AI Agents and the Future of Alt Data

AI agents need real-time data to price markets

Alternative data has stopped being an edge. Credit card panels that were alpha a decade ago are now table stakes, and tight pricing means buyers pay alpha money for smart beta. Kirk McKeown built primary research at Glenview and SAC, then founded CarbonArc to sell insight from its data rather than the raw asset. His bet is that AI agents will need permissioned, real-time private data to reason about markets.

Takeaways

• Card data moved from alpha to table stakes. Credit card panels were alpha in 2012 to 2014 and a setup advantage by 2017. Today anyone trading a 30 to 90 day setup needs them, so investors pay alpha prices for smart beta.
• CarbonArc moves insight rather than the asset. CarbonArc built a 10-million-card panel with Facteus and holds 450 data assets and 10 petabytes. Its MCP server is going into agent search firms such as Exa and Parallel and into retail platforms like Robinhood. Financial services is about 70% of the business.
• AI agents need a grounding data layer. Brett says agents answering stock questions pulled raw inputs from sources like the Motley Fool, so they reasoned on weak material. Kirk agrees the grounding layer decides the answer, and CarbonArc can supply it without owning the customer.
• AI workflows will match analysts within years. Kirk expects systematic AI-generated workflows to match strong analysts and PMs within a couple of years, with more volatility. Reasoning models must reason against real-time transaction and sensor data, and he is doing post-training work with CarbonArc data.
• Kirk is torn on the market's endgame. One future is a hot take world where stocks achieve singularity and are no longer traded. The other is data-native markets where models trade inputs that inform the problem. Brett calls CarbonArc's long-term thesis a place in price discovery infrastructure.
• Primary research teams must act as counterparties. At Glenview Kirk's group had no P&L responsibility and was not paid on P&L, which he says limited confirmation bias. Teams should be better at calls than the investment team and understand businesses, not stocks.
• Channel checks still matter where data is thin. Only about 150 tickers have rich data, Kirk says, so most names still need primary research. His example is a San Francisco paint contractor, where Sherwin-Williams lead times and distributor price points matter, not the contractor's own demand.

Kirk's channel check playbook

• Calls ran 45 minutes and were monthly rather than quarterly, tracking where comps would shake out and how wholesalers saw momentum and promotional cadence through department stores.