Agentforce Coworker has 185,000 users. It takes 20 minutes to switch on, and longer to review what it can read
Salesforce's AI coworker runs inside existing Salesforce permissions, now across Teams, desktop and mobile. That makes the permission model, built for people, the real scope of the rollout.
Listen to this article · 5 min
AI-generated narration of the full article.
One of the customers in Salesforce’s latest Agentforce Coworker announcement is BACA Systems, a manufacturer whose Salesforce org holds about 1,300 custom objects. According to Salesforce, BACA enabled Coworker in 20 minutes. From that moment, an assistant could read across whatever those 1,300 objects expose to each user, and act on it.
That is the right way to read the headline number. Salesforce says Coworker has passed 185,000 users, with usage up 21 times in three months. Adoption at that speed is possible because there is almost nothing to build: Coworker works inside the permissions an organization already has.
What Coworker is now
Coworker is Salesforce’s assistant for employees. It plans the steps of a request, calls other Agentforce agents and workflows, and returns the result for the person to review. It reads Salesforce records, metadata and data connected through Data 360, within the user’s existing permissions.
The November release extends where it lives. Salesforce lists a Microsoft Teams integration, connections to Google Workspace and Microsoft 365, apps for iOS, Mac and Windows, conversation history and reusable Skills. Scheduled tasks and a Daily Digest are expected in December. Salesforce quotes a list price of US$50 per user per month, subject to change.
The customer examples are concrete. Fulton Bank, a regulated U.S. bank, reports more than 20 Coworker use cases in production for about 3,000 users. The Adecco Group is expanding it across 40 countries. A BACA manager describes tasks that used to take 45 to 60 minutes of switching between screens.
Inherited permissions were sized for people
“Operates within existing permissions” is the sentence that makes a 20-minute rollout safe enough to approve. It is also the sentence to examine.
Salesforce permission sets, sharing rules and field-level security were designed around what a person would plausibly do with access. A sales rep who can technically see every account in a region rarely opens more than a few dozen. Broad read access was tolerable because human attention limited its use.
An assistant removes that limit. Asked to “summarize every open issue with our top distributors”, it will read every record the user can reach, in seconds, and put the results in one answer that can be copied into a Teams chat or a Microsoft 365 document. Nothing in that sequence breaks a rule. It just exercises, at full breadth, access that was never meant to be exercised at full breadth.
We saw the same pattern when agents were found to make weak IAM fail faster: the agent does not create the overexposure, it reveals it.
What to review before the seats go out
Profiles with “view all”. Permission sets granting View All or Modify All on objects are the first place to look. They were often granted to solve one reporting problem years ago.
Sensitive fields on common objects. Compensation, health, credit and complaint fields on Contact or Account records will be included in summaries if the user can see them. Field-level security matters more than object access once a model is reading.
Where answers go. With Teams, desktop apps and Microsoft 365 connections, Coworker’s output leaves Salesforce. Data that was protected by staying inside the CRM now travels in a chat message. Label and retention policies on the destination apply, if they exist.
Skills and scheduled tasks. Reusable Skills and, from December, scheduled tasks turn one person’s request into a repeatable job. Each one needs an owner and a review date, like any automation, a point we made about Google’s coworker agents and their identity lifecycle.
The per-seat price sets the frame
At US$50 per user per month, Coworker is bought like a license, so it will be justified per seat. The more useful measure for a CIO is the opposite one: per seat, how much of the org can this assistant read, and has anyone looked? Fulton Bank’s 3,000 users and 20 use cases suggest a rollout that expanded use case by use case. That pace, more than the 20-minute switch, is the part worth copying.