Why AI Agents Can Beat the Incumbents

Episode Summary

Executive Summary: The episode argues that AI-native startups can beat incumbents when they own the end-to-end workflow, not just the system of record. Using procurement as the example, Seema Amble and Vlad Kyle explain how Leo’s agents coordinate messy work across email, contracts, finance, legal, and suppliers, progressively earning trust from retrieval to autonomous negotiation. The discussion centers on exceptions, multi-agent orchestration, human-in-the-loop deployment, and why procurement is a high-impact vertical AI opportunity.

Main Topics: Why AI-native startups can still beat incumbents: Seema argues incumbents have data, distribution, and system-of-record control, but startups can win by owning the full job-to-be-done across fragmented workflows that extend beyond the incumbent product. Procurement as an end-to-end workflow problem: Vlad explains that procurement is not just the final price in an ERP; the real work happens across emails, spreadsheets, legal review, finance, engineering analysis, and supplier negotiation. Agent maturity: retrieval, process, policy, principal: The conversation lays out four agent levels, from simple retrieval and workflow automation to judgment-based policy decisions and principal-level tradeoffs that require business discretion. Trust and human-in-the-loop deployment: Both speakers stress that enterprises will not adopt fully autonomous agents immediately; trust is built through human oversight, evals, and gradual expansion from simple tasks to high-stakes decisions. Multi-agent systems and exception handling: Leo positions its product as a multi-agent system that coordinates different tasks and data sources, with a major focus on handling exceptions rather than only happy-path automation. Direct vs. indirect procurement: Vlad distinguishes low-risk, high-volume indirect spend from strategic direct procurement for aircraft, robots, and data centers, where pricing, quality, and timing have huge operational consequences. Vertical AI moats and customer customization: Seema and Vlad discuss durability in vertical AI as emerging from customer dependency, learning loops, and product depth—not from a simple feature or chatbot layer.

Key Arguments: Incumbents can add models on top of their systems, but they often remain trapped inside the system of record and miss the actual work done across tools, people, and exceptions. AI-native startups win by owning the end-to-end arc of a business process, such as support resolution, legal workflow, or procurement negotiation. Procurement’s surface-level record looks simple, but the real operational work is distributed across many stakeholders and channels that incumbents do not capture well. A retrieval agent or chatbot is only the first stage; meaningful value comes when agents can apply process, policy, and principal-level judgment. Enterprises will not trust fully autonomous negotiation agents on day one, so human-in-the-loop systems are necessary to learn, de-risk, and scale adoption. The hardest and most valuable part of procurement automation is exception handling, not the happy path of ingesting an invoice or filing a document. Leo’s product is designed as a multi-agent system because real jobs require coordination across multiple steps, tools, and specialized agents. General-purpose models are useful, but vertical use cases often need harnesses, integrations, memory, workflows, and sometimes fine-tuned or outcome-based models to reach production quality. Internal builds often stall because enterprises underestimate how much context, mappings, and exception handling are required to make a workflow actually work. When both buyers and suppliers use agents, the market can shift from just price optimization to coordinated automation across the broader transaction lifecycle.

Data Points: Suppliers in aircraft procurement: Thousands - Vlad says building an aircraft requires procuring thousands of suppliers. Email delay impact: Hundreds of millions - A late supplier email can create hundreds of millions of dollars in damage if missed. Procurement workflow inputs: 30 stakeholder meetings, 500 emails, 20 Excel sheets - Vlad describes the hidden work behind a seemingly simple contract or purchase outcome. ERP-visible result: 8K - Example of a simple recorded procurement price in the system of record. Procurement effort hidden outside ERP: 80% of the problem - Vlad says process automation covers only about 20% of the job; exceptions make up most of the work. Process automation share of software market: 100% - Vlad frames invoice software as a complete market for a narrow happy-path process. Enterprise negotiations scale: 10K, 20K, 100K negotiations - Vlad describes progressively increasing trust as the agent proves itself. Fortune 500 event attendance: Over 100 procurement leaders - Attendees at Leo’s New York event for procurement leaders. Future event attendance: Around 700 people - Vlad says the next event in Munich is expected to grow substantially. Engineering ratio at Leo: 85% - Vlad says 85% of Leo’s people are engineers. Customization build time: 8 hours - A basic Leo use case can now be built by engineers within eight hours as a demo or interview exercise. Performance gap of internal builds: 70% - Vlad says DIY builds may reach around 70% performance but still require nearly all the work. Indirect procurement supplier scale: 50,000 suppliers - Vlad contrasts indirect procurement at large scale with direct procurement. Direct procurement supplier scale: 100 to 2,000 suppliers - Vlad says direct procurement is strategically concentrated among far fewer suppliers. Complex negotiation duration: Three months - Vlad describes strategic direct procurement negotiations as long and highly prepared. Price vs sales leverage: 1% savings equals 10% more sales - Vlad argues that small procurement savings can have outsized P&L impact.

Pivotal Quotes: "The opportunity for the AI native startup is to say, we're going to own that entire end-to-end arc." — Seema Amble: Explaining why startups can still win against incumbents with strong systems of record. "No company and no enterprise starts with fully autonomous negotiation agents from day one." — Vlad Kyle: On why human-in-the-loop deployment is necessary to build trust and improve agents. "Most of the work in procurement actually happens outside of this ERP or any system of record." — Vlad Kyle: Describing the gap between what enterprise software records and what procurement teams actually do.

Implications: Vertical AI winners will likely own workflows, not just records. Expect more human-in-the-loop agents, deeper exception handling, and multi-agent systems that coordinate across departments and even counterparties.