release it a few weeks later. We are doing the same now for the next step: for process agents, for fully autonomous agents, and for long-acting agents. We present it to them, find out what the problem is, and then we can implement it very quickly.

Seema: I think the interesting point about trust is this internal versus external trust. Another way to look at it is— yes, you need your client to buy procurement software and trust it to use in their internal processes. But one of the really interesting things when we first met Vlad was that they are also in the business of negotiating. So, you have to trust that the Leo agent will then interact with the third party, and that's where that element of trust comes in. And of course, I think a lot of people feel cheated by the incumbent players because they, you know, are quite limited and haven't been able to deliver on what they've advertised in the past. So, putting that aside, I don't know, maybe Vlad, I'd like to hear a little bit about how you convinced clients to trust an AI agent to negotiate.

Elina: Yes, and in doing so, can you also paint a picture of what procurement is, who your customers are, and what legacy systems they are used to?

Vlad: Yes. So when we think about procurement, we might think just about buying, or like in the B2C segment, where you just buy something. But it's actually a very intensive process that drives the economy, isn't it? And it covers many stakeholders and many departments: legal, technical, financial, and procurement. They all have to work on these solutions, and essentially when we.. I think the reason is how we convince them, as I said before: we're always a little bit ahead of the curve. We knew the technology was coming, so we started a few weeks before ChatGPT broke ground. We've already heard about all the problems and hype in, say, the tech bubble, and we've been able to offer that to businesses. We quickly realized that this could be an interesting use case, for example, chatbots, information search agents, or document processing. We identified the problem and quickly implemented it into production. Obviously, there is a human factor of trust: we promise something and actually implement it in our work. But there is also a product perspective, where we have a lot of experience: no company or enterprise starts with fully autonomous negotiation agents from day one. Nobody does this because they don't trust us and the technology from the start. So we take a simple approach with human involvement, and it helps us tremendously. We have, say, the perfect negotiating agent who is a great procurement specialist, but we don't know exactly how a particular Fortune 10 company works. With this approach, we feed our agent feedback and knowledge, and then they start to trust us to negotiate $ 10,000, $20,000, or $100,000. Um, and you also have other, longer-term agents. So, when we're talking about multi-million dollar negotiations where complex 3D models and technical drawings need to be analyzed. Um, we intentionally always involve experts there, right? The agent may work for a few hours, and then we ask for feedback from the cost estimating engineer, after which he performs the next part of the work. Um, and so on.

Elina: Perhaps it is worth dwelling on this in more detail. How exactly, where do you involve a person during the negotiation process? You mentioned a cost reduction engineer, but if you 're negotiating a deal, is that mostly about cost data, or is there anything else where people are involved?

Vlad: Again, it depends on the level of negotiation, right? We have to distinguish between negotiations that businesses never did because of lack of resources, but now, by using agents, they can get savings they never even knew they had, right? Before agents or Leo came along, they just ignored anything under 50k, right? This could be a kind of life hack for other startups. You can just send a large company an invoice for 40 thousand. They probably won't bargain because they don't have time for it. Unless they have Leo agents, then we will negotiate against you. Otherwise, they will simply pay this bill. And, of course , what is the risk in not negotiating at all ? So what is the risk now of using a failed agent to negotiate? Almost zero, right? We may lose something in the negotiations, but it's still better than nothing. But still , we are talking about business relationships, and they are not always just about costs and money, right? You may not be spending a lot on a supplier, but you need that business relationship, right? A good example would be podcasts or marketing services. It's probably only a fraction of the cost, but you already have a clear business relationship with someone who is setting up the studio, and you don't want outsiders doing it, because they already know how the A16Z works and how you want to record everything. That's why we use a human-centered approach here, as procurement professionals care about relationships, the tone of communication, and how it works. But mostly autonomously. We also have another group of agents where we always involve a person, and it's like a multi-step negotiation process. It's also, well, it's not just about the price, right? Is this also, say, how the contract is drawn up? So we are talking about legal issues. How, for example, is the collection system designed? So we're talking about finances. Obviously, the cost structure, so we're actually talking about cost engineering. Next, we talk about commercial terms. Um, this is procurement. Um, and these are n't people from the back office. These are highly qualified specialists who have in-depth knowledge of the specific processes of a particular company and a particular industry. Um, and they're infusing these long-term Leo agents with these insights. This is another example of how procurement has historically been more limited, and now it, so to speak, touches on legal issues, finances, a bunch of different software systems, people, both narrow specialists and generalists.

Elina: Yes. Um, can we pin this to a specific client, or rather, not a client, but a specific vertical? For example, I am a drone manufacturer. I make humanoid robots or something like that. How many parts do I have to, you know, order and purchase? How many factories do I interact with? How many suppliers do I deal with? Just all these things. If you want to choose, Vlad, maybe a vertical that, you know, just manages all of this complexity with Leo, and just walk us through their experience. I think it would really help illustrate everything you're touching on.

Vlad: Again, when we, like Leo, talk about procurement, we don't mean laptops or pencils. This, we believe, is also resolved by Leo agents, but it's simple. We decided this 3 years ago. Um, we're talking about when you want to build an airplane, robots, drones, or even, like we're doing a podcast right now about the hype around AI, but even AI has to be built. Um, and construction means procurement. For example, someone needs, if you build an airplane, you have to bring in thousands of suppliers. You need to build a factory to create this plane. Um, and even very small frictions can have a crazy impact, right? So, if you are running a very large project, like building a data center or an airplane, and one specific detail is delayed by 2 weeks, it can lead to losses of hundreds of millions of dollars and postpone the entire project. Therefore, all these decisions must be coordinated. One of the steps is to determine what exactly you need, which suppliers to work with, and who is the best to get that part. But when you have everything decided, there is operational work “ behind the scenes”, which may seem boring and unnecessary, but operational work is when someone sends a confirmation: “ Sorry, this part will arrive 2 weeks later”. And this is just one of the 500 emails in Outlook or Gmail from the purchasing manager. And if they miss this letter, hundreds of millions in losses. Done. And this happens regularly because the only thing they record in their accounting system is just the date. Right? That is, not this Wednesday, but next Wednesday. This is what you see in your system. But you don't see that maybe that's normal. Or maybe it's a loss of $100 million. Someone has to solve this , and that's what our agents do, right? They don't just receive information , they make decisions: does it affect anything? What exactly is the impact? And how can we solve this?

Elina: Yes. Ahem. When you are so deeply integrated into the physical world, what problems in the physical world can you influence? Some things, in my opinion, are simply impossible to solve: let's say a shipment arrives, some goods fall off the ship, or a street is blocked, or something else. There are things you can't influence, but there are also things you can. So where can you intervene and where will it really make a difference?

Vlad: Yes, but it's really all a matter of probability, isn't it? Obviously, you can't change the situation if the cargo on the ship has been damaged. As with every example you gave, you can't change it, but if you have all the context, you can predict it. Mm. Because you can predict how reliable the supplier is. Of course? So, there are ways to protect the goods you ship. And if you have the full context, you have one supplier that is 20% out of stock and another that is only 1% out of stock. And maybe the one with the 20% shortage is 10 times cheaper, but in this case you're better off paying ten times more because you have a higher chance that the product will actually arrive. This is the most powerful thing because you have context beyond just one enterprise. We've talked about many stakeholders, but there's also the context of the outside world, right? So the agent must have the context of all the news that appears. Maybe even have a context for betting on Polymarket, like: "Okay, these crashes will happen." Then information from the supplier side, from the seller side, from the demand side. And combining all these contexts, I wouldn't say that there are any limitations in the long run. Of course, today we have different levels of probability, but we can help throughout the process, and that's what we're building at Leo. This is much bigger than just internal procurement. It's more about the interaction between companies, about how businesses do business with each other. That is, both from the buyer's side and from the supplier's side.

Elina: You described Leo as a multi-agent system. So can you describe what the different agents do?

Vlad: At one level, Leo agents cover all four categories that Sima mentioned in her article. And this again depends on the risk and complexity. So we use them all. That is, several agents. But the other thing is that to get the job done from start to finish, these agents have to exchange information with each other. They have to do this in a very specific order. And when we talk about a multi-agent system, that's essentially what we're doing. We solve this problem from beginning to end. And since the human level of task execution involves eight people, eight stakeholders, three departments, and five different software tools, we need to cover all of that to get the job done completely. And these agents must then communicate with each other. Only with the help of a multi-agent system can the work be completed from start to finish. When we started, we obviously started with a search engine that was more like a co- pilot, about 3 years ago. But the next step was one agent. However, we quickly realized that you can't have a successful negotiation without a contract agent or an agent who monitors the news and all the things I described earlier. So, this is what we define as a multi-agent system. So if you have a bolt, for example, an airline needs to procure a bolt, let's say Boeing needs to procure a bolt. What exactly is the process for purchasing this bolt? And where exactly does Leo appear in this process ? Yes. This is one of those purchases that we can make completely autonomously. And we can do this thanks to this multi-agent system. So, first of all, someone has a need, right? They need to be informed about this somehow. And even this part is extremely difficult. You need to call someone. Maybe you open your laptop because you're a construction worker, for example. You only open your laptop once every two weeks, and now you are required to work in SAP or another ERP system. So you can't even apply. Here's how we do it very simply. You can take a photo, upload a code, or an Excel file—that's all you need to know about procurement. No one outside the purchasing department cares about categories, general ledger accounts, or framework agreements, no one is interested. We in the procurement world are concerned about this, but no one else is. And then our agents take action and check the warehouse stocks. They are clarifying the situation by asking another factory. Can we get these bolts internally? No, okay, then I'll contact the purchasing agent to see if we have any internal suppliers. Do we have external suppliers? Another agent should then prepare a request for quotation (RFQ) and send it via email. After that, a bunch of letters arrive. Some of them are complete nonsense. Some are simply written in the letter itself. Some of them are PDF files, others are Excel spreadsheets. We receive this information. Then we move on to the next step. Perhaps, based on our price comparison, there is an opportunity for negotiation. And then we have agents who essentially decide what to do next. Negotiation can mean strategic negotiations involving a person. This could mean autonomous negotiations. These can be auctions and electronic bidding. Calling a specific agent for negotiations. And then the entire process from A to Z: order confirmation, shipment tracking, invoices. And we are able to do this completely autonomously, taking into account the entire context. And, of course , the next important step is to do this for more complex details when we talk about direct procurement, where we also work.

Elina: Yes. What is direct procurement?

Vlad: So, look, everything I just described is about automation, right? So, you can run this process completely autonomously and get even more savings as a result, right? We look at it this way: okay, what does this task look like from A to Z and what exactly needs to be done? What do they do 1,000 times a day but would actually like to do zero times? We've launched fully autonomous agents, but there are also opportunities to do things that they don't do at all right now, but if a business did that 1,000 times a day, it would have a huge impact on the bottom line. Autonomous negotiations on costs that have never been discussed before. This applies to indirect purchases— for example, maintenance, building a factory ( as in the Bold example), as well as laptops, pencils, marketing services, or setting up this podcast studio. These are all indirect costs, and we also have direct components. It's like building an airplane: there are all these suppliers whose parts you actually need to build an airplane, a drone, or a robot. And here we are not talking about 50,000 suppliers, but about 100 or a maximum of 2,000 suppliers. Um, and they are extremely strategically important, and you have maybe one supplier with a billion in spending. Therefore, you don't want to conduct autonomous negotiations. You want to have negotiations that last 3 months, where you are incredibly prepared and where your engineers analyze: okay, what is the situation in the aluminum industry? What is the situation in the oil industry? Hmm, how did the price change? Um, so you're actually checking all these drawings, you're quality controlling this part. And this is where it gets really exciting to implement agents.

Elina: Yes, and for something like that, you would probably have engineering experts and other procurement professionals as the front office, and the agent would be more of a back office. Is this an idea? Is the agent: "Actually, you're sitting at a table, shaking hands, and it's a robot instead of a human negotiating?" "Um, is it.. So, what is it? Is it more back office or still front office?"

Vlad: This, this, this is obviously more back office now, because you, you, you have these complex multi-million dollar negotiations. And this, this is again a great example. 90% of the work is preparation. The end result that you see in your accounting system is, "Oh, instead of $1 billion, I paid $900 million ." "Although it took 3 months of preparation and 10 people working on it full-time." Um, and obviously this happens in, um, the back office. But we actually have a few use cases where, um, it helps in real time as well. So, think about, um, let's say we're negotiating right now and I have perfect preparation. Just like with those, um, those notes that we have here. Um, imagine, while we're negotiating, I have real-time insights popping up on my screen where you tell me that the oil index has moved 10%. So, it grew by 10%. So that's why we need to raise prices by 10%. And I would immediately get a pop-up message: "It's true." "Oil has increased by 10%, but the product contains only 30% oil." So you, you, you shouldn't raise the price by 10%, maybe only by 4%. Um, so, yes. These are truly exciting use cases, even in real life.

Elina: Yes, we know that companies like Harvey and Decagon are currently actively retraining models. Um, what exactly are the base models you use and how do you approach things like training or tuning?

Vlad: Yes. We think it's possible...we use a lot of models from all the vendors, and we really see it as a commodity, right? So they have a good general business purpose, like reading, creating PDFs, and then creating Excel spreadsheets and all that. But we also believe that for some use cases, you can achieve a tremendous amount by combining the base model with certain tools, and perhaps get the job done 100%. But there are also cases where you can have the best base model and the best toolkit, whatever that means, but you'll still only achieve 80%. And, I would say, a good example of this is what cost- cutting engineers do during negotiations, right? They analyze the drawings. And then they determine how much this part should actually cost. That's why it's called proper cost modeling. And here, of course, there is an opportunity where we are already thinking and have started to train the model to reach 100% in this part. Another example is price comparison, where imagine you have a price quote, and in a perfect world you would just drag it somewhere and get the perfect price. But all this information is not publicly available, is it? This is all proprietary data based in one enterprise or across multiple enterprises. Therefore, general-purpose models cannot learn from this. Um, so we're, you know, thinking about not just teaching LLM; I think we're seeing models like Jeff emerging now, where you train them on text data, but the output is actually some kind of summary or just an ideal price. Um, and you can't do it with standard tools, because if you give me a quote from BCG and from McKinsey, they can do the same job, but the price can be ten times different. And I will have no idea which is better. But if you give it to a purchasing manager, he will immediately have a gut feeling. Like, okay, this suggestion makes sense. Um, I think good examples are again content creation. The question always comes up: I don't know how much I should pay someone to make a video. Um, but there's an intuition behind it, if I ask another video maker how to do it. But if you ask them to write down the rules, they won't be able to do it because it's just a feeling and instinct. So this is where we see a lot of opportunity, actually, in training an agent, but maybe not a classic LLM, but rather exactly what companies like Jeff's models are doing now, focusing on the outcome.

Seema: So, we have price benchmarking and short-run cost modeling.

Vlad: Yes.

Elina: Um, Sima, we talked a little bit about how labs are moving into industry work or partnering with market leaders to do that. Um, when you think about what a sustainable vertical AI company looks like, what qualities do you look for?

Seema: First, it's ownership of the entire workflow that we're talking about, creating this data asset and being able to do things that in many cases no one has done before. Given this, I think we talk a lot about "moats" ( competitive advantages ). It is very, very difficult to predict your " moat" for the future. If you look at all the best companies, in the early stages they just thought, “Okay, I’m gaining customer trust, I’m selling to them more, and there’s a lot of opportunity here,” rather than “ Okay, I’ll do these six steps, then I’ll move on to step seven, and then we’ll have a competitive advantage .” So I think we talk a lot about security and durability, and part of that is customer engagement, more dependencies arise, they appreciate it, and you do most of the work. And here it really is: you know, the old CRM company was a log for all deals, the new AI sales agent actually takes over a significant part of the sales preparation process, outbound processes, processing incoming requests and doing a lot of work. Um, the company, the customer in general depends on this product, and this is a very important signal for creating a "moat" around the business. And everything we talk about in terms of customer acquisition, network effects, etc., is kind of a consequence of that initial customer usage and the value of the product.

Elina: Um, Vlad, have you had any conversations with customers or potential customers who asked you, “Why should I buy your product?” “Why can’t I just plug into the model and do it myself?” Or use any existing accounting system I have, plus a model. What do you tell them and how do you convince them to use Leo?

Vlad: Yes, one hundred percent, and that's a fair question, right? Even if you look at Leo from the inside, the first case three years ago that went kind of viral in the procurement world was simply about getting a price quote and entering that information into SAP. It was technically simple, but it brought enormous business value. So you have this information retrieval agent that collects all the data and feeds it into SAP. This was our first product. And we had a team of engineers, obviously small, just the three of us and maybe four people who built it and then sold it. But today it's a case of applying for a job at Leo. We invite people to create this, and they have about 8 hours to do it. So what I want to say is that a product that was one of our first use cases can now be built by engineers in 8 hours. Because it's so easy to create things now. So the logical question is: if someone can build this in 8 hours...Okay, cool, but couldn't the procurement departments just build the whole thing in 2 months? And the answer is: yes, you can build it in 8 hours, but you will only achieve, say, 70% productivity. And the problem is that 70% productivity or accuracy, or however you measure it—it depends on the task—doesn't mean 70% automation. So, this might mean that you have 70% productivity, but you still need to do 100% of the work, because 70% is not that much. So again, all people have to verify the data. So you might even have created more work than before.

Seema: Two things worth adding to what Vlad said. First, it is generally good that basic models or GPT-type products are gaining more popularity, because it means that people will also trust highly specialized products. Therefore, I believe that increased awareness, comfort, and excitement about AI tools is generally good for the market. The second thing: I was talking to a management team at a Fortune 500 company two or three weeks ago, and they mentioned that they were trying to build their own cash-raising product as a large corporate business. And after, I don't know, at least three or four months of work, they discovered that there wasn't enough context; the context was of low quality. They had a lot of recordings, a lot of screenshots. They tried to put it all together into one system, but the results weren't good enough. And then a huge question arose: we now have two different ERP systems and we were going to buy another one, so who is going to update all the mappings and check if everything is working? And we keep talking about exception handling, you now have to map that to a completely different system, a different way of doing things, and I think they quickly realized that in-house development didn't make sense. So we hear stories like this all the time where people say, "Okay, I'll do the development in-house, " and then, " Wait, this is no different than what corporations have always tried to do on their own in the past." But, um, I think corporations in general are realizing that there's their core competency, and then there's the development of internal tools, and they should focus on the former.

Vlad: Yes. Yes, that makes sense. Ahem. Yes. That's exactly what I meant: Sure, they can reach 80%, but that last 20% really matters. Um, and they matter in order to go into production. That's why you need this base, right? You need all these integrations, memory, workflows, and sometimes industry data. Um, to do that. And, as the Pareto principle says, those 20% can require 80% of the effort. Um, or they take 80% of the effort. Um, so to all those Fortune 500 companies: you can do it. But then, say, procurement processes or AI procurement agents should become one of your core competencies. And you should evaluate whether it makes sense for you to make this one of your key skills.

Elina: Vlad, I'm curious if you see suppliers starting to use agents or AI, and what do you think will happen when both buyers and suppliers are fully equipped with AI?

Vlad: Mhm. Yes. We are 100% sure that in the future there will be agents on both sides. Um, and that's quite logical. But, strangely enough, we see that when we look at the supplier side, it's essentially the seller side, right? We see that the seller side has always been ahead of the purchasing side. But now , as we look at the suppliers of these Fortune 500 companies, we see that this is no longer the case. They may have advanced, say, in video recording or using tools like Granola, but they don't have agents that automate the work. And the cool thing about shopping is that it's "unsexy," right? Sales is attractive, procurement is not, but this is the only process where procurement is the opposing party. But now the good thing is that in these industrial companies, Fortune 500 companies, procurement has more power over the supplier. Because you, as a typical example, for example, an automotive component supplier, dictate to your suppliers what they should use, what the quality should be, how they should respond to a specific request for quotation (RFQ). So now there is an opportunity if we serve the purchasing department, they can dictate what the supplier should use. Why don't we also promote agents to them that help automate the work, covering both sides of the transaction?

Seema: Olena, I know we talked earlier today about how you can have two sides on one platform? How does it even work? I think this could even extend to legal work, right? Even if these are two very conflicting sides , right? But if you have two law firms with clients who have different interests, but both benefit from knowing: here's the latest project, here's where we are with open issues, here are the things that were agreed upon, and just keep track of that. This doesn't really exist now, right? This is all created by people, and this coordination work could be done by an agent.

Elina: Yes. Yes. It's great to think about how both sides are perhaps developing in parallel. One side may move a little faster, as you said, Vlad, but over time, maybe people will just be on the same platform, and it will be much better coordinated for everyone.

Vlad: 100%, because, as we mentioned at the beginning, right? So the obvious question is: Okay, we also talked a little bit about negotiation agents. What if both sides had negotiating agents? And the price is 100% the moment where it's a zero-sum game or where they have different interests. But we also discussed that the price is the result of 5,000 other tasks that have taken place. And for these 5,000 other tasks, they have the same incentives. Sales wants to have as little friction as possible. Buyers want to get to market quickly, right? Again , going back to building airplanes, building data centers. You want this data center built as quickly as possible. You don't want it to be built 6 months late just because it takes so long to analyze all the responses from suppliers. Um, and you want to speed up sales as well. So, just like for all those 500 or 5,000 other tasks , the incentive is exactly the same. And so you can deploy, or Leo can deploy agents on both sides, automating all these other tasks. Hmm, that's the beauty of it.

Elina: I think there used to be a logic that you shouldn't tailor software too much to one end user or client. But I think that with big language models and AI in general , it might become easier to customize a product without slowing down your business. So I'm wondering if you observe anything similar, Vlad or Simo? And what does this actually mean for end software buyers?

Seema: I think, overall, a lot of work is currently being done to implement solutions on the ground. Part of this is because the state of customer data and understanding customer N is much more complex than understanding N plus 1. So we're in the early stages of implementation, and that's why there are still a lot of people involved in this product. Incidentally, this is harder for market leaders to implement because they are not set up for such a feedback loop ; they have implementation teams , but it's more of a secondary process, not something that fuels product development. The beauty of AI is that it learns over time —what we call learning cycles—and if you have the right evaluation process in place, you can do increasingly complex tasks, and part of that is automating the deployment itself. So you can.. and I'd be interested to hear how Vlad does it, but a lot of our companies are doing it pretty quickly, where customization: (A) is handled automatically, and (B) the customer can configure the parameters themselves through the software, unlike how it used to be when you brought in Accenture to customize SAP, and now you have an implementation team that helps with the architecture, but ultimately everything is driven by software.

Vlad: Yes. Well, that's the reason why, when you look at the organizational structure of Leo, about 85% of the people are engineers. And even if you look at those who don't have an engineering position, they mostly have an engineering background , and the reason is that we obviously don't want to be a consulting company, right? So we make sure we have the best agents in indirect, direct procurement and finance, and in those areas. But then, as you mentioned, there's a lot of on-site deployment work if you're working with large companies, because they have different nuances in their processes. But we work as you mentioned: we build the product in a way that reduces the need for customization, and also provides a high level of self-service. So the work of our FDEs and implementation engineers is , on the one hand, to create opportunities for self-service, and on the other hand, to automate our own work. So their KPI is for you to see this happen many times. Your job is literally to automate yourself. And then, once you've automated yourself, you move on to the next task. I think this is the same approach that Google used. So, yes, I agree with you 100%. And that's exactly what we build and how we do it.

Elina: There was a very big event recently related to accounting systems, and we're not talking about Dreamforce. We're talking about the Bots and Buyers summit that Leo hosted in New York. I would like to hear real stories from the scene and find out what you see among buyers. What are people passionate about? What do people expect in the future? What do people ask you for? Can you tell us any stories from that day and that event?

Vlad: Well, this was the third time we held this event. We held it in New York, just a few blocks from our office, and over 100 procurement leaders came. And we made sure that when we hold such events, we only invite high-class professionals, right? So, senior managers , purchasing directors, vice presidents. And there are two things that are very different, or why they are so impressed. The first thing is how procurement has worked over the last 26-27 years, because there have been a lot of tools, right? You can see all of these technology landscapes on LinkedIn. You will see about 500 procurement tools there. But if you talk to procurement people, it's a very painful topic. I challenge anyone to find someone who loves working with procurement. There simply are none. It's safe to say that people hate working with procurement. I'm talking about those who make requests, about suppliers, and even the people in procurement themselves hate working with this department. So what happens if there are 1000 tools? And the reason is that all these tools.. They just made the process a little more efficient. That's all. But they never changed how these people actually work. It's just crazy that they still work in email , Microsoft Teams, Excel spreadsheets, and PowerPoint. This is their main channel of work, where there is no artificial intelligence. Of course. And this is the same part. And the other thing is we give them a very broad interagency view of procurement. Right? We don't say, "Look at this cool billing feature we built." We look at how someone needs something, and at the end of the day it ends up on your desk—it could cover indirect costs, direct costs, logistics, finance, and you see how we do it. We also bring this into the physical world, because AI agents—that's very abstract. That's why we create stands, they even exist in our offices, particularly in New York, where you can walk around and try out all these agents in practice. And that's exactly what people like, the next event will be for about 700 people. So you can imagine how massive it is.

Elina: Buyers went on a rampage. Yes.

Vlad: Yes, it will happen , and that event will be held in Munich. Yes. So, we hold them, for example, in Europe, Munich and New York all the time.

Elina: Yes. Well, if you're listening to us and you're in the procurement business, you know where to turn. Maybe I have one question, one question from me. What do you think it takes to get people excited about shopping? Are these agents? Are these people? Is this a time saver? Or something else?

Seema: You mentioned this. Procurement is one of those things that I remember people not liking, it's an area with universally low customer satisfaction. I remember talking to a guy in procurement probably 7 or 8 years ago when I was studying this category. And he said, "I hate talking about this product." I use, you know, this outdated accounting system. I'm on Coupa and don't want to buy anything else. I don't want to talk about it. This is normal. This was the most dissatisfied customer I have ever spoken to out of millions of similar calls. But I'm curious what you think really makes people excited about this category.

Vlad: Yes. So, I mean, that's the reason I love procurement, it's not so much what happened, but how people react to it. Isn't that right ? If you talk to people, they are very irritated. So, this is a very emotional topic. But let's be honest, B2B—that's us, okay? But it's very emotional. So that's good. And if you combine that with something boring and highly specialized , that's also an advantage, because it's easier for us to impress these people, right? Because the last real revolution they saw was 20 years ago. Well, and maybe a better user interface 10 years ago. But nothing more happened. And so you have a boring but highly emotional field, plus a crazy impact on business. Yes? It seems unnecessary, but as I have already given examples. This obviously has a huge impact on profits and losses, and also affects the entire economy. Isn't that right? So, we're talking about how data centers are built , how airplanes are built , how cars are built, uh, how drones are built. So it's extremely important to, uh, have a fixed procurement process, uh , not just to implement or build something. Uh, but also when you talk about when we, when you look at the competitive environment. So to get a 1% margin increase, uh, you need to make 10 % more revenue, 10% more sales. Uh, so if you can get even 1% savings, that's 10% of the sales, uh, that you have to make, uh, to get the same result on your income statement. Uh , so this is extremely important. And if you combine all three of those things, you get, uh, a trillion-dollar business opportunity. This, this is my opinion. How about procurement, but there are probably other things that are emotional, boring, but have a huge impact on the business.

Elina: Well, Vlad, thank you very much for joining us. Uh , that was a lot of fun .