Alex Lieberman spends much of his working life inside large companies, and that has made him doubt a celebrated group of AI startups. Lieberman works at Tenex, a firm that builds AI systems for big enterprises. On an episode of the Latent Space podcast, recorded at AI Engineer New York and published October 21, 2026, he said he and his Tenex colleague CJ Hess had been discussing "vertical" AI just before the conversation began. Vertical AI means products built for one industry or one profession, such as law or medicine. He said the two of them had been "expressing an overall bearishness" about such products.
The pushback came almost at once.
Lieberman's case against vertical AI
Lieberman gave two reasons. The first came from what Tenex sees inside its clients. In large companies, he said, "there is so much nuance and so much inertia," with many systems and people that have to be wrangled.
The second was about what buyers now expect. "We are in an age where people are expecting one of one software solutions," Lieberman said, meaning software built to fit a single customer. He said this explains why "every software company is basically building up a team of FDEs." FDEs are forward-deployed engineers, who work closely with a customer to adapt the product to that customer's needs. He described the trend as a "convergence happening of services and software."
That led him to two questions. Software businesses are prized for high margins, because selling one more copy costs little, while custom work for each client costs a lot. So how could vertical products "do this at software margins if everyone expects a one of one solution?" And: "why should we not expect like frontier models to just gobble them up?" Frontier models are the most capable general-purpose systems from labs such as OpenAI and Anthropic.
"Generally bearish" and the response
The discussion then made the question concrete. Harvey for lawyers and OpenEvidence for doctors were named as classic examples of companies winning their domains, and Lieberman was asked whether he was bearish on them. He said he was "generally bearish on them."
The response was blunt: so far, it went, he had been wrong.
The first counterargument was data. Access to data that ChatGPT will not have was described as a moat, the industry term for an advantage that competitors cannot easily copy. Anthropic and OpenAI, the argument granted, are acquiring the same data sources, but are not as dedicated as the specialists. OpenEvidence shows what such access can look like. In February 2025 it announced a multiyear content agreement with NEJM Group. The deal covers material published since 1990 in the New England Journal of Medicine and its sister publications, and that material informs answers in OpenEvidence's medical search product.
The second counterargument was focus. Specialist companies, the argument went, build interfaces and features customized for their users because those are the only users they serve, and tabular review in legal AI, offered by both Harvey and its rival Legora, was the example. Legora's Tabular Review puts documents in rows and a lawyer's questions in columns, and each cell can be expanded to show the reasoning and source material behind the answer. Harvey has published a product demonstration of the same approach, in which five questions are run across 800 emails to produce 4,000 data points for building a case chronology, with links back to each source document. It is a demo, not a measured accuracy study.
Lieberman summed this up as "just focus, basically." The discussion added that specialists care about their customers' last mile, and that once professionals are trained on a piece of software, it tends never to go away.
The Bloomberg example
To illustrate the point, the conversation turned to an experience from a finance career. The same information was always available from Nasdaq and the New York Stock Exchange, so why pay $50,000 a year for Bloomberg? Partly because all of one's peers used it, the argument went, and partly because Bloomberg consolidated and specialized the information to avoid mistakes. It was also tied into a trader's trading system, chat and news.
Lieberman had a different explanation: "I think people love the DMs in Bloomberg." The reply conceded that point but maintained that, as a trader, the tools themselves get real use.
The email app Superhuman came up as another example. Everyone has Gmail, the argument ran, but Superhuman is a premium Gmail for people who live in email all day. It may not have been the largest startup in the world, but it shows there is a market for a better specialist tool. In this view, demand for software and for tools to do one's job is huge: general assistants such as Claude and ChatGPT will meet some of it, and for the rest, people will want something better.
The discussion then found common ground. The total market would grow and a rising tide would lift all boats, the argument went, so the remaining question was how much of the market each kind of product would win. Lieberman said he "totally" agreed.
Agents as the product, or agents on top of a product
The conversation then drew a sharper line between kinds of vertical products. Apps rebuilt from the ground up around AI, with Superhuman and its "cool hotkeys" as the example, were held up as products worth liking. Products whose main offering is an AI agent doing work in one industry, meaning software that carries out tasks on its own, were expected to struggle unless they have a strategic moat such as compliance or proprietary data. The reasons given were that customers can spin up their own agents quickly, those agents keep improving as models get better, and many of the tasks will simply fall within what general models can do.
Products that build an application and tools around the work, change how that work is done and add an agent on top were expected to succeed. That echoed an argument made earlier in the episode about changing the underlying process before handing it to an agent.
Bring your own agent
The segment closed with a related, less provocative view than Lieberman's. The idea is that people do not want to use a vendor's own agent; they want to use the agent they already have and plug the vendor's data into it. Every vendor wants customers to pay $20 a month for its own agent, the argument went, but customers cannot keep stacking up such subscriptions. The expectation voiced in the episode was that a small set of companies will own the interface, with everyone else plugging into it.
As a sign of this, the discussion pointed to Salesforce, described as one of the "system of record" companies, whose databases hold a business's official customer and sales records. As characterized in the episode, Salesforce is effectively telling customers they will not use its agents and is offering APIs and MCP connections for their own tools instead. MCP, the Model Context Protocol, is a standard way for AI assistants to connect to outside tools and data.
Salesforce's own announcement is more specific. Its August 25, 2026 Headless 360 announcement describes opening its business functions to outside agents through MCP while keeping existing identities, permissions, validation rules and business logic in place. Its examples include an agent updating sales opportunities, retrieving contracts or launching workflows without opening Salesforce's usual interface. When the announcement was made, the Headless 360 MCP server was in open beta and the Data 360 MCP server was generally available.