AI Agents vs. Incumbents in Enterprise Procurement
A16Z partners and Leo CEO Vlad Kyle analyze how AI-native startups outperform incumbents by owning end-to-end workflows rather than just system-of-record data. The discussion covers the four-tier agent maturity model, the strategic value of exception handling, and why internal builds often fail at the final 20% of performance.
The Shift from System of Record to End-to-End Ownership
The traditional software model, where incumbents own the system of record, is being disrupted by AI-native startups that own the entire workflow. As A16Z partner Seema Amble and Leo CEO Vlad Kyle explain, incumbents are limited by their siloed data and internal incentive structures. They can layer AI models onto existing tools but struggle to execute complex, cross-departmental tasks that involve legal, finance, and external supplier interactions. AI-native startups, however, can orchestrate these disparate data sources to complete the "end-to-end arc" of a job, such as resolving a complex procurement dispute or negotiating a multi-million dollar contract.
The Four-Tier Agent Maturity Model
The discussion outlines a critical framework for evaluating AI capabilities: retrieval, process, policy, and principal agents. Most incumbents remain stuck at the retrieval stage, offering chatbots that pull data from their own databases. True value emerges at the policy and principal levels, where agents apply judgment to ambiguous situations, such as determining if a service outage warrants compensation or how to negotiate a price based on market volatility. Leo’s strategy involves starting with human-in-the-loop process agents to build trust and data, then scaling to autonomous principal agents for high-stakes negotiations.
The Economics of Exception Handling
A key insight is that happy-path automation is a commodity. The real economic value in procurement lies in exception handling—managing fraud, mismatches, and supply chain disruptions. These tasks represent 80% of the actual work but are poorly served by traditional software. By focusing on these high-friction areas, AI agents can capture savings that were previously ignored due to lack of capacity. For example, automating negotiations for low-value items that humans previously skipped can yield significant aggregate savings.
Strategic Implications for Enterprises
Enterprises attempting to build internal AI tools often discover that reaching 100% reliability is prohibitively expensive. While they may achieve 70-80% accuracy, the remaining errors require full human review, negating efficiency gains. Specialized AI vendors, with their focus on vertical data and continuous learning loops, can reach production-grade reliability faster. The future of enterprise software is not just about better dashboards, but about agents that actively execute work, coordinate across departments, and interface with external parties, fundamentally changing the cost structure and speed of business operations.
Key insights
-
Incumbents are constrained by their system-of-record limitations and internal organizational conflicts, preventing them from delivering end-to-end AI solutions that cross departmental boundaries.
Impact: AI-native startups can capture market share by offering holistic workflow automation that incumbents cannot replicate without cannibalizing their existing product lines.
-
The majority of procurement value is hidden in exception handling and complex negotiations, not in routine transaction processing, which is already commoditized.
Impact: Focusing on high-friction, high-judgment tasks allows AI companies to demonstrate clear ROI and differentiate from basic automation tools.
-
Trust in autonomous AI agents is built incrementally through human-in-the-loop interactions, which provide the necessary feedback data to improve agent performance and reliability.
Impact: Startups should design onboarding processes that start with supervised autonomy to reduce customer risk and accelerate adoption of fully autonomous features.
-
Internal AI builds often fail at the final 20% of accuracy, where the cost of human verification negates the benefits of automation, making specialized external vendors more viable.
Impact: Enterprises should evaluate the total cost of ownership for internal builds versus specialized AI agents, particularly for complex, high-stakes workflows.
-
Multi-agent systems are necessary to handle the complexity of modern procurement, requiring coordination between specialized agents for sourcing, legal, and financial analysis.
Impact: Investors and builders should prioritize multi-agent orchestration capabilities over single-agent improvements to address real-world enterprise complexity.
Action items
-
Map your current workflow to identify where work happens outside the system of record, such as in emails or spreadsheets, and prioritize automating these cross-system interactions.
Impact: This reveals the true scope of AI opportunity and helps identify where incumbents are blind, allowing for a stronger value proposition.
-
Implement a human-in-the-loop strategy for new AI deployments to build trust and generate high-quality feedback data before scaling to full autonomy.
Impact: This reduces customer risk and accelerates the learning curve, leading to faster adoption and higher retention rates.
-
Focus AI development efforts on exception handling and judgment-based tasks rather than happy-path automation, as these areas offer the highest ROI and differentiation.
Impact: This positions the product as a strategic partner rather than a commodity tool, justifying premium pricing and deeper customer integration.
-
Evaluate the total cost of internal AI builds by accounting for the human verification required for the remaining 20-30% of errors, and compare this against specialized vendor solutions.
Impact: This provides a clear financial argument for outsourcing complex AI workflows to specialized vendors who can achieve higher reliability at lower cost.
-
Develop multi-agent architectures that allow specialized agents to communicate and coordinate, ensuring the system can handle end-to-end workflows involving multiple departments and external parties.
Impact: This enables the automation of complex, high-value tasks that single agents cannot handle, unlocking new revenue streams and operational efficiencies.
Quotes
“The opportunity for the AI native startup is to say, we're going to own that entire end-to-end arc.”
“No company and no enterprise starts with fully autonomous negotiation agents from day one. Why? Because they don't trust us and they don't trust the technology from day one.”
“It's actually only like 20% of the work of the job to be done of the problem. Because 80% of the problem is like, what if the invoice is fraudulent?”