Elena Burger

1. Key Themes

The Work Lives Outside the System of Record

The core argument for AI-native startups is that the incumbent's system of record captures only the result of work, not the work itself. Incumbents are bounded by their own database, while the real job spans email, spreadsheets, contracts, engineering data, and supplier conversations.

Vladimir Keil (Leo): "You see in your ERP system 8K for aluminum, but you don't see that maybe the supplier did like a pushback and asked for like 10K. You don't see that like a cost engineer had run three weeks of Excel sheets and 3D modeling to find out the prices of the part." [00:05:12] He concludes: "Most of the work in procurement actually happens outside of this ERP or any system of record." [00:05:36]

Seema Amble (a16z) frames the opportunity: "The legacy incumbent is limited to their system of record and that record that they have, and they're not completing the end-to-end job." [00:03:12] and "The opportunity for the AI-native startup is to say, we're going to own that entire end-to-end arc." [00:00:10]

The Four-Agent Taxonomy: Retrieval, Process, Policy, Principal

Seema lays out a ladder of agent capability defined by the judgment required. Retrieval agents pull information, process agents apply rules, policy agents interpret ambiguous rules, and principal agents weigh relationship trade-offs beyond policy.

Seema Amble: "A year ago, I made this meme, which was the slap on a chatbot strategy, which is essentially like all the incumbents effectively had a chatbot that sat on top of the system record." [00:06:32] On the top rung: "The principal agent, you're actually weighing, okay, should we offer more compensation? Because that was a pretty terrible outage and we want to reserve the relationship and it's worth doing more beyond even what a process or policy is." [00:07:59] Her read on incumbents: "They're very much still limited to, I would say, retrieval and a little bit of process. They've not gotten into more judgment." [00:08:28]

Why Incumbents Are Structurally Held Back

Incumbents have distribution and trust, but face internal conflict between selling workflow tools and selling outcomes.

Seema Amble: "If you're resolving a customer support issue end to end versus providing a workflow for a support agent, a human agent. Those are different buyers... those two teams are in conflict." [00:09:50] Vladimir adds the trust asymmetry: "Incumbents already have the trust, but this also means they can destroy the trust if they ship something like too early and the product maybe doesn't work." [00:11:49]

The Last 20% Is the Moat: Exceptions Are the Real Job

The happy path is a commodity; the exception handling is where value and defensibility sit. This also explains why internal builds stall.

Vladimir Keil: On invoice software, "that's like 100% of the software market... But it's actually only like 20% of the work of the job to be done... 80% of the problem is like, what if the invoice is fraudulent? What if there's like a mismatch?" [00:12:14] On DIY: "You can build this in eight hours... but you will only reach 70% of the performance... 70% of performance... doesn't mean 70% automation." [00:40:53]

Earning Trust Through Human-in-the-Loop Feedback Flywheels

No enterprise starts with fully autonomous agents. Human-in-the-loop is both a trust-building mechanism and a data-collection engine for customer-specific learning.

Vladimir Keil: "No company and no enterprise starts with fully autonomous negotiation agents from day one... By having this human in the loop approach, we are feeding our agent with all the feedback and all the learnings. And then they suddenly trust us for like 10K negotiations, 20K negotiations, 100K negotiations." [00:00:24]

Moats Are Hard to Forecast; Dependency Is the Signal

Rather than predicting a moat, look for whether the customer becomes dependent on the product doing the work.

Seema Amble: "It's really, really hard to forecast your moat going forward. If you look back at all the best businesses at the early stages, they were just thinking about, okay, I'm winning customer trust. I'm selling more to them." [00:00:49] And: "Old CRM company was a log for all the deals. New sales AI agent is actually owning a lot of the sales prep process and the outbound process... The overall customer is dependent on that product. And that's like a really important signal of getting to the moat." [00:38:31]

Multi-Agent Systems Mirror Cross-Functional Human Work

A single agent cannot complete end-to-end jobs that involve many stakeholders, departments, and tools.

Vladimir Keil: "A human level task involves eight people, eight stakeholders, and maybe like three departments and five different software tools... Only with a multi-agent system you can do a job end to end." [00:26:52] He notes the evolution: "You can't solve a negotiation even without having a contract agent without maybe having an agent looking at the news." [00:26:52]

Buyer-Led Adoption Pulls Suppliers Onto the Platform (Agents on Both Sides)

Procurement holds power over suppliers in industrial settings, creating a path to own both sides of the transaction. Agent-to-agent negotiation is less zero-sum than assumed because price is just one outcome of thousands of aligned tasks.

Vladimir Keil: "In those industrial companies... procurement has the bigger power to the supplier... why aren't we also pushing them to Leo agents, that are also helping them automate the work, owning then both sides of the transaction." [00:45:41] And: "Price is the outcome of like 5,000 different other tasks that happened. And on those 5,000 other tasks, they have the same incentive." [00:47:25]

Forward-Deployed Engineering as an Automate-Yourself Function

Customization is shifting from consultants to software, with FDEs measured on automating their own work.

Vladimir Keil: "The job of our FDEs and forward deployed engineers is on the one hand making it self-service, but internally is automating their own job... literally your job is like to automate yourself. And then if you automate yourself, you go to the next task." [00:51:34] Seema adds: "A lot of our companies... more of the customization is being handled in an automated way and the customer is able to turn the knobs and levers." [00:50:04]

Proprietary Outcome Data Justifies Fine-Tuning Beyond LLMs

General models are treated as a commodity; the differentiated layer is proprietary, tacit-knowledge-based pricing intelligence.

Vladimir Keil: "We use multiple models from all providers and we really see this as a commodity." [00:34:28] But for price benchmarking: "Those are all proprietary data based on one enterprise across multiple enterprises. So general purpose models can't train their models on that." [00:36:22] He is exploring models trained on outcomes: "If you ask them to write down the rules, they can't do this because it's just gut feeling and instinct." [00:37:17]

2. Contrarian Perspectives

Boring, Hated Categories Are the Best Markets

Most investors chase glamorous categories. Vladimir argues the opposite: unsexy, emotionally charged, long-neglected categories are where AI can most easily amaze buyers and capture outsized impact.

Vladimir Keil: "You have boring, highly emotional, and then plus crazy business impact... and then you have like a trillion dollar business opportunity." [00:57:25] On neglect: "The really last revolution they have seen is like 20 years ago. And then maybe a nicer user interface 10 years ago. But nothing else happened." [00:57:00] Seema corroborates the pain: a past procurement customer was "the most disgruntled customer call I've ever done out of like millions of them." [00:56:11]

The Real Prize Is Negotiating Spend Nobody Was Negotiating

Most pitches focus on automating existing work. Vladimir points to work that humans never did at all, where the downside of a mediocre agent is near zero.

Vladimir Keil: "They just didn't care about everything which happened below 50K... You can just send an enterprise an invoice for 40K. They will probably not negotiate because they don't have the capacity to do so... What is the risk now of having a bad negotiation agent? Nearly zero." [00:17:52] And: "What are they doing zero times a day? But if a business would do this thousand times a day, that would have a crazy P&L impact." [00:30:57]

Buyer and Seller Agents Aren't Purely Adversarial

Conventional wisdom says agent-vs-agent negotiation is zero-sum and will devolve into an arms race. Vladimir argues price is the only zero-sum variable among thousands of shared-incentive tasks.

Vladimir Keil: "Sales wants to have as little friction as possible. Buyers want to have a really fast time to market... All of those 5,000 other tasks, the incentive is exactly the same." [00:47:54] Seema extends this to law: "Two law firms with clients with different interests, but both benefit from knowing... here's the latest draft. Here are things that have been agreed upon." [00:46:27]

Sellers' Sales-Side AI Lags Procurement in Industrial Markets

The default assumption is that sales is ahead of procurement in AI adoption. Leo sees the opposite among large industrial suppliers.

Vladimir Keil: "The sales side was always ahead of the procurement side. But what we're now seeing with those suppliers for these Fortune 500 companies, this actually is not true... they're advanced and, let's say, video recordings and using tools like Granola, but not really having agents deploy, automating the work." [00:44:48]

Being Slightly Ahead of the Curve Beats Being Right Early

Rather than betting on a distant vision, Leo's edge was pitching the next capability customers would pay for and shipping it within weeks.

Vladimir Keil: "Three years ago, it was just retrieving a document. Not impressive at all today. But three years ago, this was crazy impressive. So we pitched this to customers. We find out, okay, that's a real problem... And then we were able to ship this some weeks later." [00:13:36]

3. Companies Identified

Leo

AI agent platform for enterprise procurement (indirect and direct), founded by Vladimir Keil. Mentioned as the focal company; spans retrieval through autonomous negotiation agents. Growing event series (third event had 100+ procurement leaders; next expected around 700 people in Munich). Org is "85% of the people are engineers" [00:50:45]. Quote, Vladimir Keil: "We span across all of those categories and it really depends on the complexity and the risk our agents take... we already have use cases where we run fully autonomously." [00:10:57]

Salesforce (Agentforce)

CRM incumbent. Mentioned as an example of incumbent distribution advantage and as the target of the "Cloudforce" (Claude plus Salesforce) hypothetical. Quote, Seema Amble: "When Agentforce launched, it was very easy for customers to say, yeah, I'm going to sign up for the Salesforce agent, especially if it was offered at almost no extra cost." [00:09:21]

Anthropic / OpenAI

Frontier labs that incumbents partner with to bolt on model capability. Quote, Seema Amble: "Some of them are saying, okay, let me partner with OpenAI, Anthropic, one of the labs, and try to build out, take that model capability and complement what they have." [00:08:54]

Harvey

Legal AI company; cited as an example of a vertical AI company fine-tuning models. Quote, Elena Burger: "We know that companies like Harvey and Decagon are really fine tuning models now." [00:34:13]

Decagon

Customer support AI company; cited alongside Harvey as fine-tuning models. Quote, Elena Burger: "Companies like Harvey and Decagon are really fine tuning models now." [00:34:13]

SAP and Oracle

ERP systems of record that Leo's early invoice and quote agents push data into. Quote, Vladimir Keil: "You match this across other documents, and then you push this back to SAP Oracle, like a very clear process." [00:12:14]

Coupa

Legacy procurement system of record, cited for customer dissatisfaction. Quote, Seema Amble: "I use this legacy system of record, I'm on Coupa and I don't want to buy anything else." [00:56:11]

Granola

AI meeting notes tool used by suppliers, cited as the extent of supplier AI adoption. Quote, Vladimir Keil: "Using tools like Granola... but not really having agents deploy, automating the work." [00:45:17]

Polymarket

Prediction market named as a potential external context source for disruption forecasting. Quote, Vladimir Keil: "Maybe even having context of some bad sale on Polymarket. Okay, those disruptions are going to happen." [00:25:19]

Boeing, Airbus-type Aerospace Builders (Boeing named)

Used as the example for complex direct procurement. Quote, Seema Amble: "An airline company needs to procure a bolt for, say, Boeing needs to procure a bolt." [00:27:50]

Accenture

Cited as the old model of SAP customization via consultants. Quote, Seema Amble: "You brought in Accenture to do your SAP customization." [00:50:19]

Google

Cited as a reference for the practice of automating your own job. Quote, Vladimir Keil: "I think it's also an approach that Google or so is doing." [00:51:34]

BCG and McKinsey

Used as an example of two vendor quotes whose prices could differ 10x, illustrating tacit pricing judgment. Quote, Vladimir Keil: "If you would give me a quote from BCG and a quote from McKinsey, they could do the exact same work, but this could be a 10X different price." [00:36:22]

4. People Identified

Vladimir Keil

Co-founder and CEO of Leo. Highlighted for staying one step ahead of the capability curve, building a trust-earning product motion, and designing an engineer-heavy, anti-consulting org. Quote: "We were able to pitch the next generation of agents... pitching this to them, finding out it's a problem. And then we're able to ship this very fast." [00:13:11]

Seema Amble

Partner at a16z; author of "The Incumbents Are Coming," which introduced the retrieval/process/policy/principal agent framework. Quote: "The legacy incumbent is limited to their system of record... and they're not completing the end-to-end job." [00:03:12]

Alex Rampell

a16z partner; source of the "distribution versus innovation" framing of incumbent-vs-startup competition. Quote, Seema Amble: "It's a fight between distribution and innovation, which to take my partner Alex Rampel's phrase." [00:02:49]

a16z host of the episode; drove the discussion on customization and agent-to-agent dynamics. Quote: "LLMs and AI in general... it might increasingly be possible to customize without slowing yourself down as a business too much." [00:48:40]

5. Operating Insights

Turn Your First Product Into the Hiring Test

Vladimir uses the company's own first product as a take-home: engineers get eight hours to rebuild it. This both screens talent and continuously recalibrates the team on how fast the build frontier is moving, and what is therefore no longer defensible.

Vladimir Keil: "We give this to people to build this and they have eight hours to do so... a product that one of our first use cases can now be somehow built by engineers within eight hours." [00:40:23]

Make Customer Autonomy a Dial, Set by Risk and Complexity

Rather than a binary autonomous/non-autonomous product, calibrate human involvement by deal size, relationship sensitivity, and complexity. Low-risk work runs fully autonomous; strategic work keeps experts in the loop and uses their feedback to train the system.

Vladimir Keil: "It depends on the budget approval, how complex it is, how risky this is." [00:10:57] And for multimillion-dollar negotiations: "We on purpose have always experts in the loop... there's an agent running for multiple hours and then we ask for feedback of the cost engineer, and then it does the next work." [00:17:03]

Staff Like a Product Company, Not a Services Company

Keeping 85% engineers and giving FDEs the KPI of automating their own work prevents the drift into consulting that plagues enterprise AI vendors.

Vladimir Keil: "Obviously, we don't want to be a consulting company, right? So we make sure that we have overall the best agents." [00:50:45]

Build Physical, Cross-Department Experiences to Sell Abstract Agents

Because "AI agents, that's very abstract," Leo builds walk-through booths and hosts curated C-level events, framing the pitch around cross-department outcomes instead of single features.

Vladimir Keil: "We're not talking about 'look at this crazy invoice feature that we developed,' but more like someone needs something. And in the end you have it on your table." [00:54:33] And: "We're building up booths... you can walk through the booths and experience all of those agents really hands on." [00:54:56]

Pitch the Next Capability, Then Ship Within Weeks

Sell what's slightly beyond current expectations once you've validated the problem, then deliver quickly. This builds trust and keeps the product on the frontier.

Vladimir Keil: "We pitched this to customers. We find out, okay, that's a real problem. That's a real use case they would pay for. And then we were able to ship this some weeks later." [00:13:11]

6. Overlooked Insights

Context Beyond the Enterprise Walls Is a Latent Data Moat

Briefly, Vladimir describes agents combining internal stakeholder context with external signals (news, supplier reliability, even prediction markets) to predict and price risk. This reframes procurement from cost-cutting to probabilistic supply-chain risk management, where paying 10x for a more reliable supplier can be the rational choice.

Vladimir Keil: "You have one supplier where 20% of the goods are missing and then 1% of the goods is missing. And maybe this one with 20% is 10X cheaper, but for this use case, it's fine for you to pay 10X the amount because you have a higher probability that this thing actually arrives." [00:24:52] He adds that it is "much bigger than just procurement into a company. It's more like intra company and how enterprises are doing business with each other." [00:25:48]

The Real Failure Mode of Procurement Is a Missed Email, and Systems of Record Store Only the Date

A throwaway operational detail carries the whole thesis. The system of record logs a shipping date, but the cause and consequence of a slip live in one of hundreds of unread emails, and a missed email can mean a hundred-million-dollar project delay. The wedge is not automation of tasks but continuous monitoring and judgment over unstructured communication.

Vladimir Keil: "This is one of 500 emails in the Outlook or Gmail of a procurement manager. And if they missed this email, hundreds of millions of damage done... the only thing they store in their system of record is then just the date." [00:22:55]