Seema Amble, a fintech investor at Andreessen Horowitz, posted a long thread on X on September 3 arguing that AI agents make incumbent “systems of record” more valuable, not less, while leaving room for vertical AI-native startups to win specific jobs. Her trigger was Salesforce’s new Claudeforce integration with Anthropic, which lets users work with Salesforce data from inside Claude instead of opening the Salesforce app. The stakes are which side of enterprise software captures the value as agents take over more of the work: the incumbent that owns the underlying data, or the challenger that owns the interface.
Amble’s core claim is that owning a record does not mean owning a job. A signed contract is not itself the legal matter, she wrote; a support ticket is not the customer’s actual problem, and a Salesforce opportunity is not the sale. She built a four-tier ladder of application agents: retrieval assistants that summarize and answer questions, process agents that update records under fixed rules, policy agents that apply an organization’s playbooks to ambiguous cases, and principal agents that make open-ended strategic calls. Most incumbent products, in her account, have moved from the bottom rung to the middle over the past year: Docusign’s Iris now reviews contracts, Atlassian’s Rovo routes requests, Klaviyo’s Composer builds and assesses marketing campaigns.
It is worth stating plainly what Amble is: an investor at a16z who backs, and profits from, the vertical AI-native startups she describes as having a durable path to winning. Her thread reads as market analysis, but it doubles as a case for why the category she funds still has room to compete against both Salesforce-style incumbents and Anthropic’s general-purpose Claude. That does not make her argument wrong. It does mean every claim in it should be read as the view of someone with a position in the outcome, not a neutral survey of the field.
The mechanism Amble offers for why focus beats breadth is a learning loop specific to a narrow job. A vertical startup that owns an entire workflow, she argues, sees the corrections an expert makes and the reasons behind them, not just the finished record. In her framing, a closed ticket records how things ended, while the hypotheses a support team raised and discarded stay invisible. She points to Harvey, the legal AI startup, which built roughly 1,750 synthetic legal-task environments with expert rubrics (each scoring against about 50 criteria) to train its Tenet model before it had years of customer data to draw on. According to Amble, that let Harvey manufacture a training curriculum rather than wait to accumulate one.
Amble’s own framework, though, understates how the fight usually gets settled. The incumbent-versus-startup question in enterprise software rarely turns on which model reasons better. It turns on distribution and data access: who already has the login the customer uses every day, and who controls the permissions to act on that data. By her own account, Claudeforce leaves Salesforce holding the CRM data and the write permissions while Claude becomes a new front door. That is a distribution fight over an existing workflow, not a contest decided by model quality, and the more durable test of any vertical AI-native claim is whether the startup can get inside a workflow the customer already runs, not whether its underlying model has a cleverer learning loop.
Amble names four conditions she says predict a strong vertical AI market: an expert can quickly judge whether the AI’s output was right, the work requires real judgment rather than fixed rules, the job recurs often enough to generate training signal, and a startup can expand from one task into an entire job. She argues legal, tax, and accounting work clear that bar, and extends the logic to industrial quality-control work, where a manufacturing defect investigation crosses multiple systems that no single incumbent owns end to end.
The unresolved question in Amble’s own thread is whether Anthropic’s push toward general-purpose memory closes the gap she is describing. She concedes the labs are already building memory systems that let a general agent recall a customer’s past preferences, and argues this differs from learning because memory does not tell an agent whether a specific action was good or why an expert corrected it. Whether that distinction survives the next generation of frontier models is exactly the kind of claim that benefits from independent evaluation, and Amble’s thread offers none.
For operators evaluating a vertical AI vendor pitch over the next quarter, the question to ask is not which model the vendor has fine-tuned. It is whether the vendor already sits inside the workflow your team runs today, or whether it is asking you to route data out to a new interface that a Salesforce or Anthropic-level incumbent could absorb within a year.
Seema Amble, a fintech investor at Andreessen Horowitz, writing on X on September 3, 2026.
