Two Sigma Ventures reposted this
Don't Sell Work, Give It Away “Sell work not software” has been a common trope among VCs, but it seems to be aging quite poorly. The problem with this thesis is that it causes you to underwrite a TAM that’s roughly proportional to existing markets for human labor ($100Bs). But of course, customers won't pay "x% of human costs" in 7 years. They'll pay the prevailing rate of AI agents, which will be orders of magnitude cheaper than current human rates. My favorite new wave of startups is giving away agents at cost, with their actual revenue model predicated on adjacent offerings enabled by at-cost AI agents: 1) Financing - Credit: Offer agentic workflows for unstructured docs (e.g., invoices or inventory) at cost. Use that data to better price loans, inventory financing, etc. - Also applies to underwriting risk (insurance), equity (buying ownership), fraud (corporate cards), etc. 2) Marketplaces, Brokers, & Protocols - AI-native brokers: Refer to Jared Rosner's excellent piece. - Bonus points if non-users are precluded from participating in an industry, like Bloomberg IB. 3) Recommendation Systems - Having access to a customer’s docs could create a MUCH better ads, marketing, or co-selling service. 4) Vertical Integration with Operational Excellence - This also includes roll-ups and buyouts. Most VCs should stay away, as it requires a level of operational/financial sophistication that most early-stage investors are poorly equipped for. Of course, there are still opportunities for durable revenue by selling agentic software alone. But it’s difficult to imagine durability in agentic software revenue unless it's built on one of the following moats: 1) Too-Cheap-to-Replace - If the marginal cost per user is really small ($5-10), it's just an annoyance for the user to switch. - In this case, the number of potential users is all that matters. (Cursor?) 2) Physical Data Collection or Generation -Deploy proprietary sensors in the customer's environment/site that collect previously unmeasurable data. or -Generate novel datasets in-house. 3) Regulatory - Real-time policy adherence. (Valuable when the model providers can't incorporate policy updates quickly enough in their public API.) - Limited approved vendors. - Risk of fine or financial loss to the customer. 4) Platform -Allow (ideally non-software) 3rd-party services to build on your agentic workflows. There are other exceptions, so feel free to throw them in the comments! Last, but not least, there's brand & trust. Nothing new for me to say here. A lot of startups are claiming that they’ll be sticky due to the “moats” below, but these seem a lot weaker in an agent-centric world: 1) System of record 2) Cognitive switching costs (in the enterprise) 3) "Embedding deeply" in enterprise workflows If these moats do persist, they would strongly advantage incumbents, anyways. cc: Vin Sachidananda, PhD, Two Sigma Ventures Message me at nikhil@tsv.vc! https://lnkd.in/eH3dnnKY