$100T is managed by “human duct tape” | E2308

Approximately $100 trillion in global assets is managed using QuickBooks, Bill.com, and Excel, with fund administrators acting as human middlemen who withhold client data and force fund CFOs to email requests for their own financial information. Quarterly reporting to limited partners such as Harvard and Yale endowments is delivered on a lag, and institutional knowledge disappears when a fund accountant leaves. Hanover Park, founded in 2024 by Chris Halanzik, was built to replace this infrastructure.

Hanover Park is structured as an AI-native services company rather than a standalone SaaS product, on the explicit reasoning that a pure software tool would be commoditized by Claude and ChatGPT. The company is building an ERP for funds as a system of record, with AI agents handling accounting and financial reporting layered on top. It employs no product managers and no designers, only engineers working alongside fund accountants who serve dual roles covering client services and informing product development. The team is approximately 50 people in New York City and doubled in a single quarter before recording.

The complexity of the underlying work is significant. Large funds like Blackstone can have hundreds of legal entities within a single fund, each requiring separate profit and loss allocations, and every limited partner can carry different economic terms governed by the limited partnership agreement. Hanover Park addresses institutional knowledge loss by building AI agents with persistent memory. A migration for a venture capital fund with roughly 20 entities and several funds was completed in six days, with the limited partner portal live at that point. A comparable migration on a traditional platform would take approximately 24 months. The one-click migration capability was described as not feasible a year before recording and only became possible three to six months prior, enabled specifically by Claude Opus 4.6. Human review teams still check agent outputs given the accuracy requirements for institutional limited partners, and migrating a large fund would take closer to 30 days rather than six. The limiting factor is context complexity of a given fund rather than model intelligence.

Hanover Park grew from $1 billion in assets under management to $15 billion in roughly 12 months and reached $20 billion by approximately 15 months total. The company raised $27 million when AUM was at $15 billion in March. It charges basis points off AUM rather than per-transaction fees, bundling all services into one price, though the specific rate is not publicly disclosed. At a hypothetical 25 basis points on $20 billion AUM, annualized revenue would be approximately $50 million. The company currently has 10,000 institutional limited partners on platform and is focused exclusively on closed-end funds including venture capital, private equity, and private credit.

SaaS pricing has evolved from boxed software to per-seat to active-user to usage-based token models, with outcome-based pricing predicted as the next step. Slack charges on an active-user basis and sends monthly utilization reports to build customer trust. Figma charges only for editors rather than viewers and applies a pricing model its CEO Dylan Field does not personally favor because customers want it. AI token costs are creating the same budget anxiety that SaaS seat costs once did but at much larger scale, and enterprise software buying cycles that took months in the SaaS era are now compressed to weeks or days. The bottom-up SaaS adoption model, where individual users spread tools organically before enterprise contracts follow, was cited as a key driver of faster sales, with Granola attracting venture investment partly because employees were already using it without being directed to.

Enterprise security risk is higher in the AI era than in the SaaS era because stolen data can train models that replicate at scale rather than simply being leaked. Jason Calacanis described Google algorithmically demoting Mahalo around 2007 alongside eHow and Answers.com, removing 80 to 90 percent of traffic overnight and forcing layoffs of 100 writers, then approximately five years later extracting those same publishers' answers into its own one-box results. The argument made in the episode is that Google has consistently optimized for monetization at the expense of users and publishers, and that whichever company beats Google at AI becomes the most important company in the world because Google controlling both AI and search would make it the arbiter of truth and speech.