Robert F. Smith

Asked on camera to answer the narrative that AI is about to destroy software companies, Smith opened by conceding part of it. Software that repurposes data already available in the public marketplace and resells it has no right to exist, he said, and AI "will eat those companies. No, no question about it."¹ He owns roughly ninety enterprise software companies outright, which is the only reason the concession is worth anything.

Smith is the chairman, founder, and chief executive of Vista Equity Partners, a private equity firm that buys enterprise software companies and then operates them. In a January 2026 interview he described the firm as 25 years old, with assets over a hundred billion dollars and about ninety portfolio companies.¹ That seat is what makes his account of AI and software different in kind from the rest of the argument: the public-market investors, venture investors, and defending founders who carry most of the debate are all adjacent to the outcome, while Smith is the buyer of record with the profit and loss statements in front of him. It is also the reason to discount him, since the interview is a fundraising-adjacent segment aimed at family offices and registered investment advisers, and every figure in it is self-reported and unaudited.

Background

Smith began his career at Goldman Sachs in investment banking before founding Vista.¹ The firm's method has stayed narrow ever since: one industry, enterprise software, bought outright and run rather than traded. He is candid that the fundraising channel is a byproduct of the operating model rather than a marketing program, since founders who sold their companies to Vista, took the liquidity event, watched the returns, and then asked to invest built the high-net-worth base over 25 years, and the firm says it caps that allocation.¹

The three-category sort

His distinguishing intellectual move is that he refuses to defend the whole category. Pressed on whether AI kills software, he compresses his answer into a list: "agentic enterprise software, rule of 70, and no right to exist."¹ The first category is software whose agents can perform the tasks inside a workflow the vendor already owns, at higher frequency and higher precision than a person using a tool, which produces the thesis he states twice in the same conversation, that AI will enable enterprise software to eat services. The second is the margin case, filed here as the rule of 70, his own coinage for a replacement to the old rule of 40. The third he gives away without a hedge, and the reader is directed to the file on no right to exist for why a concession against interest is the strongest evidence available inside a promotional interview.

The dividing line between the categories is a single phrase he repeats twice: "those companies that have sovereignty and dominion over the workflows and data sets have an opportunity to become part one or part two."¹ The mechanism he offers for why that dominion holds is a statistic he attributes loosely: "less than 1% of enterprise data actually is in the environment that has been used to train these large foundational models. Less than 1%. I think IBM put a study out about that."¹ The attribution is his own hedge, and the number carries a great deal of the argument.

The factory

Vista's operating claim is that the conversion is not a project but a standing capability, and its credibility rests on a prior run rather than the current one. The firm built a factory roughly 25 years ago to move enterprise software companies off customer-owned hardware and into the cloud, says it converted more businesses that way than any other institution, and measured a two-and-a-half to three times pickup in economic rent per conversion from eliminating the hardware refresh cycle and the systems-administrator layer.¹ The second factory, moving cloud companies to generative AI, was stood up more than two and a half years before the interview; thirty-plus companies had been through it and were producing revenue, with the remaining thirty to forty due over the following quarter.¹

What the factory manufactures, on his account, is not agents but precision. Agents are cheap; hitting the reliability enterprises require is not. His illustration has become the archive's clearest statement of the enterprise precision threshold: 93% precision is fine for booking a Thai restaurant, and "does not work in banking. It does not work in insurance. It does not work in automotive."¹

How he operates

Three habits show through. He sorts before he argues, which is what separates his position from a simple defense of his asset class. He reads the capital cycle as a supply chain rather than a bubble: infrastructure "is a cost of goods for us and as more capital flows into it, it becomes a lower cost of goods," so the overbuild everyone else prices as risk he books as gross margin on his own agents.¹ And he instruments himself the way he instruments the portfolio, running a personal agent across roughly twenty monitored categories daily, queryable across the ninety companies for operating metrics and adoption rates, and retuned every week with his team.¹

He is also unusually specific about the numbers that cut against the enthusiasm. Agentic coding delivers 30 to 50 percent productivity on new code and somewhere between 2 and 12 percent on existing code, a gap of roughly an order of magnitude that most greenfield-only figures conceal.¹ On unit economics he supplies the ratio an owner would actually track: "20 cents worth of inference can lead to five, eight, ten dollars of savings."¹ On employment he does not soften the conclusion, saying some job categories will disappear while offering the standard consolation that the person using AI will replace your job.¹

Where things stand

Smith reads the enterprise software repricing as an entry point rather than a warning, calling the decline from the 2020 to 2021 run-up "a wonderful buying opportunity especially as you're transforming businesses to become agentic."¹ The reader is directed to the file on Brad Gerstner, who is standing on the other side of exactly that trade and who puts the same sector in Warren Buffett's too-hard basket, warning that anyone buying the dip "still could be catching a falling knife."² The disagreement is not about the price, since both men agree software got cheaper. It is about whether the replacement question can be handicapped at all, and Smith's answer is that the too-hard verdict is a failure to disaggregate a sector into its three parts. The parallel breaks in a way worth checking: if his categories separated as cleanly as he claims, public markets full of category-one software companies should already be sorting them away from category-three ones, and an undifferentiated sector-wide derating is evidence they are not.

No plate is assigned here. Vista is assembled rather than grown, and his stated edge is a read on which assets are buyable that other allocators decline to underwrite, which is what put an earlier reading on this file. What the file cannot support is that the return is fixed in the purchase: on his own account it is manufactured afterward, by the factory, and measured per conversion. The record stays open until the factory reading is tested against a second independent source.

Key facts

• Chairman, founder, and chief executive of Vista Equity Partners, described in a January 2026 interview as 25 years old, with assets over a hundred billion dollars across about ninety enterprise software companies.
• Began his career at Goldman Sachs in investment banking before founding Vista.
• Sorts enterprise software into three categories, agentic, rule of 70, and no right to exist, and concedes the third to AI without hedging.
• States the survival test as "sovereignty and dominion over the workflows and data sets," repeated twice in the same interview.
• Reports a two-and-a-half to three times economic rent pickup from Vista's earlier on-premise-to-cloud conversion factory, the precedent on which the generative AI factory rests.
• Puts agentic coding productivity at 30 to 50 percent on new code and 2 to 12 percent on existing code, and per-unit agent economics at 20 cents of inference against five to ten dollars of savings.
• Treats infrastructure overbuild as a falling cost of goods for his portfolio rather than as bubble risk.
• Runs a personal agent across roughly twenty monitored categories daily, retuned weekly with his team.