AI-first, not AI-feature: Sophia Zhao on the moats that survive the next model

"If you just have a feature, I worry that might be easily killed at the next model update of any of these big foundation players."

That line from Sophia Zhao is the whole investing thesis in one sentence. Sophia is a partner at Alumni Ventures, helping run the firm's new AI First Fund — a Seed-to-Series-A fund built around companies where AI is the foundation, not a bolt-on. We sat down with her on The QAI Podcast, and the conversation kept returning to one question every founder should be asking: when the next model ships, does your product get stronger, or does it get deleted?

Alumni Ventures is an unusual vantage point for that question. With over 1,600 portfolio companies and a habit of co-investing alongside Sequoia, a16z, Y Combinator, Founders Fund, and Khosla, the firm sees a very wide top of funnel. Its pitch, in Sophia's words, isn't money. "We don't bring the capital, we bring the network to help our portfolios grow." It also does something most venture funds don't: it gives accredited retail investors access to deal flow that's historically been walled off to institutions.

What "AI-first" actually means

The term has lost its edge. As Boyuan put it, everyone now calls themselves AI-native, and the phrase "gets thrown around quite a bit." Sophia skipped the definition and pointed at two portfolio companies instead.

The first was One Mind, which announced a $30M Series A the day we recorded, alongside Battery, Wing, and Primary Venture Partners. Founded by Amanda Kallo, One Mind builds what Sophia called "an AI superhuman teammate that actually does a sales job and not just assisting a sales staff." Not a co-pilot, not a wrapper — "a full digital teammate with a face, voice, personality, and a full go-to-market brain" that works around the clock, joins video calls, gives demos, and takes actions. The wedge is a mundane failure of B2B go-to-market: buyers wait days for answers that should take minutes, and "a lot of buyers choose whoever responds first."

The second was Tama AI, a voice-AI company for car dealerships that Alumni backed at seed — and that a16z later "took the entire series around." The stat Sophia cited makes a sleepy industry suddenly interesting: dealerships miss 45% of their inbound calls. Tama's agents are "trained on the data and workflow specific to that particular dealership," not one-size-fits-all. Her observation underneath it is worth keeping: "some of the best companies are always in the small lives" — boring, capital-intensive industries nobody's glamorizing.

The pattern across both: AI isn't doing the assistant's job. It's doing the job.

The four moats she looks for

When Boyuan asked what signals an AI product is actually investable, not just usable, Sophia gave a concrete checklist. Four things build defensibility.

A data moat. "All of these foundation models, they basically exhausted all the publicly available data out there." The edge now is proprietary data you can keep accessing and train a small or medium model on — "that will give you defensibility against these big LLMs."

A niche industry. Her examples were Harvey and Legora, both AI tools for law firms, defensible because they're built around an industry "that is highly sensitive to confidentiality" — the kind of space that's hard for OpenAI or Anthropic to walk into directly.

Partnerships. A proprietary relationship with a key player in your target industry — she named Nvidia — that other entrants can't replicate.

Workflow integration. This one was the most personal. "I hate to download a new app and open it every time I need to use it. I will love for it to be already integrated as part of my workflow." HubSpot's AI agent, Gmail's AI features — "I don't need to form a new habit." A tool that lives where the work already happens doesn't fight for adoption.

Underneath all four is the oldest signal there is: pull, not push. "If you build something people want to use that solves a person's pain point, people will be coming over and be like, please take my money."

Bubble talk, and where AGI actually is

We didn't dodge the hard question. An MIT report found 95% of generative-AI pilots are failing, and much of the funding momentum looks circular — hyperscalers pour billions into AI startups, which spend most of it buying compute back from those same hyperscalers. Are we near a top?

Sophia was honest that she's still working it out: "we're still formulating how we think about this." But she made two points that cut through the froth. First, the adoption gap is real and human-sized. As a power user she feels she lives "inside a bit of a bubble," while the people she meets volunteering to teach financial literacy at a high school say "I only heard about ChatGPT, I use it sometimes, but I don't use it that much." Second, beneath the hype there's "real revenue from a lot of the companies that we back."

On AGI she was un-hyped. Borrowing the SAE self-driving scale of 0 to 5, she put today's AI "probably between 1 and 2" — powerful, but without "genuine reasoning, planning, and world modeling." The fully autonomous stage is "coming for sure," but it isn't here.

The takeaway for founders

The throughline is future-proofing. For vertical SaaS, Sophia sees "a huge advantage where you can move fast by building on top of existing foundation models" — with a warning: think a few steps ahead about "what will the foundation models look like in the next six months so that you are not building a product just for now." When she notes that OpenAI hired over 100 investment bankers to work on a finance product, the message lands. The model companies are walking up the stack toward you. The defensible move is to build the moat — data, niche, partnership, integration — before they get there.

That's also why funding alone was never the point of this conversation. The bar to raise is higher now precisely because the tools to ship are better; everyone can move fast, so the contest is who builds best and finds product-market fit first. At QAI Ventures we'd put it the same way: funding is table stakes, and the edge is customers, code, and distribution. Sophia's four moats are the founder-side version of that idea.

Watch the full conversation with Sophia Zhao on YouTube, and find more episodes of The QAI Podcast at /podcast.