Why Building an AI Agent Is Easier Than Deploying One

Strategic Context & Problem Statement

• Procurement historically operates in isolated "boxes" but now touches legal, finance, engineering, and disparate software systems; missing a single email about a two-week delay in part delivery can cause hundreds of millions in damages.
• Incumbent software vendors are limited by their "systems of record," which capture only final results (e.g., a signed price) rather than the complex, unstructured work behind them (30 stakeholder meetings, 500 emails, 20 Excel sheets, 3D modeling).
• 80% of the procurement "job to be done" occurs outside standard ERPs; standard systems only capture the "happy path" (100% of the software market) while ignoring exception handling (20% of the work but 80% of the problem).

The Incumbent vs. AI-Native Startup Dynamic

• Incumbents struggle to own the end-to-end arc because internal incentives conflict between selling workflow tools (for humans) versus resolving work entirely (autonomous agents).
• Legacy vendors often hold back on advanced agents due to fear of eroding customer trust if they ship premature autonomous features, whereas startups can earn trust through iterative human-in-the-loop adoption.
• A "slap on a chatbot" strategy is insufficient; incumbents often remain stuck in Retrieval Agents (information lookup) and Process Agents (applying strict rules) rather than Policy Agents (judgment-based application) or Principal Agents (strategic relationship management).
• Seema Ambul notes that while incumbents have distribution and trust, they face "classic incumbent issues" where different VPs and orgs sell conflicting products, preventing cohesive end-to-end ownership.

Leo's Multi-Agent Architecture & Capability Spectrum

• Leo deploys a spectrum of four agent types: Retrieval (information), Process (rule-based execution), Policy (judgment application), and Principal (strategic decision-making).
• Direct Procurement (e.g., aircraft, robotics parts): Involves multi-million dollar negotiations, complex 3D model analysis, and long-running agents with experts in the loop; these are not fully autonomous due to high stakes.
• Indirect Procurement (e.g., laptops, MRO parts): Can be fully autonomous for transactions previously ignored (e.g., under $50k) where capacity was lacking, turning "non-negotiated" spend into savings.
• Real-time Intervention: Agents can pop up insights during negotiations (e.g., oil price index up 10%, but product only contains 30% oil, suggesting a 4% price increase cap) to prevent bad deals.
• The company utilizes a multi-agent system where specialized agents (sourcing, RFQ, negotiation, invoice, logistics) communicate to execute tasks end-to-end, mimicking the 8-stakeholder, 3-department, 5-tool complexity of human teams.

Adoption Strategy & Trust Building

• Enterprises do not start with fully autonomous agents due to trust deficits; Leo utilizes a human-in-the-loop approach where humans provide feedback to agents, allowing the system to scale from 10,000 to 100,000+ successful negotiations.
• Customers were convinced to adopt the technology because Leo was "ahead of the curve," pitching problems and solutions before the broader market hype (e.g., ChatGP) matured.
• Customization is handled via a self-service model where 85% of Leo's workforce are engineers; Forward Deployed Engineers (FDEs) focus on automating their own jobs to reduce customization time.
• Internal builds by Fortune 500 companies often fail (reaching only 70% performance) due to poor context, siloed data, and lack of exception handling; the "last 20% of performance" requires the 80% of effort incumbents lack.

Market Dynamics & Future Outlook

• Supplier-AI Integration: Suppliers are currently less advanced than buyers but will adopt agents; Leo aims to push both sides of the transaction onto the same platform, aligning incentives (e.g., faster time-to-market) despite adversarial price negotiations.
• Economic Impact: A 1% increase in procurement savings equals a 10% increase in sales revenue for P&L impact, making procurement a "highly emotional, boring, and high-impact" category with a trillion-dollar opportunity.
• Model Strategy: Leo uses multiple foundation models as commodities but plans to fine-tune specifically for outcomes like "should-cost modeling" and price benchmarking where general models fail on proprietary data.
• Moat Definition: Durability comes from owning the end-to-end work, creating dependencies where the customer relies on the agent for complex tasks, rather than just data assets.
• Upcoming Events: Leo hosted a procurement leaders summit in New York with 100+ C-level attendees; the next event in Munich is expected to grow to 700 participants, driven by frustration with legacy tools that improved efficiency but not the actual user experience.