Interview, Fireside Chat
Why Building an AI Agent Is Easier Than Deploying One
- An AI-native startup is projected to capture the entire end-to-end procurement arc by overcoming legacy incumbents' limitations to system-of-record functionality, though enterprise adoption of fully autonomous agents will not occur from day one as organizations lack initial trust.
- Deployment strategy involves an initial human-in-the-loop phase where feedback trains agents to handle escalating negotiation volumes of 10,000, 20,000, and 100,000 cases, eventually transitioning from simple process agents to exception handling and long-running autonomous agents to address the 80% of problems involving fraud and mismatches.
- Differentiation strategies include targeting Fortune 500 suppliers who currently lack agent deployment to automate response standards, expanding beyond intra-company processes to buyer-supplier interactions, and utilizing multi-agent systems to coordinate tasks across eight stakeholders, three departments, and five software tools.
- Operational scope will distinguish between indirect procurement (MRO, laptops, marketing services) handled by fully autonomous agents and direct procurement involving strategic suppliers with up to $1 billion in spend, where autonomous negotiations are replaced by multi-month human-engineer analysis of technical drawings and industry trends.
- The technology will enable non-procurement staff to initiate demands via photo, quote, or document uploads without requiring SAP or ERP access, while agents will execute inventory checks, RFQ drafting, price benchmarking, and negotiation using context from news, prediction markets, and supplier reliability data.
- Customization will shift toward self-service "knobs and levers" allowing customers to build simple retrieval agents in eight hours, though reaching 100% production performance requires addressing the final 20% of work through specific integrations, memory, workflows, and vertical data to handle exceptions.
- Specific high-stakes scenarios involve complex negotiations for items like aircraft parts where a two-week delay can cause hundreds of millions in damages, prompting expert-in-the-loop protocols for cost engineering feedback on deals involving 3D models and technical drawings.
- Real-time capabilities will include pop-up insights on index changes, such as a 10% increase in oil prices, to calculate immediate price adjustments based on product composition, alongside adversarial coordination in legal work to track open issues and reduce friction for both buyers and sellers.
- Financial incentives drive the opportunity as one percent savings in procurement equates to ten percent of sales, enabling enterprises that previously ignored deals below $50,000 to negotiate via agents due to capacity constraints, while human-in-the-loop approaches remain necessary for relationship-sensitive sectors like podcast studios.
- The company plans to host events growing from 100 to 700 attendees to demonstrate hands-on agent functionality, while incumbents attempting to partner with model labs face internal conflicts where workflow product VPs and end-to-end resolution product VPs have divergent incentives.
- Business risks include the difficulty of forecasting durable moats in early-stage trust-focused markets, the potential for incumbents to destroy existing trust by shipping premature products, and the high cost of failure if agents miss supplier disruptions, alongside the limitation that 70% automation does not equal 70% of total performance.
- The market outlook suggests a trillion-dollar opportunity where forward-looking engineers automate their own jobs to reduce customization, while suppliers of large enterprises are expected to adopt agents driven by procurement department standards rather than current capabilities in video recording tools.
- Future development includes training models on outcome-based data like perfect prices rather than just general LLMs to capture proprietary insights, with a belief that both buyer and supplier sides will eventually deploy agents to automate the entire transaction, even when disagreeing on price.