Chris Halanish calls the $100 trillion asset management industry “human duct tape.” The stack - QuickBooks, Excel, manual entry - is fragile. When a fund accountant quits, institutional memory vanishes. Hanover Park, Halanish’s firm, built AI agents with long-term memory to automate fund administration, capital calls, and data cleaning. They migrated a fund with 20 entities in six days; a process engineers thought impossible a year ago.
"The shift to AI-powered data cleaning and ontology mapping became viable only 3-6 months ago, enabled by OpenAI's Opus model series."
- Chris Halanish, This Week in Startups
Nathaniel Whittemore sees the same trend reorganizing all work. He outlines five new job archetypes: Prototyper, Builder, Sweeper, Grower, and Maintainer. The barrier between idea and function vanishes. In the agentic era, an HR professional needing a specific expense reporting tool no longer files a ticket with IT; they build it themselves. Back offices become hubs of micro-product development.
This moves software value from human-centric UI to the underlying data and logic. The a16z Show notes enterprise giants like SAP and Salesforce remain unkillable because they codify internal business rules, not just data. Steven Sinovsky argues that displacing SAP from a global automaker would dissolve the business; the stickiness is regulatory compliance and decades of edge-case customizations.
"Enterprise software stickiness comes from human UI muscle memory, organizational workflows, being the single source of truth, and the inertia of ongoing payment."
- Steven Sinovsky, The a16z Show
The pivot is toward headless software, where AI agents, not humans, are the primary users. Agents change interaction from workflow to analysis. Sinovsky defines an agent as a program that takes a long time to run and might not finish - essentially a bug rebranded as a feature. They don’t care about button placement; they care about data access and context stored in unstructured documents.
Startups avoid head-on competition with incumbents. Instead, they target the gaps between functional silos, acting as a layer of translation between sales and finance. Hanover Park charges funds a transparent, all-inclusive fee based on assets under management, abandoning seat-based pricing. This aligns incentives: the vendor becomes a partner in the fund’s growth, not a shelfware seller.
The change is structural. Goose, originally a standalone Mac app, pivoted to become the Goose Development Kit, a model-agnostic agent harness donated to the Linux Foundation. Steve Lee argues this creates a public-good framework, preventing vendor lock-in with OpenAI or Anthropic. The cost of making drops to near zero. Every department gets a maker.
