Sarah Tavel on AI Applications, Model Economics and Venture Investing

Sarah Tavel tells Harry Stebbings that frontier AI models appear likely to require immense investment and, in her view, may be supplied by a small number of companies. She focuses on applications because she believes products that own the user relationship can create and capture substantial value.

Tavel describes joining Benchmark after Peter Fenton approached her while she was at Greylock. She initially declined because she felt loyal to her colleagues. Zillow CEO Rich Barton encouraged her to spend time with the Benchmark team and learn how they practiced venture investing.

She found the small, equal partnership and its commitment to founders distinctive. After getting to know the team, she joined; at the time of the interview she had spent nearly seven years there.

Asked what makes Fenton effective, Tavel cites his relentless curiosity, reading and willingness to learn outside his immediate field. She also admires his understanding of what motivates people and his ability to help founders become stronger leaders.

The lesson she wishes she had learned earlier is the importance of “why now.” A real technological or market catalyst can act like a current carrying a company forward, giving it room to make and recover from mistakes. Without one, founders may expend enormous effort for little progress.

Stebbings describes investing in BeReal around a thesis of authenticity but wonders whether that need was sustainable. Tavel agrees authenticity mattered, yet argues that the product competed for attention against TikTok and Instagram, particularly highly engaging algorithmic short-form video.

On AI, Tavel says the first wave often reinforced incumbents. Companies such as Notion and Adobe could integrate model APIs into existing workflows, offering users new capabilities without requiring them to adopt an entirely separate product.

She nevertheless sees a disruptive opportunity when startups sell completed work rather than a tool that makes an employee more productive. A company delivering an outcome has a different unit of value and pricing from an incumbent accustomed to selling software seats.

Stebbings asks whether models are capable of doing enterprise work end to end. Tavel calls it a spectrum: some specific tasks in areas like recruiting, sales and HR operations can already be automated, while more complex work remains difficult.

A human in the loop can bridge that gap. Tavel prefers a provider whose own staff perform quality assurance, rather than asking the customer to supply that labor, until models can handle more of the work themselves. This is her business-model preference, not a claim that every task can be fully automated.

Asked whether established service firms can use the same tools, Tavel says hungry challengers may be more willing to pass cost savings to customers. Faster delivery and a different price structure could help those challengers win share, although incumbents can adopt technology too.

Tavel argues that the application layer can capture considerable value because the application owns the relationship with the end user. As an investor in that layer, she favors companies that become more useful as their underlying models improve, potentially taking on more work and reducing the cost of human review.

Stebbings challenges her on distribution: an incumbent may bundle a feature that is only slightly worse than a startup’s. Tavel agrees this is a serious problem for an early application whose value comes overwhelmingly from an underlying model that everyone can access.

She is more excited by companies that own more of a workflow and build an experience not available directly from a model API. The distinction is not just an attractive interface; it is the amount of useful work the company does around the model.

For large enterprises, Tavel says it was too early at the time to distinguish lasting adoption from experimentation. She observes more adoption among digital-native, mid-market and smaller businesses eager to grow and gain share.

Tavel uses DeepL, a Benchmark investment, as an example of an enduring proposition: timely translation that customers find useful enough to replace some human translation work. She combines that first-principles assessment with cohort evidence rather than assuming early revenue proves permanence.

The evidence she looks for includes depth of use and continued engagement. Early cohorts are often too young for certainty, she acknowledges, but they can offer hints of whether customers would be disappointed if the product disappeared.

Competition among similar AI startups makes differentiation unusually hard. Tavel looks for founders with urgency, ambition and competitive energy who can navigate a crowded market. Her expectation of intense competition is an investment judgment, not a prediction of which company will win.

Valuations may be higher than traditional Series A expectations, Tavel says, if the opportunity is genuinely larger. She argues that selling a substantial amount of work could address more spending than selling a modest improvement in an employee’s productivity. Her 10-to-50-times illustration is a thesis, not a measured market result.

Stebbings worries that repeated large financings will dilute investors. Tavel says the rationale for taking dilution is that capital can let an exceptional team buy compute and train a model that creates a durable advantage in a very large market. Whether that thesis works remains to be seen.

She also warns against deploying capital from fear of missing out. Benchmark asks whether an opportunity could become an independent, enduring company and develop a way to escape competition, such as a network effect. She says each partner makes only one or two new commitments a year.

Tavel observes that business software has rarely relied on classic network effects. Instead, companies can build scale by serving a specific need, expanding distribution, adding features and increasing revenue from existing customers. She expects those economics to remain relevant to business-focused AI products.

On specialized versus multimodal models, Tavel suggests that a focused team such as ElevenLabs can devote more attention to its particular modality than a provider concentrating on its core frontier model. She presents this as a possible advantage, not proof that incumbents cannot catch up.

DeepL illustrates her distinction between a model and the surrounding workflow. She says the company developed its own language model and added customer integrations and vocabulary controls, building value beyond the initial copy-and-paste translation experience. Benchmark is an investor in DeepL.

Tavel expects frontier training to remain expensive because demand for compute and power rises as developers pursue more capable models. Even if chips and research become more efficient, she believes the race for the frontier could produce an oligopoly; meanwhile competitive pressure may reduce prices to end customers.

On open versus closed models, she says it is uncertain how a company recovers the cost of openly releasing a frontier model. Her tentative view is that the frontier may remain closed, while open models can be sufficient for many tasks. She notes that Meta’s investment in Llama could change that picture.

Her biggest concern is competition. Many startups are pursuing the same attractive opportunities, and battles for market share can require large amounts of capital. Some markets may support multiple winners, but the distribution of value is not yet clear.

Asked how Benchmark wins deals, Tavel describes the firm’s product as direct partnership by a general partner. She contrasts it with delegating functions to an internal platform team, saying direct involvement helps a partner understand the founder and the company’s needs.

Stebbings asks whether a specialist recruiter might be better qualified. Tavel argues that knowing the CEO, the team and the role deeply, and having an owner’s commitment to the outcome, matters alongside recruiting expertise. She says repeated recruiting work has also improved her own skills.

Benchmark’s equal partnership, she says, gives partners a shared economic interest in each other’s investments. That can make winning a deal and supporting a founder a collective effort rather than the responsibility of one partner alone.

When an investment fails, Tavel often finds that the apparent “why now” was less powerful or durable than she believed. She says a board member should do more than cheer: ask difficult questions, identify future scaling problems and help the CEO make incrementally better decisions.

A board member might help a struggling executive find mentorship or recognize that a different hire is needed. Tavel believes small improvements in decisions and recruiting compound, while emphasizing that the founders and their team perform the overwhelming majority of the work.

Tavel says Benchmark’s focused model will not suit every founder. Some prefer a broader platform or do not want a board member at that stage; she regards those as legitimate choices rather than failures. The firm may also decline to engage when a round is too large for its model.

She says Benchmark rarely departs from its board-involvement model, but can be flexible about round size when it strongly wants to partner with a founder. In assessing price, she asks whether a company can escape competition through a founder’s skill, network effects or other durable advantages.

Tavel warns against using a lower price to manufacture conviction in an investment. Benchmark concentrates on its initial investment and, she says, often invests little afterward. Her description reflects the firm’s stated approach and its commercial interests.

She proposes that existing investors should “earn pro rata” participation in later rounds by continuing to contribute. In her view, automatic demands for allocation can make it harder to bring in a new partner and increase founder dilution. Stebbings adds that dependence on investor reserves can discourage founders from being candid about difficulties.

Tavel names her long relationship with Chainalysis, where Benchmark led the Series A and she became the first board member, as especially meaningful. She says the experience helped her learn as the founders and the company grew. This is a company in her investment portfolio.

Her biggest missed opportunity, she says, was Ethereum. While at Greylock, she became excited about the potential of smart contracts but did not act. The lesson she draws is to find a way to act when she sees a technology she believes could be fundamentally disruptive.

She recalls a founder whose intrinsic motivation convinced her that they would push through obstacles; she says the company became remarkable without naming it. Asked what she had changed her mind about, she says she had not previously understood how present antisemitism was.

Tavel explicitly distinguishes support for Palestinians from antisemitism. She argues that repeated public expression can change what people feel able to say and can influence beliefs. The conversation then turns to the approaching 2024 US election, about which she expresses concern for democracy and says she does not closely follow polling.

Parenthood has not changed how Tavel evaluates investments, she says, but it has increased the opportunity cost of her time. She enjoys meeting founders even without a potential deal, yet now weighs those meetings against time with her children.

Asked about a recent investment, Tavel does not name the company because the deal had not been announced. She says she looks for a strong desire to work with a founder and closes by thanking Stebbings for the conversation.