The incumbents—Salesforce, Docusign, Atlassian, Klaviyo—aren't getting displaced by AI. They're moving up the value chain. Where they used to just store data and maybe slap on a chatbot, they're now building agents that take action: Docusign's Iris reviews contracts, Atlassian's Rovo routes requests, Klaviyo's Composer builds marketing campaigns. The real tension is whether these legacy systems stay the core interface or become invisible backends. Claudeforce shows one possibility: Claude sits on top as the front door while Salesforce owns the CRM data underneath. This unbundling creates a wedge for startups, but only if they can do something better than either the incumbent's specialized agent or Claude's general capabilities. The key insight is that jobs are bigger than the records these systems manage. A contract isn't the legal matter, a ticket isn't customer resolution, and an opportunity isn't the sale. The full work crosses applications, teams, and companies—different parties hold different pieces. General-purpose agents like Claude hit a wall here: they can theoretically access everything through integrations, but they face latency issues, data format mismatches between systems, and no access to information held by external parties. Vertical startups have room to win by owning more of the job end-to-end, which lets them see the decisions, corrections, and outcomes that actually matter. Where vertical startups build real advantage is through learning loops that general-purpose models can't replicate. A completed contract shows what got signed but not what alternatives were considered or why exceptions happened. A closed ticket shows resolution but not the hypotheses the support team tested. Vertical companies can manufacture a curriculum for the profession (how good experts do the work) and then use production data to learn how a specific firm operates (its templates, risk thresholds, escalation rules). The learning happens through the harness—the right context, tools, workflows, and evals—not just from hoarding historical data. Long-running agents that work over days with multiple checkpoints benefit most from this loop because each intermediate decision can be evaluated and improved.