@seema_amble - Incumbents vs. Vertical AI
@seema_amble on X The piece argues that systems of record like Salesforce won't be made irrelevant by AI, but vertical AI startups can still win by owning the whole job across systems and learning from expert corrections. It uses Harvey's Tenet work as the model for building that edge.
- Claudeforce lets users work with Salesforce CRM from inside Claude, making the general agent the front door while Salesforce keeps the data and actions.
- Incumbents mostly sit at the retrieval and process stages of the agent hierarchy, and their judgment stays bounded by the record they own, while the real job crosses applications and companies.
- Harvey post-trained an open-weight model on about 1,750 synthetic legal-task environments, averaging around 50 expert criteria each, rather than waiting years to collect customer data.
- Strong vertical markets share four traits: experts can judge output quickly, the work needs judgment, it repeats often, and a startup can start with one assignment and grow into the full job.
The core argument is that incumbent systems of record like Salesforce get more valuable as AI agents do more work through them, not less. The pitch is that a customer can pipe CRM or contract data into Claude or Codex and skip buying an AI-native app. Salesforce's Claudeforce, built with Anthropic, is the example: you work with Salesforce data from inside Claude without opening Salesforce. The author grants that incumbents are not sitting still. Docusign's Iris reviews contracts against a legal team's playbook, Atlassian's Rovo routes requests, and Klaviyo's Composer builds campaigns and scores their performance. Meanwhile, general agents are learning to coordinate across apps, which could make them a front-door layer sitting above the systems of record. That unbundling, the author says, opens a gap for AI-native startups to rebundle around a specific job.
The piece's main framework sorts application agents into four tiers by how much judgment they need: retrieval assistants, process agents, policy agents, and principal agents. Most incumbents were stuck at retrieval a year ago and are now climbing into process and some policy work. But that judgment stays bounded by the record the incumbent owns. The real customer job, the author argues, crosses applications, teams, and companies. A contract is not the legal matter, a ticket is not the resolution, and a sale is not the opportunity. Even a general agent with access to every internal app runs into latency through MCP, conflicting descriptions of the same customer across tools, and no visibility into what outside parties such as suppliers and buyers know. So the opening for vertical startups is to own the whole job across those boundaries rather than a local slice of it.
The strongest claim is about learning. A vertical company that owns more of the job can see the decisions, corrections, and expert overrides that produced the final result, and feed them into a learning loop around the model and its harness. Completed records alone make a weak curriculum: a signed contract shows what was agreed but not the alternatives rejected, and a closed ticket hides the hypotheses the support team tested. The author distinguishes learning how a good professional does the work, which can be taught through expert-built assignments and quality standards, from learning how one firm works, its templates, thresholds, and escalation rules, which improves through its own corrections over time. Memory alone won't hold up as a moat, since the labs are building general memory anyway, and long-running agents make the checkpointed learning loop matter even more.