Prebuilt is useful when it gives the work a shape
Starting from a blank prompt is a bad way to introduce most people to AI at work. A support lead does not need a lecture on prompting; they need the return policy, account history and escalation route to meet in one place without making them rebuild the same request every morning.
Salesforce says each new agent comes with role-specific skills, actions and data models, and can be tailored to a company’s own processes. Its Agent Script documentation describes a way to combine model reasoning with deterministic rules. Those are vendor claims, not an independent test of how any particular deployment performs. Still, the direction makes sense. Familiar work needs familiar boundaries.
The catch is that a ready-made role can blur the difference between a common case and a real decision. “Employee support” may cover a password reset and a disputed leave record. “Customer service” may cover a tracking question and a shipment that never arrived. Packaging the first case well does not authorize the second one to disappear into a queue.
The handoff is where the product earns its name
Salesforce says its new long-horizon runtime will let its outbound-sales agent pursue a goal over days or weeks, retaining context and adjusting to feedback. That can be helpful. It also makes the handoff more important, not less.
Before a team gives any AI assistant a recurring job, write down what it must leave behind when it stops:
- what it saw and what it could not verify - actions it took, drafts it prepared, and changes it made - the next person’s actual choice, rather than a vague “needs review” - the customer, account or case that could be affected if the guess is wrong - the point at which the assistant should stop rather than keep trying
That is not paperwork for its own sake. It is how someone avoids spending their evening reconstructing a cheerful automated thread after the situation has become expensive or personal.
Try one boring job before you buy the story
A small team does not need seven agents to learn whether this category is useful. Pick one repeatable job with a clear end: route routine access requests, prepare first-pass return replies, or collect the missing details before a human follow-up. Keep the existing process for comparison.
Then look past the volume number. How often did the work come back? How many people had to reread the thread? Did anyone get a wrong answer that sounded settled? Did the person on the other side have to explain the whole problem again?
If the assistant removes a boring loop and leaves the strange cases legible, keep going. If it creates a nicer front door to the same pile of follow-up, call that what it is. The category does not need more proof that an AI can start work. It needs proof that the work ends better.