Jeff Seibert on Product Pivots, AI Infrastructure and Angel Investing

Jeff Seibert, founder and CEO of Digits and former Twitter product leader, joins Harry Stebbings. They disclose that Stebbings later became a major investor in Digits, a material relationship for a conversation about the company. Seibert recalls loving building as a child, first with Lego and then through programming.

At Twitter, Seibert learned that aggregate data can obscure very different user needs. Sports fans, public figures and users seeking anonymity may want different experiences; he argues against designing for an imagined average user. He prefers building for a problem one knows personally, while checking that others actually share it.

He values a rapid product cadence and criticizes waiting weeks for data before every decision. In his view, founders also need deliberate execution: decisions about time, hiring and priorities should follow a conviction that can later be examined. He presents this as a management philosophy, not proof that speed always produces better results.

Seibert says individual contributors can become managers without having practiced management, and leaders may receive too little candid feedback. At startups he has tried a two-month trial for new managers, with the option to return to an individual role. He believes compensation and recognition should support strong individual contributors as well as managers.

Digits runs weekly sprints and Friday retrospectives. Teams describe “anchors” that slowed them down and “breezes” that helped, then surface important points to a company-wide discussion. Seibert says this encourages feedback and incremental improvement. He also celebrates small wins so the team can feel progress even while larger ambitions remain unmet.

Digits began with Seibert’s frustration at having real-time product analytics at Crashlytics but late, difficult-to-understand financial statements. He says the goal was to make accounting timely and intuitive for founders. The company is his own business, and Stebbings is an investor, so their positive assessment is interested.

Seibert says Digits spent about three years struggling with data quality and bookkeeping automation. In 2021 it pivoted to reporting and collaboration tools; he reports adoption by roughly 1,000 accounting firms and 5,000 downstream businesses. After advances in language models, the team returned to its original automation ambition. These adoption and capability claims are his account, not independently verified here.

On when to pivot, Seibert says the founder must judge whether a credible path remains and have enough cash to pursue a new direction. He suggests at least a year of runway, while Stebbings questions whether that leaves enough time to build and fundraise. Seibert calls it difficult and argues a pivot requires commitment rather than a few half-hearted trials.

He says he brought Digits’ leadership together, laid out the problems and options, and aligned on a new direction quickly. He thinks founders must earn team trust and keep investors and the board informed rather than surprise them. Stebbings and Seibert both question whether a team can fully commit to a pivot it fundamentally rejects.

Seibert credits Benchmark investor Peter Fenton with recognizing opportunities in markets where customers dislike the status quo. Accounting, in his telling, resembled an underserved market where even a modestly better experience could matter. This is an investor’s and founder’s interpretation, not a measured comparison of accounting products.

Stebbings asks whether reliance on external AI providers creates risk. Seibert views AI as a fast-evolving tool and says weekly planning helps the company adapt, but the customer problem should remain the focus. He compares the period to earlier platform shifts in the web and mobile.

In this November 2023 discussion, Seibert predicts language models will become more commoditized and that open alternatives will approach proprietary models for many uses. He expects an open base model plus fine-tuning for specific tasks. These are time-specific forecasts, not established outcomes.

He says Digits trains its own models and has worked on accuracy because finance applications cannot tolerate invented figures. He distinguishes large-scale foundation-model training from fine-tuning on smaller, high-quality task-specific data. His statement about Digits’ accuracy is a company claim without independent evaluation in the interview.

Seibert distinguishes a thin application that merely scripts a model from a defensible product that solves a deeper customer problem. He predicts OpenAI may become more like infrastructure, hosting models and providing capabilities to other businesses, while warning that startups can be overtaken if they merely fill a gap on a model provider’s roadmap. The prediction reflects his interests as an AI application founder.

He notes the challenge of compute costs and speculates Apple could combine on-device models with its own chips and privacy-oriented design. At the time he thinks Apple is behind but has an opportunity if sufficiently capable models can run locally. This is a hypothetical future, not a claim about the performance or privacy of any specific deployment.

Asked about sensitive enterprise data, Seibert compares concerns about AI providers with earlier concerns about cloud services. He expects clearer restrictions and trust to develop, while warning against adopting technology merely because it is fashionable. His prediction does not establish any provider’s actual data practices.

He thinks consulting for AI implementation may be a large market but is more excited by products that automate substantive workflows. He predicts adoption could be faster than mobile because users need not buy new hardware and can interact through familiar interfaces. Stebbings notes transitions can take longer than expected; the rate remains uncertain.

Seibert considers Google especially exposed if AI changes search and argues it should be willing to disrupt its own business. He expects compute to become faster and cheaper but says he is not a hardware expert. He also predicts AI will be absorbed into existing pricing models rather than necessarily creating a new one; these are 2023 opinions.

Seibert says clean, high-quality training data is difficult to obtain and describes Digits as having a proprietary data set of 100 million financial transactions. He says its terms allow data use to improve the product and that the company trains internally rather than sharing that data externally. These are his statements about Digits’ practices, not an independent legal or privacy assessment.

He speculates that a consumer photo service could have aided model training, then says “presumably” when asked about permission. That is not evidence of the provider’s actual use of private photos. He also discusses potential cloud-company acquisitions and AI features at incumbents as speculative competitive judgments.

Stebbings questions the claim that 90% of early investments go to zero. Seibert shares his own angel portfolio: since 2014 he had made 97 investments, of which 30 had failed outright and 19 were still marked at 1x. He says approximately ten holdings mattered most and most gains remained on paper; very few had returned cash. These are portfolio-specific, self-reported figures.

Asked whether he trusts those marks, Seibert says some companies may still carry 2021 valuations despite lower secondary-market prices. He describes some prices as down substantially, but does not identify the companies or provide a representative data set. A paper valuation, a secondary quote and realized proceeds are different measures.

Seibert says his most successful liquidity outcomes have come from private secondary sales. Some holdings that reached an IPO subsequently fell sharply during lockups before he could sell. He recommends disciplined check sizes and a diversified portfolio after two larger, high-conviction angel investments went to zero; this is his personal strategy, not investment advice.

He describes investing $10,000 in a founder’s earlier project that eventually became Alchemy and selling on the secondary market for a reported 200x return at the height of the crypto boom. He says that realized sale paid for the rest of his angel investments, and emphasizes founder persistence and timing. This is a self-reported realized result, distinct from remaining paper gains.

Stebbings recalls declining an opportunity to sell shares in another company and later regretting it. Seibert describes taking a 3x cash return on part of a position whose remaining paper value subsequently rose much further; he is uncertain whether the paper value will be realized. On asking a founder for cash back, he says the answer depends on the founder’s conviction, performance and circumstances.

Seibert expects pressure on valuations and fundraising after the difficult 2023 market, and sees a potential opportunity for early-stage companies to recruit strong people. Stebbings notes that established employers may offer more cash compensation, which can keep talent in place. These are forecasts and competing considerations at the time, not verified later outcomes.

Looking across his portfolio, Seibert says the most successful founders are often deeply focused on customers and personally connected to the problem, not necessarily the most technical. Stebbings suggests frequent childhood moves as another possible pattern. Both treat these as personal observations rather than a demonstrated causal rule.

In quick-fire discussion, Seibert voices concern about climate change and near-term risks, while the two discuss unequal responsibility and effects across countries. His specific timeline and the emissions comparison offered in the conversation are views and estimates, not independently established by the interview.

He predicts AI will raise productivity without replacing many jobs, drawing on historical technology transitions. That is a contested forecast, not evidence that particular workers or sectors will be unaffected. He also recommends startups focus primarily on customers rather than competitors, and make reversible decisions promptly so teams can keep learning.

Seibert says a year earlier he had expected useful AI to be years away. Looking toward 2033, he hopes Digits will become a widely used, real-time accounting platform. His climate remarks and product ambitions are forecasts from the interview, and the latter directly concerns his own company.