990 is a significant player in the execution layer of the agent stack. While many companies focus on the reasoning or interface layers, 990 is concerned with how an agent actually interacts with legacy, non-API-friendly enterprise software. They are essentially building the connective tissue between advanced large language models and the rigid world of industrial operations.
For the broader ecosystem, 990 represents the push toward verticalized agents. They demonstrate that the most valuable agent applications might not be general-purpose assistants, but highly specialized workers that understand the nuances of specific industries like supply chain or manufacturing. Their work pushes forward the standards for agent reliability and system integration in environments where failure has physical-world consequences.
The promise of artificial intelligence in the enterprise often hits a wall at the front door of the operations department. While marketing and engineering teams have adopted LLM-based tools for content generation and code assistance, the core business functions—procurement, logistics, and supply chain management—remain trapped in legacy ERP systems. These systems, such as SAP or Oracle, were designed as databases of record rather than platforms for automation. 990, a San Francisco-based startup, is building AI agents designed to bridge this gap by navigating the complexity of back-office workflows.
The company focuses on "AI-native operations." This is not a rebranding of Robotic Process Automation (RPA), which typically relies on rigid, rule-based scripts to move data between windows. Instead, 990 builds agents capable of understanding the underlying logic of a business process. If a shipment is delayed or an invoice contains a discrepancy, these agents can reason through the necessary steps to resolve the issue, communicating with suppliers or updating internal records without a human needing to map every possible branch of the decision tree.
The technical approach at 990 is informed by the background of its founder, Suhas Sunder. Sunder spent nearly a decade at Palantir, where he worked on large-scale data integration and operational challenges for major industrial and government clients. This experience is visible in 990's focus on messy, real-world data environments. Unlike generic agent frameworks that struggle when a database schema changes or a web interface shifts, 990 is designed for the high-stakes environments of manufacturing and logistics where data is often fragmented and siloed.
For these companies, the cost of human error in data entry or process management is significant. A misfiled customs form or a mismanaged inventory count can halt a production line. 990 positions its agents as a way to harden these processes. The agents are built to interact directly with the systems where the work actually happens, rather than acting as a simple chat interface that sits on top of the data.
990 enters a market that is increasingly crowded with startups claiming to automate the enterprise. However, its specific focus on the operational layer differentiates it from the flood of sales and marketing automation tools. The company targets mid-to-large enterprises in heavy industries—sectors that have historically been slow to adopt new software because of the sheer complexity of their existing infrastructure.
By focusing on the execution layer—actually performing the tasks within the ERP rather than just summarizing information—990 is making a bet that the future of enterprise AI is in the autonomous agent model. This requires a level of reliability and visibility that many consumer-grade AI tools cannot provide. The company is currently scaling its team in San Francisco and working with early design partners to refine its agents for specific industrial use cases. As the agent ecosystem matures, the success of 990 will likely depend on its ability to prove that its agents can be trusted with the core operational data of a business.
Autonomous agents for managing back-office ERP and supply chain workflows.
990 is hiring.