a16z: After Claudeforce, Vertical AI Startups Must Win with 'Learning Loops' and 'Data Assets'

The announcement of the Claudeforce partnership between Salesforce and Anthropic has brought renewed attention to the importance of Incumbent Systems of Record. Existing enterprises are using AI to control data access and build their own agents, expanding their scope beyond simple information storage to actual task execution (Action). However, combining general-purpose agents like Claude or Codex with existing systems is insufficient to fully resolve complex tasks.

Limitations of General-Purpose Agents and Opportunities for Vertical Startups

General-purpose agents can access multiple systems, but they face limitations in task execution due to latency issues when using MCP and data inconsistencies between applications. Additionally, they cannot automatically access information held by external parties (e.g., buyers and suppliers). In contrast, vertical AI-native startups can secure deeper access, intentional data assets, and specific context by focusing on particular tasks.

The Importance of Learning Loops

The core competitive advantage of startups lies in learning loops. Completed records alone are insufficient as a training curriculum for tasks; signed contracts or closed tickets show only the final outcome, excluding considered alternatives or reasons for exceptions. Startups must improve model performance through learning loops that use expert feedback and outcomes to refine subsequent attempts.

Harvey's Case and Market Evaluation Criteria

Legal AI startup Harvey post-trained open-weight models using synthetic data and expert-generated data without using customer data. By manufacturing a curriculum for model training and evaluation through the creation of approximately 1,750 legal task environments, Harvey achieved high accuracy without accumulating years of customer history.

a16z presented four axes for evaluating good vertical AI markets:

• Can experts quickly judge AI accuracy and explain how to improve it?
• Is the task difficult enough that judgment, rather than simple rules, is critical?
• Do tasks occur frequently enough for the product to learn?
• Can the startup expand from starting with a single task to performing the entire workflow?

Ultimately, even if existing enterprises own the records and Labs control the front door, vertical AI companies can win by performing the work itself best.