A RevOps lead who spent last quarter scoping an internal agent to read opportunities out of Salesforce and draft the follow-up now has a vendor answer to compare it against. Andreessen Horowitz describes Claudeforce, the Salesforce product built with Anthropic, as letting you work with the CRM from inside Claude without opening Salesforce, and then appends its own "or so they claim!" to the sentence [1][2].
That parenthesis is the rollout question. What is being pitched is a front door; what the same passage concedes is that Salesforce still controls the underlying CRM data and the actions taken on it, with the interface and the system of record coming apart [6]. The front door is what gets demoed, and the data control is what turns up in your access review.
The economic half of the argument is thinner than the product half. The essay says incumbents have an incentive to control and charge for access to their data and to push their own agents forward [5]. It names no price, rate card, or access term in either place it ran [3]. This is a venture firm's essay, not a pricing page, and the two versions carry the same text, so it is one document published twice rather than two reads on the same market [1][16][17]. The incentive can be modeled; the invoice cannot, at least not yet.
The part worth taking to a planning meeting is the sort into four kinds of agent by how much autonomy and judgment each needs: retrieval, process, policy, principal [7]. A year ago most incumbent AI products sat at retrieval, working as chatbots and analytical tools over existing records [8]. Now more of them act inside the workflows, permissions and rules already in the product, and a few apply narrow human-defined policy, the example being Docusign's Iris running a legal team's playbook [9][10]. Count the tiers the essay puts a shipped product in and you get three of four; principal agents, the ones making open-ended calls on strategy and risk, have no incumbent example named [2].
The boundary the argument leans on is not technical. Incumbents can pull in outside data and add workflows, but their work expands outward from the record they already own, and judgment gets harder as soon as the decision depends on information that record does not hold [11][12].
So run two axes over each workflow on your agent roadmap. Does the work need data the incumbent's record does not hold, and does it need judgment a written playbook does not cover? Inside the record and inside a playbook is where Iris, Rovo and Composer already operate, and the honest answer is to buy and wait [3][4]. Inside the record but needing judgment is the incumbent's declared direction of travel, so a build there is a bet on their release schedule. Outside the record but rule-following is integration plumbing, and the essay's own point is that a general agent like Claude can reconstruct that from the underlying tools [14]. Outside the record and requiring judgment is the quadrant where the full job crosses applications, teams and companies whose parties hold different information and want different outcomes [13], which is what the essay contrasts with the incumbent's "local system of work" [15]. It is the only quadrant its logic leaves open, and the hardest one to show anybody in a thirty-minute call.
