What VCs Really Think About Personal AI Agents | E2347

Investors described personal AI agents as a potentially major market over the next few years because they can run tasks in parallel and alert people when judgment is needed. They cited teams of five using thousands of agents, a chip-design company planning a 20-person team, and a 19-year-old running 50 agents. The scale is contested: one investor relayed hedge-fund skepticism about venture valuations near $10 billion for a company said not to be making money, while another saw a broad consumer opportunity. Speakers estimated that Facebook has roughly 3.5 billion users and Google, including YouTube, about 4 billion—around 70–75% of internet-connected people.

Raw intelligence does not make an agent useful or safe. Agents need context, permissions, guardrails, trust, and help integrating into people’s routines. Gmail and Apple have intimate user relationships and distribution that could make access difficult for new entrants; competitors may also need advertising networks, commercial relationships, wallets, or the ability to offer free or low-cost services. Specialized agents could still win through domain expertise. One investor questioned whether a roughly $10 billion entry valuation for Instinct was justified given the possibility that OpenAI, Google, Meta, or Apple could lead the consumer interface. The investor did not expect Muse to reach a billion daily users, but saw room for hundreds of millions.

Trust is an immediate constraint on adoption. One speaker avoided Instinct rather than grant access to all their data; another would not provide credit-card or Social Security information after a reported incident in which an agent exposed bank-account information in a company Slack. In a medical-bill test, Muse declined to pay using tokenized card data, while Instinct paid with a card stored in plain text, which the speaker said was not ideal. These experiences illustrate the trade-off between convenience and control, and the risk of agent fatigue if people must repeatedly share data and train new products. An agent embedded in daily workflows could become sticky, but handling intimate information requires trust and security.

The clearest near-term uses discussed were administrative and financial chores: paperwork involving accountants, healthcare, and insurers; splitting party expenses using purchases, an attendee spreadsheet, and arrival times; and using Claude with Granola, calendar, and email inputs as a project-research chief of staff. One speaker would pay to delegate low-stakes tasks such as calling venues for a child’s birthday party. Agents could also reduce friction in refunds, insurance claims, and rebates. A speaker described an agent escalating a six-hour flight-delay claim to a person and then a manager, while sending updates; they saw value if it could secure the refund owed. Another asked Instinct to seek a €600 delayed-flight refund, but the synopsis gives no outcome. Easier transfers of deposits could pressure banks to offer better rates, while businesses that do not accommodate agents could leave openings for startups.

Investors said AI lets founders build and iterate with smaller teams and less money, but does not replace the need to set a vision, recruit, and sell. They back both deep domain experts and curious outsiders who question industry assumptions, and said a founder’s beliefs about frontier models shape the opportunities they pursue. One firm shifted from roughly 80% horizontal SaaS to 80% vertical AI, arguing that specialized data can improve on general models. Examples included EvenUp’s personal-injury case and insurance-negotiation data, Aldot Health’s personalized biofeedback for early detection, and Kanu Galati’s investments in tax-focused Pramana Labs and financial-code developer Logos Research. Investors remain divided on whether data businesses can sustain an advantage as models learn from supplied data; recurring services embedded in workflows may be needed, and one investor was unsure whether demand for post-training data and evaluations would plateau.

The discussion also weighed AI’s scientific promise against practical limits. Dee Dee Das said OpenAI’s math result, its earlier Navier–Stokes result, and other research showed progress across four of the six unsolved Millennium Prize problems. Kanu Galati said formal proof-building helps explain such advances, while less formal real-world domains are harder to capture precisely. Chris Farmer expected AI-enabled research to improve understanding of the brain, complex systems, and cancer, including by learning from failed experiments. But a proposed cancer therapy still needs human safety evidence, and better battery chemistry still needs manufacturing and testing. Investors also noted that compute constraints can favor large labs able to spend billions, even as advances in data, algorithms, chips, and efficiency could lower costs and widen access. SaaS multiples have compressed with the end of COVID-era, zero-interest-rate valuations and concern about AI disintermediation, but speakers rejected the claim that SaaS is dead: AI may shrink seats and weaken switching costs, while products that extend from systems of record into intelligence and execution could grow beyond historical limits.