Salesforce's customer system can now be run from inside Claude. A seller asks Claude to get ready for a call or fix the pipeline, and Claude reads the live records and makes the changes, within the seller's own permissions. Claude was already working inside Salesforce's own agents; this goes the other way, with Salesforce as the system underneath and Claude as the place where the work happens. It is a beta for paid Claude plans, and companies need Salesforce's approval to join.
The first step changes who uses the software
The usual argument is whether AI will kill software companies. A better question is what happens to everything between the company's records and the work itself. Most business software was built to help people pass work around: a queue so tickets reach the right person, a schedule so shifts are covered, a review tool so a manager can check a sample. Handing that same stack to an agent changes who clicks the buttons and little else, which is what happened when factories swapped one engine for another.
Access is not know-how
Giving an agent access to a system gives it tools, and it does not teach it the job. That is why the launch comes with 37 prebuilt sales skills, including meeting prep and deal health review. A skill is a package of instructions and examples that Claude loads when a request calls for it, so it does not need to be told again each time. Anyone can write one. A recruiter's screening steps, a shop's returns rules and a hospital finance team's month-end checks are all the kind of know-how that fits in one. The part of a job that is hard to copy is how your own business decides things, and that is now something you can write down and hand to an agent.
What happens to the six apps
Take a support team at a mid-sized online retailer. It probably pays for ticketing, call software, staff scheduling, quality review, a help-article tool and a customer survey tool. Each exists because people needed help coordinating.
Now picture the whole job given to a group of agents. One answers the customer. A second reads every finished conversation and marks where it went well or badly. A third notices that forty tickets this week were about the same late delivery, and drafts a fix to the help article. A fourth tests that fix against old tickets before any customer sees it. The next customer then meets a slightly better system. A human reviewer who reads a few dozen tickets a week sees a tiny slice of what happened, and an agent reads all of it, so the learning loop that used to be a quarterly project can run every day.
In that setup the queue, the review tool and the survey tool become steps inside one system instead of separate products with their own logins. Most of them are priced per human user, so a company with fewer humans in the loop will pay for them differently, or stop paying. Salesforce cut 4,000 support jobs while selling this, which shows which direction the work is moving.
Who wins
The records themselves stay valuable, because an agent needs a trusted place to read and write. The apps that only moved work between people are the exposed ones, and the companies that rebuild a whole function around agents will gain the most from it, as the owners who redesigned their factories did.
When a vendor shows an agent, ask which of the steps around the work disappear. If the answer is none, the company has bought a new engine for the old factory.