Corporate procurement is becoming the key battleground where AI-native startups are taking on traditional software giants.

On an October 2 podcast interview by renowned venture capital firm a16z, host Elena Burger engaged in a deep conversation with a16z partner Seema Amble and Vladimir Keil, co-founder and CEO of enterprise AI procurement agent startup Lio, covering core market concerns such as how AI application companies can compete against traditional software giants, the penetration path of AI agents in real business environments, and financial returns.

a16z partner Seema Amble stated that traditional software giants are constrained by the boundaries of their systems of record, compounded by internal team interest conflicts, making it difficult for them to truly advance to higher-order autonomous judgment AI agents—which is precisely where the core opportunity lies for AI-native startups.

She pointed out that the key to winning the endgame is "owning the end-to-end process," rather than merely embedding a chatbot into existing workflows.

Vladimir Keil argued that in enterprise procurement processes, approximately 80% of substantive work occurs outside systems of record such as ERP, and it is precisely this blind spot that makes it difficult for traditional software vendors to achieve true end-to-end automation by "layering AI models on top of existing systems."

Vladimir Keil stated that if a procurement manager misses a delivery delay notification for a single component, it could cause hundreds of millions of dollars in project losses, and Lio's multi-agent system is specifically designed for such high-risk scenarios.

The Real Business World: 80% of Work Happens Outside ERP Systems

In the highly traditional trillion-dollar procurement market, this "end-to-end" pain point is particularly evident. The market tends to focus only on the final procurement price shown in the ERP system, while ignoring the enormous friction costs behind it.

Lio CEO Vladimir Keil pointed out directly:

You'll see in the ERP system that purchasing aluminum cost $8,000, but you can't see the 30 stakeholder meetings, 500 emails, and 20 Excel spreadsheets that may have occurred behind it. The vast majority of procurement work actually happens outside this ERP or any system of record.

In large complex projects like building aircraft or data centers, minor supply chain frictions can bring catastrophic financial consequences. Vladimir stated:

If a specific part arrives two weeks late, it could cause hundreds of millions of dollars in losses and delay the entire project. And this could simply be because the procurement manager missed one confirmation email among 500 in Outlook or Gmail.

Traditional systems can only record a "arriving next Wednesday" date, but cannot judge like an AI Agent whether there are hundreds of millions of dollars in default risks lurking behind it and proactively intervene.

If a year ago the market hotspot was "chatbots" layered on top of software, the market focus has now shifted to Agents with autonomous decision-making capabilities. Seema categorized current Agents into four stages: Retrieval Agents, Process Agents, Policy Agents, and Principal Agents.

Unlocking Overlooked Profit Margins: The Unnoticed Long Tail Market Below $50,000

Beyond improving efficiency, the market's biggest concern is whether AI applications can directly improve corporate income statements. In this interview, Vladimir revealed the incremental business imagination space brought by AI Agents: handling expenditures that enterprises never negotiated due to limited manpower.

Vladimir disclosed:

Before Agents, enterprises didn't care at all about procurement below $50,000. This is also a 'hack' for other startups—if you send a large enterprise a $40,000 bill, they may not even negotiate with you, because they don't have the manpower capacity to do so.

Vladimir further pointed out:

But with Lio's Agent, we will automatically negotiate with you. For enterprises, the risk of having an even imperfect negotiation Agent is almost zero, because it's always better than not negotiating at all.

From a macro financial perspective, this accumulation of small savings will unleash astonishing leverage effects.

Vladimir used a set of striking data to explain the business value of procurement AI to investors:

In the current competitive landscape, to improve your profit margin by 1%, you need to increase revenue by 10%. So if you can achieve just 1% procurement cost reduction through AI, the result on the income statement is equivalent to the effort of increasing sales by 10%. This is extremely important.

Building Trust and Moats: Human-Machine Collaboration Is the Necessary Path

Facing business negotiations involving large amounts of money, how can enterprises dare to hand authority to AI? Vladimir stated directly:

No one will establish and use a fully autonomous negotiation Agent on day one. Because they won't trust us from day one, nor will they trust this technology.

Lio's approach is to adopt a "Human-in-the-loop" model, absorbing feedback by handling $10,000 and $20,000 negotiations, gradually winning customer trust.

Specifically, AI agents generate recommendations in each round of negotiation, which are reviewed and confirmed by humans before execution. This process continuously provides feedback and learning data to the system, allowing the agent to gradually accumulate specific enterprise process preferences and decision boundaries.

As trust accumulates, customers begin to expand the scope of autonomous negotiation authorization from thousands of dollars to tens of thousands, even hundreds of thousands of dollars in procurement.

For more strategic direct procurement, such as long-term supplier contracts worth hundreds of millions of dollars, Lio always retains professional engineer intervention nodes in the system.

Agents can integrate external data such as aluminum indices and oil price trends in real-time during multi-hour negotiation preparation, and push insights to human negotiators instantly at the negotiation table:

The other party says oil prices have risen 10%, therefore demanding a 10% price increase, but the system will immediately pop up a prompt: oil-based raw materials account for only 30% of this component, so the actual reasonable increase is approximately 4%.

Vertical AI's Moat Lies in "Owning the Work Itself"

Facing the challenge of "why not just use general models to build it yourself," a16z partner Seema Amble cited a Fortune 500 company case: the company attempted to build its own accounts receivable collection system, but after months discovered that internal data quality was insufficient and cross-ERP system mapping maintenance costs were extremely high, ultimately having to abandon the effort.

She stated:

An enterprise's core competitiveness is its business, not building internal AI tools. 70% to 80% accuracy seems good, but in real scenarios it may mean "still need manual review of 100% of data," which not only fails to reduce workload but may create more trouble.

Vladimir Keil pointed out that from a moat perspective, the key advantage of procurement AI comes from data accumulation and depth of process embedding.

As the scale of negotiations handled by agents jumps from tens of thousands to millions of dollars, the enterprise-specific knowledge and industry price benchmark data accumulated by the system will form continuously strengthening competitive barriers.

As Amble concluded, truly enduring vertical AI companies are judged not by "how many defensive steps I designed," but by "whether customers truly rely on this product to get work done"—this dependency relationship itself is the most important moat signal.

Full Podcast Transcript Below:

Vladimir Keil

If you're building an aircraft, you need to procure materials from thousands of suppliers. Suppose someone sends an email saying: "Sorry, this part will be delayed two weeks." If the procurement person misses this email, the loss could be hundreds of millions of dollars.

Seema Amble

Procurement historically may have been more like a relatively closed module, but now it extends to legal, finance, and numerous different software systems and related personnel. For AI-native startups, the opportunity is: we want to own the entire end-to-end business process.

Vladimir Keil

No enterprise will deploy a fully autonomous negotiation agent from day one. Why? Because they initially trust neither us nor this technology. It is through the "human-in-the-loop" approach that we continuously train agents with feedback and experience, and slowly, they begin to trust us to handle $10,000 negotiations, $20,000 negotiations, even $100,000 negotiations.

Elena Burger

When you think about what a vertical AI company with lasting vitality should look like, what traits do you pay attention to?

Seema Amble

It's hard to predict future moats. If you look back at the early stages of the best companies in history...

Welcome back to the a16z podcast.

Elena Burger

I'm Elena Burger, and today I've invited two guests: a16z partner Seema Amble, and Lio co-founder and CEO Vlad Keil. Lio is a company building AI agents for enterprise procurement.

Seema, you recently wrote an article titled "The Giants Are Coming," which raised a core question facing almost all AI application companies: if an established software vendor already has customers, data, and models that can run within its tool ecosystem, where is the startup's advantage?

Vlad is also here today, and he can help us better understand where startups' advantages lie. I think a good starting point is to separately examine the arguments for and against the giants. If a company can already connect capable AI agents to its existing software, what room is there for another application?

Seema Amble

Okay, let me lay some groundwork first. Our past thinking framework was: there are giants in the market and there are startups, and between them it's a battle of distribution capability versus innovation capability—borrowing my partner Alex Rampall's words.

But now a third force has emerged, which is what you pointed out: giants can layer a large model on top of their own systems, becoming even stronger competitors in the market. For example, Salesforce launched Agentforce, integrating cloud services with Salesforce, plus their existing data accumulation, and the product interface employees are already accustomed to—so why would you need another product?

I firmly believe that AI-native startups still have their value. The core reason is: traditional giants are constrained by their own systems of record—what they have is just that record, and they cannot complete end-to-end tasks.

Let me give a specific example. Suppose a customer calls and says they're still being charged after canceling their subscription. To solve this, you can't just have the customer say "I was overcharged" in a chat box and give a reply—it must also involve the billing system, the complete conversation history, and contract content. This isn't something a single system of record can cover, but rather the full knowledge and all related information surrounding that customer—things a single system of record cannot reach.

The opportunity for AI-native startups lies in declaring: we want to own the entire end-to-end process. This could be in the legal domain, such as full-process management from case filing to trial, or in procurement. The core idea is: owning the complete end-to-end work.

Elena Burger

Vlad, can you specifically explain how this manifests in the procurement scenario? Why isn't an existing giant plus a model enough? What have you observed among the enterprises you work with?

Vladimir Keil

Of course. When it comes to procurement, most people probably first think of price—what the final negotiated price is. And the recorded result always looks very simple, like just "$8,000."

This is actually the same on the sales side. Even if I go to Seema and say, "Hey, we signed a contract with another company, just look at this contract and sign it," on the surface it's very simple, but Seema can't see all the work behind it—probably 30 stakeholder meetings, 500 emails, 20 Excel spreadsheets. The procurement side is the same. What you see in the ERP system is just "aluminum: $8,000," but what you don't see is: the supplier once counter-offered $10,000; the cost engineer spent 3 weeks running countless Excel spreadsheets and 3D models to calculate the cost of the part.

So the vast majority of procurement work actually happens outside the ERP or any system of record.

Elena Burger

It's imaginable that in the workflow you're describing, various agents can complete many different types of tasks. I think this leads perfectly to the next question: a year ago, most giants were only releasing chatbots, which was basically all they could offer. But Seema, in your article, you outlined four different types of agents: Retrieval Agents, Process Agents, Policy Agents, and Principal Agents. Can you walk us through these four types and explain how they differ in the level of judgment required?

Seema Amble

Sure. About a year ago, I made a meme called "the strategy of slapping a chatbot on it"—meaning that the major giants were basically just putting a chatbot on top of their systems of record, used for retrieving information or doing some simple analysis. This corresponds exactly to the first category: Retrieval Agents.

Let me use a customer support example to illustrate what Retrieval Agents, Process Agents, Policy Agents, and Principal Agents each are.

Suppose you're a customer, your service has been interrupted, and you call saying, "I want to apply for compensation."

Retrieval Agent (what giants may already have): can look up what the contract terms are, confirm whether an interruption actually occurred, and verify relevant information. It extracts customer information from the database and presents it in summary form.

Process Agent is the second step in the agent sequence: it can automatically walk through the credit approval process based on the policy manual, determining "according to our policy, service was interrupted during this period, therefore you are entitled to compensation," and directly complete the billing processing. The entire process requires no judgment, purely process execution.

Policy Agent goes further: it doesn't just execute processes, it makes judgments. For example, "this service was interrupted for 20 minutes, which is sufficient to be classified as a major interruption event"—here it's exercising judgment, because there's no strictly defined standard.

Principal Agent is the final stage: it weighs considerations, such as "this interruption was quite severe; to maintain the customer relationship, should we offer additional compensation beyond what the process and policy require?"

These four levels matter because most giants, even if they're doing anything, are still at level one. Now at best they're making marketing claims about evolving toward Process Agents and Policy Agents, meaning having more judgment capability. But look at what they're actually shipping—those workflow agents that help you sign documents, enter information from transcriptions, and such—they're essentially still limited to retrieval and a small amount of process, without truly entering the judgment level.

Of course, there are various structural reasons behind why they can't move forward. Some giants are trying to partner with labs like OpenAI and Anthropic to supplement their capabilities with model capabilities, but overall, progress remains limited.

Elena Burger

Are you willing to talk about why some giants are holding back?

Seema Amble

Sure. I'd say they're not actively holding back but passively constrained—I believe they all want to push forward fully, but there are roughly two reasons.

First, they have distribution advantages. They have customer trust and can use that as a foundation to sell more products to customers. Take Salesforce—when Agentforce launched, customers easily accepted it, especially with almost no additional cost. So they have a trust foundation, distribution channels, and product activation is often quite smooth.

Second, they face serious internal incentive conflicts. Once they start developing more complex agents, it conflicts with existing products—existing products provide workflows, while new agents handle complete work outcomes. These are two different products, aimed at different buyers. For example, "end-to-end customer service resolution" and "workflow support for human customer service agents"—these two have different buyers. Not to mention there may be internal competition between these two teams. From a sales perspective, what exactly are you selling to customers? And it's often the classic big company disease: two VPs in different departments, each selling their own products, completely unable to reach consensus on incentives and responsibility division. These are inevitable structural dilemmas for giants.

Elena Burger

That makes a lot of sense. Vlad, on the spectrum of Retrieval Agents, Process Agents, Policy Agents, and Principal Agents, where is Lio currently?

Vladimir Keil

We actually span all of these levels, depending on risk and complexity. Some scenarios we've achieved fully autonomous operation, while others retain "human-in-the-loop"—depending on the scale of budget approval, complexity, and risk level.

Coming back to the trust issue Seema mentioned—you're talking about internal trust, but I think there's also an external trust dimension: how do you convince customers to leap from "just retrieving some information and running some processes" to "making decisions fully autonomously"? This isn't just a product-level issue, it's a people issue. They need to trust you. We as a startup/growth-stage company must earn that trust. And while giants already have a trust foundation, this also means that if they release too early and the product underperforms, trust will equally be destroyed. This is precisely where our opportunity lies.

One thing that surprised us greatly: Process Agents have almost become a "side quest" for us. Take invoice processing—retrieving information, cross-document matching, pushing results back to SAP or Oracle—this is a very clear process. We discovered that this actually covers 100% of the software market, meaning today's invoice software all does this, but it actually only completes 20% of the overall task, because the remaining 80% of the problem is: what if the invoice has issues? What if there are inconsistencies? What if it doesn't follow the ideal path?

So we quickly pivoted to the next level—exception handling. The reason we were able to convince customers, I think, is largely because we always stay just slightly ahead of the curve. We can paint a vision of next-generation agents for customers. For example, three years ago, "retrieving information from documents" was already amazing enough. We pitched this solution to customers, found that these were real problems they genuinely cared about and real use cases, and then we could get it into production within weeks. Now, we're using the same approach to advance Process Agents, fully autonomous agents, and long-running agents one by one.

Seema Amble

There's an interesting angle about trust: internal trust versus external trust. Customers need to trust procurement software and use it for internal processes—that's one layer. But when we first met Vlad, one thing was particularly impressive—they're not just handling internal processes, they're also conducting negotiations. This means you need to trust Lio agents to interact directly with third parties—that's another layer.

Of course, many people have been burned by giants' past exaggerated marketing, but setting that aside, Vlad, I'd love to hear how you convince customers to trust an AI agent to lead negotiations?

Elena Burger

Right, while discussing this, you could also describe the full picture of procurement work: who are your customers? What kind of systems did they use before?

Vladimir Keil

When it comes to procurement, people might only think of "buying things," like B2C. But actually, this is a complex process that runs the entire economic system, involving numerous stakeholders and many departments—legal, cost engineering, procurement, finance, everyone participates together in making these decisions.

As for how we win customer trust, the core is still what I said earlier: we always stay slightly ahead of the times.

We started before ChatGPT launched—in fact, we launched a few weeks before ChatGPT, so we caught the technology wave early. When the industry hype period came, we could already show enterprise customers practical, workable solutions, quickly validate use cases, and rapidly get into production.

Of course, there's also the human factor—trust is built on whether you can truly deliver on your promises. But from a product level, we designed a progressive trust-building mechanism, the core being the "human-in-the-loop" email approval process. No enterprise will deploy a fully autonomous negotiation agent on day one—because they trust neither us nor this technology. So we offer a very natural entry point: human-in-the-loop.

This is extremely valuable for us too. We may have a theoretically "perfect negotiation agent," but we don't understand how a specific Fortune 10 company actually operates. Through the "human-in-the-loop" mechanism, customer feedback and experience continuously feed into our agents, allowing them to keep learning. Slowly, they begin to trust agents to handle $10,000 negotiations, then $20,000, $100,000.

For long-cycle negotiations involving millions of dollars—such as analyzing complex 3D models and technical drawings—we always have experts in the loop. After agents run for hours, we request feedback from cost engineers, then proceed to the next step. This is intentional design.

Seema Amble

On the negotiation side, where exactly do you place human intervention? You mentioned cost engineers, but in the back-and-forth consultation process, is human intervention mainly around cost engineering-type data, or are there other aspects?

Vladimir Keil

This also depends on the level of negotiation. There's one type of negotiation that enterprises simply never had the capacity to do before—by deploying agents, they can capture savings that were previously completely ignored. For example, procurement below $50,000 was previously never negotiated because there was simply no processing capacity.

Here's a finding that might be useful for other startups: you can send a large enterprise a $40,000 invoice, and they likely won't negotiate, because they don't have a process to handle it—unless they have Lio agents, then we'll come negotiate with you. Before this, they usually just pay directly. And in these cases, the risk of not negotiating is almost zero—even if the negotiation result isn't perfect, it's always better than doing nothing.

But then again, business relationships aren't always just about cost. Some supplier relationships have value beyond the procurement amount—like podcast recording services, which may account for a tiny fraction of spending, but you need to maintain a long-term cooperative relationship with the supplier because they understand how you work and can't be easily replaced.

In these cases, we designed a human-involved approach: procurement staff focus on relationship maintenance, on the tone and manner of communication, while most of the process still runs autonomously.

Another type of negotiation always requires human-in-the-loop participation. Negotiation isn't just about price—it also involves contract term design (legal), payment term design (finance), cost structure (cost engineering), and commercial terms (procurement). These are professionally trained back-office personnel who possess deep expertise specific to particular processes, companies, and industries—it's precisely their insights that continuously feed Lio's long-running agents.

Seema Amble

This comes back to what I said earlier: procurement historically may have been more like a relatively closed module, but now it extends to legal, finance, and numerous different software systems and personnel—both specialists and generalists.

Elena Burger

Can we ground this in a specific vertical industry? For example, I'm a drone manufacturer, or a humanoid robot manufacturer—how many components do I need to procure, how many factories do I need to interface with, how many suppliers are involved? Vlad, could you pick a vertical industry you're currently serving that's using Lio to manage this kind of complexity, and walk us through their actual experience? I think this would make the whole picture more concrete.

Vladimir Keil

When we talk about procurement at Lio, we're not talking about laptops and pencils—those problems we solved three years ago. We're talking about manufacturing aircraft, robots, drones, and even the AI we're discussing today—because AI also needs to be built, and building means procurement. Building a data center requires procuring countless components.

Even minimal friction can have enormous cascading effects. For example, in a large project—building a data center or an aircraft—if a specific part arrives two weeks late, it could cause hundreds of millions of dollars in losses and delay the entire project.

All these decisions need to be coordinated. You need to figure out what you need, which suppliers to work with, who the most suitable suppliers are. But once these decisions are made, there's still a massive amount of operational back-office work. It may sound boring, but "operations" means: someone sends you an email saying "Sorry, this part will be two weeks late"—this is just one of 500 emails in the procurement manager's inbox. If missed, it's hundreds of millions of dollars in losses.

The bigger problem is that what's stored in the ERP system is just a date change, which looks trivial, but behind it could mean a hundred million dollars in losses. What our agents do every day is not just retrieve this information, but judge: will this have an impact? How big is the impact? How should we respond?

Elena Burger

When you're deeply embedded in the physical world's supply chain, which problems can you intervene in, and which are beyond your control? For example, if a shipment has an incident at sea, or a strait closes—some things you can do something about, some things you can't change. Where's the boundary?

Vladimir Keil

Essentially, this is a probability problem. You truly cannot change losses that have already occurred. But if you have sufficient contextual information, you can predict in advance—for example, a certain supplier's goods have a 20% defect rate, while another's is only 1%. The one with 20% defect rate might cost one-tenth of the other, but for a specific scenario, paying 10 times the price for higher delivery reliability is completely commercially reasonable.

The power behind this is: what you have isn't just internal enterprise context, but also external world information—market news, dynamics from both supply and demand sides, even prediction market probability assessments of certain types of disruptions. Combining this information allows for far more precise decisions than manual judgment.

In the long run, I don't think there are insurmountable limitations, just that prediction probabilities currently vary across different types of problems. What we're building isn't just enterprise-internal procurement automation, but something more macro—the way enterprises collaborate with each other, between buyers and sellers.

Elena Burger

You've described Lio as a multi-agent system. Can you specifically explain what each agent does?

Vladimir Keil

On one hand, Lio's agents span all four categories mentioned in Seema's article, chosen based on risk and complexity.

On the other hand, to complete an end-to-end task, agents need to share information and operate collaboratively in a specific sequence. When we say "multi-agent system," this is essentially what we mean.

Because a human-level task may involve eight stakeholders, three departments, five software tools—we need to cover all of these to truly complete the end-to-end work. These agents must communicate with each other, and only a multi-agent system can achieve this.

We initially started with retrieval-based, copilot-like approaches, then evolved to single agents, but quickly found it wasn't enough—even a single negotiation requires contract agents, news analysis agents, and others to participate collaboratively to complete the full work. This is our definition of a multi-agent system.

Seema Amble

Then let's get specific: suppose Boeing needs to procure a bolt. What does this procurement process specifically look like? Where does Lio intervene?

Vladimir Keil

Procuring bolts is actually one of the scenarios we can run fully autonomously, precisely through the multi-agent system.

The entire process starts with demand generation, and this step itself is already quite complex. Imagine a worker at a construction site who may only open his laptop once every two weeks, yet is required to use SAP or another ERP system to submit a demand request. This itself is enormous friction.

We made this step extremely simple: take a photo, or upload a part number and Excel spreadsheet, that's it. Those complex concepts in procurement systems—categories, general ledger accounts, framework contracts—nobody cares about them, and people outside the procurement department shouldn't need to understand them.

Next, our agents take over: checking inventory, asking whether other factories can transfer internally; if not, the procurement agent queries whether there are internal or external suppliers; then another agent drafts an RFQ and sends it via email; after which a flood of responses comes in—some content entirely in email body, some PDFs, some Excel, various formats. We retrieve this information, then proceed to the next step: judging based on price benchmarks whether there's room for negotiation. If there is, the agent decides the next strategy—it could be a strategic negotiation with human involvement, a fully autonomous negotiation, or an electronic auction. The corresponding specialized agent executes it, then completes subsequent order confirmation, logistics tracking, and invoice processing—the entire process runs autonomously end to end.

Of course, this is just the indirect procurement scenario—bolts and similar standard parts. In the direct procurement domain, the situation is more complex, and it's also an area we're deeply cultivating.

Elena Burger

What is direct procurement?

Vladimir Keil

Let me first clarify the difference between indirect and direct procurement.

The type of scenario I just described has automation as its core goal—running the procurement process fully autonomously and continuously uncovering savings along the way.

Our way of looking at these scenarios is: find things that need to be done a thousand times a day but ideally wouldn't need to be done at all; also find things that are currently done zero times but would have enormous impact on the income statement if done a thousand times a day—such as initiating autonomous negotiations on expenditures that have never been negotiated.

Indirect procurement includes: MRO consumables, factory construction, standard parts like bolts, as well as laptops, pencils, marketing services, like building a podcast recording studio—all of these fall under indirect procurement.

Direct procurement is different. When you're manufacturing an aircraft, you need core components that actually constitute the aircraft, from hundreds of strategic suppliers. Here we're not talking about 50,000 suppliers, but 100 to 2,000 suppliers, each potentially involving billion-dollar-level procurement amounts.

In this scenario, you don't want fully autonomous negotiation. You need three months of deep preparation—you need to check aluminum indices, oil price trends, analyze price changes, engineers need to verify technical drawings and component quality one by one. This is the truly exciting scenario for deploying AI agents.

Elena Burger

In these scenarios, I'd guess it's roughly expert engineers and procurement staff in the front, agents in the back? Or is it that agents are actually sitting at the negotiation table, really shaking hands with the other party? Are agents more like back-office support, or have they already moved to the front?

Vladimir Keil

Currently, in most cases it's still back-office. For those complex multi-million-level negotiations, 90% of the work is the preparation phase. The final result presented in the system of record—say, pushing from $1 billion down to $900 million—behind it are three months of preparation work and 10 people working full-time. This work indeed happens behind the scenes.

But we also have scenarios where we intervene in real-time. Imagine we're currently in a negotiation, I've done complete preparation, just like having these discussion outlines at hand right now. If during the negotiation, my screen pops up a real-time prompt: "Oil prices have risen 10%, therefore the other party is demanding a 10% price increase," and the system immediately provides analysis: "Oil prices have indeed risen 10%, but oil costs account for only 30% of this part, so you need to accept at most a 4% increase, not 10%." These kinds of real-time assistance scenarios are equally exciting.

Elena Burger

We know companies like Harvey and Decagon are deeply fine-tuning models. What underlying models does Lio use? How do you view fine-tuning and RAG?

Vladimir Keil

We use multiple models from multiple providers. We treat these general models as a commodity—reading PDFs, generating Excel, these tasks they're fully capable of.

But we also believe that for certain specific scenarios, by combining base models with RAG, it's possible to push completion rates to near 100%. However, there are also scenarios where even with the best base model and the best RAG, you can only reach an 80% performance ceiling.

A typical example is what cost engineers do: analyzing technical drawings, then determining "how much this part should cost"—this is called "Should Cost Modeling." We believe that through fine-tuning models, it's entirely possible to break through this 80% ceiling and reach 200% performance.

Another example is price benchmarking. Ideally, you just drag a quote in, and the system immediately gives a precise market reference price. But the problem is, this information isn't public data—it's scattered across enterprises, even across enterprises' private data, and general models can't be trained on it.

Our approach is: not just training a traditional large language model, but more like new types of models like Jef—the input is text data, but the output is a result, such as directly giving the most reasonable price. RAG can't do this, because price judgment is essentially an ineffable "intuition." Just like asking two people to do the same job, BCG's quote and McKinsey's quote might differ by 10 times, but an experienced procurement manager can tell at a glance which is reasonable—this kind of judgment is hard to explicitly rules-based, but can be learned through outcome-oriented training.

This is the opportunity we see in price benchmarking and should-cost modeling: not training traditional large language models, but training specialized models with results as output.

Elena Burger

Seema, we've touched on labs going deep into industry applications. When you think about what a vertical AI company with lasting vitality should look like, what traits do you pay attention to?

Seema Amble

One key is owning the end-to-end work—which is what we've been discussing—while building data assets and achieving things no one has done before.

Regarding "moats," frankly, it's hard to predict what future moats will look like. Looking back at the early stages of all the best companies, they weren't saying "I'll follow these six steps and then I'll have a moat," but rather: "I'm earning customer trust, I'm selling them more, the opportunity is right there."

Defensibility and durability, in my view, largely come from: customers' dependence on you deepening, them becoming increasingly unable to do without you, and you taking on more and more work.

For example: older-generation CRM systems just recorded all transactions; new-generation sales AI agents truly take over sales preparation, outbound outreach, handling inbound, and a large amount of work. The entire company's dependence on such products is an important signal on the path to a moat. What we often call stickiness and network effects ultimately all stem from this initial foundation: customers' real usage and the product's real value.

Elena Burger

Vlad, have you encountered customers or potential customers asking you: "Why should I buy your product? Why can't I just connect a model myself? Or use my existing system of record plus a model?" How do you answer them, how do you convince them to choose Lio?

Vladimir Keil

This is a completely reasonable question. One of our earliest products three years ago—entering information from quotes into SAP—was technically seemingly simple, but had enormous commercial value and spread widely in procurement circles.

This feature has now become an engineering capability test question, with candidates having 8 hours to complete it. What I want to say is: what we spent an entire engineering team and weeks building back then, engineers today can reproduce in 8 hours. So the question does arise: since it can be built in 8 hours, can the procurement department build it themselves in two months?

The answer is: they can, but only to 70% performance. The problem is, 70% performance doesn't equal 70% automation. In many scenarios, 70% accuracy means you still need 100% manual review—because 70% accuracy isn't enough, no one dares to release it directly, and it might actually be more troublesome than manual processing.

Seema Amble

Building on what Vlad said, let me add two points.

First, the普及 of general models and GPT-type products is beneficial overall. People are increasingly willing to trust AI tools, which in turn makes them more receptive to vertical specialized products. This increased familiarity and trust is good for the entire market.

Second, a few weeks ago I spoke with the management team of a Fortune 500 company. They had tried to build an internal collections tool themselves. It took three or four months, and ultimately they found: data context was insufficient, quality was too poor; recordings and screenshots piled up but couldn't be integrated into a usable system; then they faced the problem of two ERP systems coexisting, about to acquire another company and needing to update all mapping relationships, with exception handling logic requiring re-adaptation each time. They ultimately realized internal self-building was not feasible.

We hear similar stories repeatedly: enterprises say they'll build in-house, then discover it's essentially no different from all historical DIY attempts. Large enterprises will eventually recognize that their core competitiveness lies elsewhere, and building internal tools isn't what they should be doing.

Vladimir Keil

This is exactly what I wanted to say: they might achieve 80%, but the last 20% is the key, and that 20% often consumes 80% of resources. From launch to true production, what's needed is complete integration capability, memory management, workflow orchestration, and sometimes vertical-specific proprietary data.

So, to those Fortune 500 companies: you can certainly build it yourself, but this means AI procurement agent capability itself needs to become one of your core competencies, and you need to carefully evaluate whether this is worth your cost investment.

Elena Burger

Vlad, have you seen the supplier side starting to use AI or agents? If both buyers and sellers are fully AI-ized, what happens?

Vladimir Keil

We 100% believe that in the future both sides of transactions will deploy agents, and this makes perfect sense.

But interestingly, what we've actually observed is unexpected. The sales side has long led the procurement side in technology adoption, but what we've seen with suppliers of large enterprise customers is not this—suppliers are at most using meeting tools like Granola, far from truly deploying agents to automate work.

Procurement isn't sexy, sales is sexy, but they're actually two ends of the same process—procurement is the other side's sales.

But now, in those large industrial enterprises, the procurement side often has stronger bargaining power over suppliers. Take the typical automotive industry—OEMs have extremely strong leverage over component suppliers, able to dictate what systems they use, what quality standards they follow, how they respond to RFQs.

This brings an opportunity: we serve the procurement side, and the procurement side can require suppliers to also use Lio's agent system, thereby enabling both parties to achieve collaborative automation on the same platform, controlling both ends of the entire transaction.

Seema Amble

Elena, this reminds me of what we discussed earlier today: how to get opposing parties to collaborate on the same platform? This is the same in the legal domain—two law firms representing clients with opposing interests, but both sides can benefit from shared information: where's the latest draft, which clauses are still open for discussion, which have been agreed upon. This kind of collaborative tracking currently relies entirely on manual work, and this type of coordination work can absolutely be handed to agents.

Elena Burger

This is fascinating—both sides may be evolving at different speeds, but ultimately, they'll likely both operate on the same platform, achieving truly efficient collaboration.

Vladimir Keil

100% agree. We discussed this earlier too: if both sides have negotiation agents, will the price game become zero-sum? But don't forget, price is just the final result, the step that appears after thousands of other tasks are completed. And on those thousands of tasks, buyers and sellers have aligned interests—sellers want the process to be as smooth as possible, buyers want products to reach the market as quickly as possible. Going back to the example of building aircraft or data centers, no one wants a six-month delay just from processing supplier responses.

It's precisely on these thousands of shared-interest tasks that Lio can serve both buyers and sellers simultaneously, helping both parties automate this work. That's the most exciting part.

Elena Burger

There used to be a logic: don't over-customize software for one specific customer. But now, the emergence of LLMs and AI has dramatically reduced the cost of customization, no longer dragging down business scaling as much. Have you seen this change, and what does it mean for end buyers of software?

Seema Amble

Overall, there's now a massive amount of "forward deployment" work happening, partly because: understanding the current customer's actual data state is far harder than predicting the next customer's needs. We're still in the early stages of deployment overall, which is why products still need a lot of human involvement.

This is precisely what makes it harder for giants—their product feedback loops weren't designed for this: implementation teams exist, but more for after-the-fact remediation rather than truly integrating into product iteration.

The beauty of AI is that it continuously learns. Over time, with the right learning loops and workflows, systems can take on increasingly complex tasks. Customization itself is gradually being automated. I'm curious how Vlad and his team do it—many of our portfolio companies are rapidly advancing, letting customers control customization parameters and toggles through software interfaces themselves, rather than needing to bring in external teams for SAP implementation. Customization will ultimately be entirely software-driven.

Vladimir Keil

This is also why Lio's team structure is 85% engineers. Even those without engineer titles mostly have engineering backgrounds. We don't want to become a consulting firm.

Our goal is to build the best agents in indirect procurement and finance, but we must also face reality: when entering enterprise customers, forward deployment work is indeed needed, because every company has its own subtle process differences.

Our approach is: minimize customization needs at the product level, while providing extensive self-service capabilities. Our forward-deployed engineers' KPI is to automate their own jobs away—once achieved, they move on to the next task, then automate again. This is similar to some of Google's methodologies. This is how we do it, and the direction we continuously pursue.

Elena Burger

There was recently an important industry event—not Dreamforce, but the "Robots and Buyers Summit" hosted by Lio in New York. Can you tell us what the scene was like? What were buyers excited about, expecting, asking you?

Vladimir Keil

This is the third time we've held this event, near our office in New York, with over 100 senior procurement leaders—CPOs, VP-level decision-makers—attending. We make sure each time to invite people who truly matter.

The event had two prominent themes.

The first theme: looking back over the past twenty to thirty years, a dizzying array of tools have emerged in procurement—on LinkedIn you can find various procurement technology landscapes listing five or six hundred tools. But if you actually ask procurement practitioners, almost none of them like using these tools. And it's not just them: demand-side departments hate dealing with procurement, suppliers hate dealing with procurement, even procurement staff themselves don't like doing procurement. Why is this? Because those tools just make the original process a bit faster, but never truly change how people work. They're actually still working in email, Teams, Excel, and PowerPoint, with almost no AI-ization.

The second theme: we provided attendees with a cross-departmental procurement panoramic view, not focused on a single invoice feature, but the entire chain from "someone needs something, to ultimately receiving it"—spanning indirect procurement, direct procurement, logistics, and finance. We made these scenarios very concrete, built physical booths, just like we do in our New York office, letting attendees actually operate and experience each agent, not just look at slides. This was everyone's favorite part.

The next event is expected to have 700 attendees, an astonishing growth rate.

Seema Amble

A carnival for procurement people (laughs).

Elena Burger

Right, the next one is in Munich?

Vladimir Keil

Yes, we do Munich for Europe and New York for the US.

Elena Burger

Okay, if you're in the procurement industry, you know where to find it.

Seema, I have a question: what do you think makes people start getting excited about procurement? Is it the agent technology itself? The time saved? Or something else? Procurement has always been an area with extremely low NPS and is disliked. I remember chatting with someone from a procurement background seven or eight years ago, and he said he really didn't like talking about his product, was using Coupa, sort of making do, completely unwilling to switch—that was the most negative customer interview I ever did. What can truly make this field genuinely exciting?

Vladimir Keil

I think this is exactly what fascinates me about procurement. What drove us to enter wasn't what happened in procurement itself, but how people react to it. They're genuinely frustrated. This shows it's a highly emotionally charged topic—clearly B2B SaaS, yet with strong emotional tension. This is a good thing.

When you combine "high emotion" and "niche and obscure," the advantage emerges: because of long-term lack of breakthroughs, any progress easily amazes people. The real technological revolution in this field happened 20 years ago, and after that maybe just the interface got a little prettier, nothing else.

And procurement has another easily overlooked characteristic: it has enormous commercial impact. On the surface it looks like a cost center, but actually, a 1% procurement price reduction is equivalent to the same profit contribution as increasing sales by 10%. And procurement's impact extends beyond the enterprise—it determines how data centers are built, how aircraft are manufactured, how cars are produced, how drones fly.

"High emotion" + "long-term stagnation" + "enormous commercial impact" combined equals a trillion-dollar market opportunity. Of course, this formula doesn't just apply to procurement—other fields meeting these three conditions are the same.

Elena Burger

Vlad, thank you so much for sharing today. This conversation was incredibly fascinating!